Add source pipeline classifier runtime

Correct source classifier activation schema
Publish source classifier model card
2026-10-18 15:31:06 +02:00 · 2026-09-15 23:17:48 -07:00 · 2026-09-15 22:55:46 -07:00 · 2026-09-15 22:36:34 -07:00 · 2026-09-15 22:00:05 -07:00 · 2026-09-15 21:32:46 -07:00
71 changed files with 3900 additions and 515 deletions
@@ -6,7 +6,7 @@
  },
  "metadata": {
    "description": "Agent skill and Claude Code plugin for the remove-ai-watermarks CLI.",
-    "version": "1.0.8"
+    "version": "1.0.9"
  },
  "plugins": [
    {
@@ -129,6 +129,10 @@ Do not classify an entire module as untestable because its main path downloads a
 - the photo-classify gate in `test_classify.py`: DEFINITELY is ridge AND MLP,
  POSSIBLY does not emit a provider, and `identify` does not import
  `remove_ai_watermarks.classify`.
+- the source-export classifier in `test_source_classify.py`: the Hub revision
+  and artifact hash are immutable, the pickle-free schema and exact GELU are
+  validated, threshold mutation changes the real decision, and `identify`
+  never consumes the result as watermark evidence.
 - the `InvisibleOptions` defaults, in `test_api.py`. When one signature promises to
  mirror another, compare them field by field rather than pinning the values you happen
  to know about, so the next field added on one side and not the other fails at the
@@ -8,7 +8,7 @@ historical data is sufficient.
 | Surface | Discovery evidence | Regression check | Removal proof |
 | --- | --- | --- | --- |
 | C2PA, EXIF, XMP, IPTC, PNG text, TC260, soft bindings, and container metadata | `scripts/corpus_gap_scan.py` plus review of unattributed and partially attributed reports | `scripts/sidecar_regression.py`; compare record-based and file-based identification when collection changes | `scripts/metadata_removal_audit.py`; signal absent after strip and decoded pixels/streams unchanged |
-| Registered visible image marks | `scripts/visible_positives.py`; metadata-derived vendor cohorts and blinded sheets for uncovered labels | `scripts/visible_eval.py`, registered examples, and full-corpus crossfire | `scripts/visible_removal_audit.py`; same-mark redetection plus visual residual review and fill-quality checks |
+| Registered visible image marks | `scripts/visible_positives.py` records strict and metadata-corroborated results, metadata cohorts, and blinded native-detail bands for uncovered cohorts plus an unbiased quiet sample | `scripts/visible_eval.py` exposes registry members without adjudicated scope, registered examples, and full-corpus crossfire | `scripts/visible_removal_audit.py`; same-mark redetection plus visual residual review and fill-quality checks |
 | Candidate China-AIGC visible marks | `scripts/vendor_cohort_harvest.py` partitions by producer identity without using pixels; inspect full-width top and bottom sheets | `scripts/vendor_mark_calibrate.py` against independently labelled positives, clean negatives, and neighbouring marks | Register only after the candidate gate survives full-corpus crossfire; then use the normal visible-removal audit |
 | Registered visible video marks | Review provider-labelled or controlled complete clips, not isolated frames alone | Full-clip tests and the local real-provider audit described in `docs/verification-plan.md` | Preserve sequence, timing, duration, and audio; second-pass detection must be quiet |
 | Locally decodable invisible image marks, including open DWT-DCT and TrustMark | Positive-control-gated corpus or constructed carriers; a negative on an uncontrolled carrier is inconclusive | `scripts/watermark_benchmark.py` and the relevant decoder tests with the required extras installed | Re-run the same decoder on output and retain a positive control in the same run |
@@ -42,6 +42,12 @@ compressed pixels, but its inventory should be reconciled with the current
 metadata constants and registry on every audit. Do not infer completeness from
 `No gap candidates` until the other rows of the matrix also ran.

+For a visible discovery pass, use a fresh report path and pass `--sheets` to
+`visible_positives.py`. Review the numbered PNG files before opening
+`MANIFEST_DO_NOT_OPEN.csv`; metadata cohorts select where a mark may occur but do
+not label the pixels. The separate `unbiased-quiet` stratum is required because a
+new provider may carry neither a mapped producer code nor readable provenance.
+
 ## When to replay old data

 Yes, old data must be rerun. Current code can learn to parse an old
@@ -7,8 +7,8 @@ name: Distribute on release
 #   - HF Space: factory-rebuild the version already pinned in the Space repo.
 #   - ComfyUI: sync, test, and publish the node against the exact library release.
 #   - ClawHub: publish the agent skill when its own version moved.
-# All wait for the freshly published sdist to appear on PyPI first, since the
-# Release event fires in parallel with publish.yml's upload.
+# All wait for the freshly published distribution to appear in both PyPI APIs
+# first, since the Release event fires in parallel with publish.yml's upload.

 on:
  release:
@@ -22,6 +22,9 @@ on:
 jobs:
  resolve:
    runs-on: ubuntu-latest
+    # Thirty attempts may spend up to 20 seconds each in the JSON request,
+    # simple-index request, and inter-attempt delay. Leave bounded headroom.
+    timeout-minutes: 25
    outputs:
      version: ${{ steps.v.outputs.version }}
      url: ${{ steps.sdist.outputs.url }}
@@ -63,26 +66,35 @@ jobs:
          path: release-identity.json
          if-no-files-found: error

-      - name: Wait for the sdist on PyPI, capture url + sha256
+      - name: Wait for the PyPI JSON and simple indexes, capture sdist metadata
        id: sdist
        run: |
          set -euo pipefail
          v="${{ steps.v.outputs.version }}"
          for i in $(seq 1 30); do
-            if rel=$(curl -fsSL "https://pypi.org/pypi/remove-ai-watermarks/$v/json" 2>/dev/null); then
-              url=$(echo "$rel" | python3 -c "import sys,json;d=json.load(sys.stdin);print(next((u['url'] for u in d['urls'] if u['packagetype']=='sdist'),''))")
-              sha=$(echo "$rel" | python3 -c "import sys,json;d=json.load(sys.stdin);print(next((u['digests']['sha256'] for u in d['urls'] if u['packagetype']=='sdist'),''))")
+            if rel=$(curl --connect-timeout 10 --max-time 20 -fsSL "https://pypi.org/pypi/remove-ai-watermarks/$v/json" 2>/dev/null); then
+              read -r url sha <<< "$(echo "$rel" | python3 -c "import sys,json;u=next((u for u in json.load(sys.stdin)['urls'] if u['packagetype']=='sdist'),{});print(u.get('url',''),u.get('digests',{}).get('sha256',''))")"
              if [ -n "$url" ] && [ -n "$sha" ]; then
-                echo "url=$url" >> "$GITHUB_OUTPUT"
-                echo "sha=$sha" >> "$GITHUB_OUTPUT"
-                echo "Found sdist for $v"
-                exit 0
+                # PyPI's JSON API can show a freshly uploaded file before the
+                # simple index used by pip has caught up. Downstream release
+                # jobs install by package version, so waiting on JSON alone
+                # creates a release-time race for them.
+                simple=$(curl --connect-timeout 10 --max-time 20 -fsSL https://pypi.org/simple/remove-ai-watermarks/ 2>/dev/null || true)
+                artifact=${url##*/}
+                if [[ "$simple" == *"$artifact"* ]]; then
+                  echo "url=$url" >> "$GITHUB_OUTPUT"
+                  echo "sha=$sha" >> "$GITHUB_OUTPUT"
+                  echo "PyPI JSON and simple indexes list $v"
+                  exit 0
+                fi
              fi
            fi
-            echo "sdist for $v not on PyPI yet (attempt $i), waiting..."
-            sleep 20
+            if [ "$i" -lt 30 ]; then
+              echo "PyPI indexes do not both list $v yet (attempt $i), waiting..."
+              sleep 20
+            fi
          done
-          echo "::error::sdist for $v did not appear on PyPI in time"
+          echo "::error::PyPI JSON and simple indexes did not both list $v in time"
          exit 1

  homebrew:
@@ -0,0 +1,64 @@
+name: Publish source classifier to Hugging Face
+
+# Manual only. The derivative model weights live in a GitHub model-freeze
+# release; the private training corpus and caches never enter this repository.
+
+on:
+  workflow_dispatch:
+    inputs:
+      confirm:
+        description: "Type publish"
+        required: true
+      weights_tag:
+        description: "GitHub Release tag holding source-pipeline-mlp.npz"
+        required: true
+        default: source-classify-freeze-2026-09-16-r2
+
+jobs:
+  publish:
+    if: ${{ github.event.inputs.confirm == 'publish' }}
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@v7
+
+      - uses: astral-sh/setup-uv@v7
+        with:
+          python-version: "3.12"
+
+      - name: Download derivative model weights
+        env:
+          GH_TOKEN: ${{ github.token }}
+          TAG: ${{ github.event.inputs.weights_tag }}
+        run: |
+          set -euo pipefail
+          mkdir -p weights
+          gh release download "$TAG" \
+            --repo "${{ github.repository }}" \
+            --pattern source-pipeline-mlp.npz \
+            --dir weights
+          test -s weights/source-pipeline-mlp.npz
+
+      - name: Verify the frozen artifact
+        run: |
+          set -euo pipefail
+          expected="$(jq -r .artifact_sha256 docs/source-classify-hf/metrics.json)"
+          test "${#expected}" -eq 64
+          echo "$expected  weights/source-pipeline-mlp.npz" | sha256sum --check
+
+      - name: Publish to Hugging Face
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          SHA: ${{ github.sha }}
+        run: |
+          set -euo pipefail
+          if [ -z "${HF_TOKEN:-}" ]; then
+            echo "HF_TOKEN secret is not set" >&2
+            exit 1
+          fi
+          uv run --no-project --with huggingface-hub \
+            python scripts/publish_source_classify_hf.py \
+            --src weights \
+            --message "Source-classify freeze (GitHub ${SHA:0:7})"
@@ -43,7 +43,7 @@ The published Agent Skill is [`skills/remove-ai-watermarks/`](skills/remove-ai-w

 [`docs/module-internals.md`](docs/module-internals.md) is the canonical per-module map, including design decisions, thresholds, calibration history, incident records, and regression guards. Read the relevant section before changing a subsystem.

-Research and current constraints are routed through [`docs/index.md`](docs/index.md), especially [`docs/known-limitations.md`](docs/known-limitations.md), [`docs/supported-signals.md`](docs/supported-signals.md), [`docs/synthid.md`](docs/synthid.md), and [`docs/watermarking-landscape.md`](docs/watermarking-landscape.md). Pixel photo classification is [`docs/photo-classify.md`](docs/photo-classify.md); it is not `identify`. Classifier research is split between general [`AI-generated image classifiers`](docs/ai-generated-image-classifiers.md) and [`SynthID source classifiers`](docs/synthid-classifiers.md). Other SynthID campaign logs are [`docs/synthid-detector-research.md`](docs/synthid-detector-research.md) and [`docs/synthid-removal-research.md`](docs/synthid-removal-research.md). The pre-split chronological archive is [`docs/synthid-detector-removal-plan.md`](docs/synthid-detector-removal-plan.md). The benchmark kernel and its pinned local oracles (AudioSeal, VideoSeal) cover the audio/video watermark cohorts and research studies: [`docs/benchmark-kernel.md`](docs/benchmark-kernel.md).
+Research and current constraints are routed through [`docs/index.md`](docs/index.md), especially [`docs/known-limitations.md`](docs/known-limitations.md), [`docs/supported-signals.md`](docs/supported-signals.md), [`docs/synthid.md`](docs/synthid.md), and [`docs/watermarking-landscape.md`](docs/watermarking-landscape.md). Pixel photo classification is [`docs/photo-classify.md`](docs/photo-classify.md); the separate OpenAI/Google/unknown export-pipeline signal is [`docs/source-classify.md`](docs/source-classify.md). Neither runs inside `identify` or constitutes SynthID detection. Classifier research is split between general [`AI-generated image classifiers`](docs/ai-generated-image-classifiers.md) and [`SynthID source classifiers`](docs/synthid-classifiers.md). Other SynthID campaign logs are [`docs/synthid-detector-research.md`](docs/synthid-detector-research.md) and [`docs/synthid-removal-research.md`](docs/synthid-removal-research.md). The pre-split chronological archive is [`docs/synthid-detector-removal-plan.md`](docs/synthid-detector-removal-plan.md). The benchmark kernel and its pinned local oracles (AudioSeal, VideoSeal) cover the audio/video watermark cohorts and research studies: [`docs/benchmark-kernel.md`](docs/benchmark-kernel.md).

 ## Data safety

@@ -3,7 +3,7 @@
 Remove AI provenance marks from images and video you generated yourself:

 - known visible labels such as the Google Gemini sparkle watermark and vendor text marks;
- invisible pixel watermarks through diffusion regeneration;
+- invisible pixel watermarks through direct local-format disruption or diffusion regeneration;
 - C2PA, EXIF, XMP, IPTC, and related AI metadata.

 Video support covers provenance identification, complete visible-plus-metadata
@@ -49,7 +49,9 @@ removal.

 Microsoft Paint and Photos InvisMark declarations are routed automatically to
 pixel regeneration. The `all` command removes both the hidden pixel watermark and
-its linked C2PA manifest; metadata stripping alone removes only the manifest.
+its linked C2PA manifest; metadata stripping alone removes only the manifest. A
+specialized Python API can inspect and disrupt the validated local
+`Watermarker.dll` payload without diffusion.

 ## Installation modes

@@ -57,15 +59,17 @@ its linked C2PA manifest; metadata stripping alone removes only the manifest.
 | --- | --- |
 | Metadata inspection and stripping | `remove-ai-watermarks` |
 | Photograph AI-versus-camera classification | `remove-ai-watermarks[classify]` |
+| OpenAI/Google/unknown source-export classification | `remove-ai-watermarks[source-classify]` |
 | Visible detection and removal | `remove-ai-watermarks[visible]` |
 | Visible video processing | `remove-ai-watermarks[video]` |
 | Video SynthID removal | `remove-ai-watermarks[video,diffusion]` |
 | Torch-free DWT-DCT detection | `remove-ai-watermarks[detect]` |
+| Direct local Paint InvisMark disruption | `remove-ai-watermarks[pixels]` |
 | Invisible image removal (needs CUDA) | `remove-ai-watermarks[qwen-zimage]` |
 | Every production feature available on the active Python | `remove-ai-watermarks[all]` |

 Lower-level and specialized extras include `pixels`, `heif`, `trustmark`,
-`migan`, `lama`, `diffusion`, and `classify-onnx`. The
+`migan`, `lama`, `diffusion`, `classify-onnx`, and `source-classify`. The
 [installation guide](docs/installation.md#feature-extras) documents their exact
 dependency composition, Python compatibility, and model requirements.

@@ -93,6 +97,19 @@ remove-ai-watermarks classify image.png

 Guide: [photo pixel classification](docs/photo-classify.md).

+For a lightweight, abstaining OpenAI/Google/unknown source-export signal after
+metadata removal, use the separate Python API. It is not a SynthID detector:
+
+```python
+import remove_ai_watermarks as raiw
+
+result = raiw.classify_source("image.png")
+print(result.label, result.reason)
+```
+
+Install `remove-ai-watermarks[source-classify]`. Guide:
+[source-pipeline classification](docs/source-classify.md).
+
 Signed provenance is the supported route for SynthID and `identify` reads it.
 There is no local SynthID pixel detector in the package. Research on a
 periodic lattice expert is in `scripts/synthid_runtime/` and
@@ -213,8 +230,10 @@ is not a product result state. The CLI therefore describes video-pixel
 regeneration instead of claiming a per-file SynthID verdict, and reports the
 copied audio watermark as `UNVERIFIED`.

-For invisible watermark removal, install the `qwen-zimage` extra. **An NVIDIA GPU
-is required**: all profiles are CUDA-only, and there is no CPU or MPS fallback.
+For invisible watermark removal through the CLI or high-level API, install the
+`qwen-zimage` extra. **An NVIDIA GPU is required**: all diffusion profiles are
+CUDA-only, and there is no CPU or MPS fallback. The specialized local Paint API
+uses the `pixels` extra without CUDA.

 ```bash
 uv tool install --force "remove-ai-watermarks[qwen-zimage]"
@@ -267,8 +286,8 @@ ControlNet, followed by SAM-masked Z-Image repair of any detected face. The
 alternative, `sdxl-zimage`, swaps the global stage for SDXL and keeps the same face
 stage. A third profile, `chroma-zimage`, uses the Apache-2.0 Chroma1 global pass
 with its own flat vendor floors; see `docs/chroma1-engine-research.md` for the
-calibration. `--pipeline auto` picks chroma-zimage for Microsoft
-provenance and qwen-zimage otherwise. All are CUDA only.
+calibration. `--pipeline auto` picks sdxl-zimage for Google, chroma-zimage for
+Microsoft, and qwen-zimage otherwise. All are CUDA only.

 ```bash
 uv tool install --force "remove-ai-watermarks[qwen-zimage]"
@@ -375,10 +394,10 @@ and keyed generation parameters are blanked without changing box sizes or
 media offsets. Other supported containers use their corresponding metadata
 path.

-Invisible removal is different. It regenerates the image through a diffusion
-pipeline to disrupt pixel and frequency domain watermarks. This changes the
-image and cannot guarantee that a proprietary verifier will reject every
-output.
+Invisible removal through the CLI and high-level API regenerates the image through
+a diffusion pipeline. Specialized callers can instead decode and disrupt the
+positively identified local Microsoft Paint format with much smaller pixel changes.
+Neither path can guarantee that every proprietary verifier will reject every output.

 See [supported signals](docs/supported-signals.md) and
 [known limitations](docs/known-limitations.md) for the full technical boundary.
@@ -473,9 +492,10 @@ invisible removal.
  The shipped profile is oracle-certified, but no public local decoder can
  certify an arbitrary output at runtime. Recheck unusually important outputs
  after provider changes.
- Invisible-watermark removal requires CUDA. All profiles refuse any other
-  device at construction rather than falling back to one that cannot run them.
-  Visible removal, metadata stripping and `identify` still run anywhere.
+- Invisible-watermark removal through the CLI and high-level API requires CUDA.
+  All diffusion profiles refuse any other device at construction rather than
+  falling back to one that cannot run them. The specialized local Paint API,
+  visible removal, metadata stripping and `identify` run without CUDA.
 - Provider watermark systems can change. Validate important outputs with the
  provider's own verifier when one is available.

@@ -490,7 +490,7 @@ The same option exists on `all` and `batch`, and as
 | `qwen-zimage` | Default. Qwen-Image-2512 global pass plus a SAM-masked Z-Image face stage |
 | `sdxl-zimage` | The same recipe and face stage on an SDXL global pass, at a higher denoise |
 | `chroma-zimage` | The same face stage on an Apache-2.0 Chroma1 global pass; explicit OpenAI use needs a higher floor than qwen-zimage |
-| `auto` | Pick the engine from provenance: chroma-zimage for Microsoft, qwen-zimage for OpenAI, Google, unknown, and an explicitly selected Meta cohort |
+| `auto` | Pick the engine from provenance: sdxl-zimage for Google, chroma-zimage for Microsoft, qwen-zimage for OpenAI, unknown, and an explicitly selected Meta cohort |

 **All four are CUDA-only.** There is no CPU or MPS profile for invisible-watermark
 removal. The former `controlnet`, `sdxl`, `qwen` and `default` profiles were removed
@@ -526,9 +526,13 @@ Install the combined extra and run only with an operator-verified manifest:
 ```bash
 uv tool install --force "remove-ai-watermarks[text-restoration]"
 remove-ai-watermarks invisible image.png -o clean.png \
-  --pipeline auto --text-manifest verified-lines.json --force
+  --pipeline qwen-zimage --text-manifest verified-lines.json --force
 ```

+The option works with `qwen-zimage`, `sdxl-zimage`, `chroma-zimage`, and `auto`.
+Each concrete profile reconstructs the donor through its own already loaded VAE;
+for Google provenance, `auto` therefore keeps the normal SDXL route.
+
 ``verified: true`` may also be set by an automated operator that attests
 machine-verified geometry: stability-gated detector boxes inside sane caps. Such
 operators should use the geometry-only schema 2, which carries no transcription
@@ -552,10 +556,11 @@ container changes remain valid while a resized or edited source fails closed. Th
 experimental helper
 `remove_ai_watermarks._internal.text_restoration.source_pixel_sha256` computes it.

-This mode is supported by `qwen-zimage`, `chroma-zimage`, and `auto` at native
-geometry with `humanize=0`, `unsharp=0`, and adaptive polish disabled. The legacy
-`sdxl-zimage` profile does not expose a verified-text VAE donor. `all` also accepts the flag,
-but its manifest must match the pixels entering the invisible stage; if visible-mark
+This mode is supported by every profile at native geometry with `humanize=0` and
+`unsharp=0`. An unset adaptive-polish option is forced off so verified text remains
+the final pixel-writing stage; an explicit request to enable it is rejected. `all`
+also accepts the flag, but its
+manifest must match the pixels entering the invisible stage; if visible-mark
 removal changes those pixels, the hash check rejects the run. One oracle verdict does
 not certify another manifest, seed, model/runtime version, or output hash.

@@ -13,6 +13,7 @@ to run the tool. Use the maintainer references only when changing the code.
 | [Supported signals](supported-signals.md) | You need to know which visible marks, metadata formats, and invisible signals are covered. |
 | [Known limitations](known-limitations.md) | You need the quality, device, format, or verification boundaries. |
 | [Photo pixel classification](photo-classify.md) | You want `classify` / `classify_pixels`: AI versus camera, optional provider. |
+| [Source-pipeline classification](source-classify.md) | You want an abstaining OpenAI/Google/unknown export-pipeline signal after metadata removal. |
 | [Photo classifier training](photo-classify-training.md) | You need the retrain pack, public sources, and what is not published. |
 | [Scope, safety, and legal notes](legal-and-safety.md) | You need the intended use and legal context. |

@@ -42,6 +43,7 @@ The current behavior is defined by the code, tests, README, and user guides.
 - [AI-generated image classifiers](ai-generated-image-classifiers.md) (photo AI-versus-camera freeze consumed by `classify_pixels`, not `identify`)
 - [Photo classifier Hugging Face card](photo-classify-hf/README.md) (Hub publication text of the 2026-08-31 freeze)
 - [SynthID source classifiers](synthid-classifiers.md) (metadata-free OpenAI/Gemini source finding and provider-lineage experiments)
+- [Source-pipeline classifier model card](source-classify-hf/README.md) (published derivative weights, aggregate evaluation, and intended-use boundary)
 - [SynthID local detector research](synthid-detector-research.md)
 - [SynthID mark removal research](synthid-removal-research.md)
 - [SynthID identity research](synthid-robust-identity-research.md)
@@ -94,11 +94,13 @@ Transformers cannot construct the Chroma tokenizer. It includes the `diffusion`
 dependencies; `diffusion` on its own covers the torch and diffusers imports but
 not the face stage or Chroma tokenizer, so it is not enough to run a removal.

-**An NVIDIA GPU is required.** `qwen-zimage`, `sdxl-zimage` and `chroma-zimage`
-are CUDA-only, and
+**An NVIDIA GPU is required for the diffusion profiles.** `qwen-zimage`,
+`sdxl-zimage` and `chroma-zimage` are CUDA-only, and
 construction refuses any other device rather than falling back to a slow or broken
-one. There is no CPU, MPS or XPU path for invisible-watermark removal. Visible-mark
-removal, metadata stripping and every `identify` command still run anywhere.
+one. There is no CPU, MPS or XPU path for those profiles. The specialized API for
+the validated local Microsoft Paint format uses the `pixels` extra and no GPU.
+Visible-mark removal, metadata stripping and every `identify` command still run
+anywhere.

 Video SynthID regeneration is a separate VAE path and does still run on CPU or MPS;
 it needs the `diffusion` extra, not this one.
@@ -119,7 +121,7 @@ application actually uses:

 | Extra | Capability | Automatically includes | Torch or model download |
 | --- | --- | --- | --- |
-| `pixels` | Shared BGR image runtime | NumPy, headless OpenCV | No |
+| `pixels` | Shared BGR image runtime and direct local Paint InvisMark disruption | NumPy, headless OpenCV | No |
 | `heif` | HEIC, HEIF, and AVIF pixel decoding | pillow-heif | No |
 | `visible` | Visible mark detection, OpenCV inpainting, and manual erasing | `pixels` | No |
 | `video` | Visible video identification/removal and timestamp preservation | `visible`, PyAV | No |
@@ -127,6 +129,7 @@ application actually uses:
 | `trustmark` | Adobe TrustMark detection on Python 3.11-3.12 | trustmark | Yes |
 | `classify` | Metadata-free photo AI-versus-camera classifier plus gated provider | `pixels`, Torch, Transformers | Yes |
 | `classify-onnx` | Optional CPU-only ONNX vision runtime for `classify` | `classify`, ONNX Runtime | Yes |
+| `source-classify` | Abstaining OpenAI/Google/unknown source-export classifier, not SynthID detection | `pixels`, Hugging Face Hub | Model download, no Torch |
 | `diffusion` | Torch and Diffusers runtime; video SynthID regeneration | `pixels`, Torch, Diffusers | Yes |
 | `migan` | MI-GAN ONNX fill backend | `visible`, ONNX Runtime | Model download, no Torch |
 | `lama` | big-LaMa ONNX fill backend | `visible`, ONNX Runtime | Model download, no Torch |
@@ -154,6 +157,7 @@ flowchart LR
    trustmark
    classify --> pixels
    onnx["classify-onnx"] --> classify
+    source["source-classify"] --> pixels
 ```

 `heif`, `trustmark`, and `text-draft` are independent branches. Combine them
@@ -20,6 +20,14 @@ The command is never started by `identify`. A no-signal provenance result stays
 unknown. This is not a clean verdict and it does not run cleanup. Full guide:
 [photo pixel classification](photo-classify.md).

+`classify_source` is a different, lightweight OpenAI/Google/unknown
+source-export classifier. It is useful only while the pixels still resemble an
+original export. A 95% resize reduced measured Google recall to 1%, and a JPEG
+quality-75 round trip made every public evaluation row abstain. Its provider
+label is not evidence that SynthID is present, and `unknown` is not evidence
+that SynthID is absent. It is also never started by `identify`. Full guide:
+[source-pipeline classification](source-classify.md).
+
 ## Visible removal

 ### Fill quality depends on the background
@@ -87,12 +95,20 @@ detections in the removed experimental mode.

 ## Invisible removal

-### Regeneration is lossy
+### Invisible processing changes pixels

-Invisible removal does not decode and delete a payload. It regenerates the
+Most invisible removal does not decode and delete a payload. It regenerates the
 image through a diffusion pipeline. Faces, text, colors, and fine detail can
 change even when the watermark is successfully disrupted.

+The local Microsoft Paint `Watermarker.dll` format is a narrower exception. Its
+direct path validates the repeated message and can require agreement with the
+signed C2PA UUID before rewriting confirmed carriers without redrawing the image.
+It still changes many pixels slightly, deliberately leaves an inverted periodic
+signal, and has no external detector that recognizes the positive control. The
+supported claim is therefore local payload invalidation, not certified removal by
+Microsoft.
+
 Text, tables, and UI screenshots are the worst case for a forced scrub: the
 model redraws glyphs as plausible-but-wrong shapes. Do not run the invisible
 stage on such content unless a locally-detectable invisible signal exists.
@@ -107,17 +123,29 @@ frozen text regions would keep the watermark
 ([text protection research](text-protection-research.md)).

 `qwen-zimage` is the default profile. `sdxl-zimage` and `chroma-zimage` keep
-the same face stage and swap the global model. `auto` picks chroma-zimage for
-Microsoft provenance and qwen-zimage otherwise. All are CUDA only.
+the same face stage and swap the global model. `auto` picks sdxl-zimage for
+Google provenance, chroma-zimage for Microsoft, and qwen-zimage otherwise.
+All are CUDA only.
 Qwen and SDXL condition the global stage on a canny edge map, which preserves
 structure but not identity or exact texture. Chroma1 is a plain strength pass
 without Canny.
-Existing face evaluations favor `qwen-zimage`, but there is no blanket fidelity
-ordering across content types. A fixed-seed, three-scene text comparison at the
-profile defaults found no stable winner: SDXL won one poster, Qwen won one, and
-the Chinese sign tied. Both are large, slow, and may still alter small text or
-difficult faces. The measurements and their OCR and oracle caveats are tracked
-in [`data/evaluations/fidelity/`](../data/evaluations/fidelity/README.md).
+There is no blanket fidelity ordering across providers or content types. At the
+current Google operating points, however, the 2026-09-15 three-carrier replication
+favored `sdxl-zimage` in whole-image metrics, localized face-crop LPIPS, and seven
+manually verified prominent CJK text regions. It retained 13/14 target lines exactly,
+versus 7/14 for Qwen and 4/14 for Chroma. ArcFace identity was less decisive against
+Qwen: SDXL won three of four scene-seed pairs and lost one, too few pairs for a
+significant result. A separate 18-face carrier at the current floors favored SDXL
+on ArcFace identity and whole-frame similarity in both seeds, while Chroma narrowly
+won face-crop LPIPS. SDXL also won a no-face/no-text photographic carrier, but Qwen
+won LPIPS on a flat graphic; one carrier cannot support content-adaptive routing.
+All three profiles regenerated fine newspaper body copy as pseudotext, so the result
+selects a relative Google default rather than promising exact face or text
+preservation. Both models are large, slow, and may still alter small text or difficult
+faces. The general fidelity corpus and its OCR caveats are tracked in
+[`data/evaluations/fidelity/`](../data/evaluations/fidelity/README.md), while the
+current-floor Google replication is recorded in
+[`module-internals.md`](module-internals.md#google-boundary-replication-2026-09-15).
 A global Z-Image Turbo prototype preserved text substantially better at low
 strength, but it has no useful cross-provider operating point and is not a
 supported profile. Automatic text restorers also remain research-only:
@@ -241,18 +269,18 @@ were measured for each.

 `qwen-zimage` reads unknown content from the resolution-adaptive denoise curve.
 Measured provider cohorts instead take flat operating points: OpenAI `0.07675`,
-Google `0.27`, and Microsoft InvisMark `0.15`. OpenAI and Microsoft add one full
+Google `0.35`, and Microsoft InvisMark `0.15`. OpenAI and Microsoft add one full
 observed cross-source boundary spread to the worst clean source; Microsoft's three
 first-clean boundaries were `0.04125`, `0.055`, and `0.095`, giving `0.14875` before
-rounding up. Google's candidate was separately repeated across three valid sources
-and three accounts. The small corpora make these operating points, not universal
-thresholds.
+rounding up. Google's source-fresh candidate cleared three independent CJK carrier
+types after the original content matrix found CJK and Cyrillic to be the limiting
+strata. The small corpora make these operating points, not universal thresholds.

 `sdxl-zimage` reads it from the C2PA issuer, on a flat ladder:

 - OpenAI: `0.15`;
- Google: `0.25`;
- unknown: `0.25`, following the stricter of the two.
+- Google: `0.50`;
+- unknown: `0.50`, following the stricter of the two.

 An SDXL global pass needs more denoise than Qwen at the same fidelity, and the
 values are flat rather than a curve because flat values are what was measured: each
@@ -268,13 +296,13 @@ established for that stage.
  boundary gives `0.20`, verified clean three times on both, while qwen-zimage
  cleared both at its existing `0.07675` operating point;
 - Microsoft InvisMark: `0.125` (below qwen's `0.15`);
- Google: `0.40` for zero-face content (above qwen's `0.27`; at this floor the
-  regeneration destroys dense text and collapses face identity -- the tradeoff
-  that motivates the adaptive arm below), or `0.125` when YuNet detects at
-  least one face (both measured face fixtures clear at 0.12 with zero
-  cross-source spread; oracle-verified 2026-08-30);
+- Google: `0.50`. A source-fresh CJK carrier remained detected at `0.40` and
+  cleared at `0.50`; that candidate then cleared independent busy-scene,
+  flat-text, and face-plus-text carriers. The old YuNet-only `0.125` face arm
+  was removed because face detection cannot distinguish easy face-only images
+  from harder face-plus-text content;
 - Meta Content Seal: `0.17` (above qwen's `0.1`);
- unknown: `0.40`, following the strictest measured cohort.
+- unknown: `0.50`, following the strictest measured cohort.

 An explicit `--strength` overrides all three. The defaults are operating points,
 not universal guarantees. Near a removal threshold, different content or a
@@ -288,8 +316,8 @@ are certified at a fixed seed. The live resolver is
 | --- | --- |
 | `qwen-zimage` | CUDA only, large model stack, and limited broad certification across seeds and content. |
 | `sdxl-zimage` | CUDA only. Its strength ladder is flat per vendor, not a resolution curve, because flat values are what was measured. |
-| `chroma-zimage` | CUDA only. Higher Google and Meta floors than qwen; Google face content uses 0.125 instead of 0.40. |
-| `auto` | CUDA only. Routes Microsoft to chroma-zimage and everything else to qwen-zimage. |
+| `chroma-zimage` | CUDA only. Higher Google and Meta floors than qwen; no safe content-adaptive Google arm exists yet. |
+| `auto` | CUDA only. Routes Google to sdxl-zimage, Microsoft to chroma-zimage, and everything else to qwen-zimage. |

 Only manually verified `vae-glyphs` is an optional production stage, and it is
 experimental rather than a default.
@@ -484,6 +484,19 @@ not added to the watermark inventory, do not trigger pixel regeneration, and do
 not suppress independently established SynthID watermark evidence. A registered
 watermark or an unknown soft-binding algorithm keeps that inference fail-safe.

+The local Paint reader treats C2PA as corroboration, not as the pixel detector.
+It extracts 4-by-4 blocks from the luminance Haar LL plane, reads the largest
+singular-value residue of the 15 DCT AC coefficients, and votes over the repeating
+12-by-12 map. A payload is actionable only when every bit has at least three
+supporting carriers and the `0x4c` prefix and GUID checksum validate. A caller can
+also require the decoded Windows-order GUID to equal the one in pristine C2PA.
+Direct disruption moves each confirmed carrier to the opposite residue center,
+encodes to a temporary file, and re-runs the reader before atomically publishing
+it. Any missing or ambiguous
+condition returns no detection so a caller can retain its existing fallback. The
+implementation contains no Microsoft binary, PDB data, or extracted writer-quality
+forest.
+
 The SDK default enables trust verification but supplies no production trust
 anchors. Consequently, an installation without an explicitly maintained C2PA
 trust bundle reports every otherwise valid signer chain as untrusted --
@@ -510,13 +523,14 @@ convert `signingCredential.untrusted` into trusted based on a vendor-name match.
 Vendor attribution comes from the registry in
 [`_internal/constants.py`](../src/remove_ai_watermarks/_internal/constants.py). Derived
 issuer and platform maps should not be maintained separately.
-For an AI C2PA claim, a recognized product in `claim_generator` takes precedence
-over the certificate issuer: an application can sign through an upstream model
-provider without becoming that provider's product. Only exact product mappings
-receive this precedence; an unknown claim generator still falls back to issuer
-attribution. An unmapped issuer org reads as unknown-signer C2PA with no platform;
-that is how Ideogram was surfaced (4 private-corpus files signed "Ideogram, Inc",
-2026-08-08) before its vendor row was added on 2026-08-27.
+A recognized product in `claim_generator` takes precedence over the certificate
+issuer: an application can sign through an upstream model provider without
+becoming that provider's product. Exact products are also useful provenance on a
+non-AI edit, such as CapCut; recognizing the product does not itself change the AI
+verdict. An unknown claim generator falls back to issuer attribution. An unmapped
+issuer org reads as unknown-signer C2PA with no platform; that is how Ideogram was
+surfaced (4 private-corpus files signed "Ideogram, Inc", 2026-08-08) before its
+vendor row was added on 2026-08-27.

 ### Metadata scanning and stripping

@@ -1102,6 +1116,30 @@ both backends explicitly (`tests/test_classify.py` pins the seam). User guide:
 [photo-classify.md](photo-classify.md). Hub card:
 [photo-classify-hf/README.md](photo-classify-hf/README.md).

+### Source-pipeline classifier
+
+[`source_classify.py`](../src/remove_ai_watermarks/source_classify.py) is a
+separate, lightweight OpenAI/Google/unknown source-export classifier. It fuses
+the existing 124-dimensional patch descriptor with the 768-dimensional
+phase-free spectrum in
+[`_internal/source_spectral.py`](../src/remove_ai_watermarks/_internal/source_spectral.py),
+then evaluates the frozen 892 -> 128 -> 64 -> 3 GELU MLP in NumPy. Provider
+labels clear independent calibrated margins; every other row is `unknown`.
+The NumPy decision treats values within `1e-5` of the stored margin as the
+inclusive training threshold, absorbing float32 BLAS differences; this matches
+the frozen training-runtime decisions on all 27,811 cached internal and
+external feature rows.
+
+The runtime pins both the immutable Hub revision and the NPZ SHA-256, loads
+with `allow_pickle=False`, and rejects any field, shape, class, activation, or
+non-finite-value drift. `RAIW_SOURCE_CLASSIFY_WEIGHTS` is the offline override.
+The `source-classify` extra adds `pixels` and Hugging Face Hub but no Torch or
+Transformers. `identify` never imports or runs this path. Its output names a
+complete source/export pipeline and is not SynthID detection. Unit contracts
+and the threshold mutation guard live in `tests/test_source_classify.py`;
+evaluation and attack limits live in
+[`synthid-classifiers.md`](synthid-classifiers.md).
+
 The CLIP loader suppresses discarded random parameter initialization with
 Transformers' `no_init_weights` context before applying the complete frozen
 state dict. Normal `CLIPModel(config)` construction initialized 427 million
@@ -1651,7 +1689,8 @@ The current profiles are `qwen-zimage` (the default), `sdxl-zimage`,
 `qwen` and `default` were removed rather than kept as a CPU path, and are
 rejected rather than aliased onward. `auto` is a deterministic per-cohort
 selection policy:
-chroma-zimage for Microsoft, qwen-zimage for OpenAI, Google, Meta, and unknown.
+sdxl-zimage for Google, chroma-zimage for Microsoft, and qwen-zimage for
+OpenAI, Meta, and unknown.
 It does not run a learned router or classify the image's genre.

 `qwen-zimage` normally resolves global denoise from image area for unknown content.
@@ -1824,8 +1863,8 @@ verification, and geometry-only operators do not need to invent text or script f

 When enabled, the one profile selected before model loading reconstructs the source
 once through its already loaded VAE, runs the ordinary global and face stages, and
-calls the shared restoration compositor. Qwen and Chroma implement this donor hook;
-SDXL does not. The optional Qwen-only fidelity anchor first blends 15% of the VAE
+calls the shared restoration compositor. Qwen, SDXL, and Chroma implement this donor
+hook. The optional Qwen-only fidelity anchor first blends 15% of the VAE
 reconstruction into the clean result; it is off by default because the blend returned
 detector-visible OpenAI SynthID in the measured poster fixtures. The compositor derives
 binary source and candidate silhouettes, groups nearby lines, uses LaMa for the initial
@@ -1839,8 +1878,9 @@ extend 8% above and 25% below. The anchored-component gate covers clipped leadin
 flourishes, trailing punctuation, icons, and descenders without walking into
 disconnected decoration or background texture.

-The stage is deliberately narrower than the engine: it rejects `sdxl-zimage`, tiles,
-resolution caps, humanize, unsharp, and adaptive polish. Those combinations change
+The stage is deliberately narrower than the engine: it rejects tiles, resolution
+caps, humanize, unsharp, and explicitly enabled adaptive polish. An unset polish
+value resolves to off in this mode, including for SDXL, because those combinations change
 geometry or final pixels after the verified layer and have no measured oracle result.
 It remains opt-in because annotations are manual and provider verdicts apply only to
 the exact tested output hashes, not to the mechanism in general.
@@ -1864,6 +1904,22 @@ the `chroma_text_restoration_study.py` harness (kept outside this repository)
 under the gitignored
 `out/text-restoration-engine-study/` directory.

+The 2026-09-15 rerun added the SDXL donor implementation and exercised all three
+profiles on the same three manifests. SDXL improved text-box MAE from 11.319 to
+7.970 and from 10.704 to 7.657 on the two OpenAI posters, while PSNR rose from
+25.83 to 29.24 dB and from 26.17 to 29.04 dB. On the tracked Google CJK carrier
+(`4affd7f27767a445db6abf741355743ba8d95108ad922c9fff045feed8492236`, already
+recorded as SynthID-positive), SDXL at the calibrated 0.50 strength improved
+text-box MAE from 10.738 to 8.564 and PSNR from 23.93 to 24.00 dB. The exact
+restored output
+`035e4919e1d84129a347006c98f19eb4431eb5772788f6d2097b94a2c6baec84`
+was submitted to the Google Gemini oracle after provenance stripping preserved
+its decoded pixels; the sanitized upload hash was
+`3e4f06274a7b735f7d06863fd54dc8146f60aeb9cc8315f86294180deb660b82`.
+The oracle returned `not_detected` on 2026-09-16. This
+certifies that one carrier, manifest, seed, runtime, and output hash; it does not
+turn the compositor into a provider-wide guarantee.
+
 A matched stage-isolation check on the 18-face Gemini portrait grid confirms the
 division of responsibility. The visible-cleaned, metadata-stripped control and the
 Z-Image face-only output were both SynthID-positive; Qwen global-only and the full
@@ -1922,22 +1978,20 @@ prompt, which was never calibrated against Chroma1), guidance 5.0, and the
 calibrated `ceil(4 / strength)` requested-step schedule.
 `CHROMA_NOMINAL_STEPS = 4` is a nominal count: Chroma rounds the start
 index, so strengths 0.09 and 0.17 execute five steps, while 0.20, 0.125,
-and 0.40 execute four. Preserve those requests rather than changing the
+and 0.50 execute four. Preserve those requests rather than changing the
 calibration to force four effective steps. There is no Canny
 conditioning: the floors were measured on a plain strength pass.

 The flat vendor floors (`CHROMA_ZIMAGE_*_STRENGTH` in watermark_profiles):
-OpenAI 0.20 and Microsoft 0.125, Google 0.40 for zero-face content / 0.125 when
-faces are detected (a content-adaptive arm from the clean face-count split
-in the calibration: both text cards need 0.25, both face fixtures clear at
-0.12, so the YuNet detector -- already loaded for the face stage -- routes
-the operating point), and Meta 0.17 (ABOVE qwen's, because Chroma1's
-per-fixture boundaries scatter wider). The matched-strength addendum in the
-research doc is the honest read: at the strength each image actually needs,
-Chroma1 regenerates better almost everywhere except face identity (which the
-inherited Z-Image face stage supplies); at the flat worst-case floors it
-destroys dense text and face identity. The Google face-content arm is the
-first shipped piece of the content-adaptive policy. A Meta arm was measured
+OpenAI 0.20, Microsoft 0.125, Google 0.50, and Meta 0.17. The source-fresh
+Google sweep supersedes the older 0.40 zero-face / 0.125 face split: fresh CJK
+remained detected at 0.40, and a face-plus-text carrier did not support treating
+face presence as evidence for the easy arm. The runtime has no independently
+measured text-content gate, so every Google carrier now uses the replicated global
+floor. The matched-strength addendum in the research doc remains useful history:
+at the strength each image actually needs, Chroma1 regenerates better almost
+everywhere except face identity (which the inherited Z-Image face stage supplies),
+but high flat floors can destroy dense text and face identity. A Meta arm was measured
 on 2026-08-30/31 (docs/chroma1-engine-research.md, expansion section) and
 does not ship: first-cleans from 0.03 to 0.10 form a continuum,
 `flat_ratio` overlaps the hard and easy clusters, and a high-flat
@@ -1973,8 +2027,9 @@ is supported by this pass, so `auto` keeps the measured provenance-cohort table.

 [`_internal/sdxl_zimage_pipeline.py`](../src/remove_ai_watermarks/_internal/sdxl_zimage_pipeline.py)
 runs the same two-stage recipe on an SDXL global pass. `SdxlZImagePipeline`
-subclasses `TwoStageZImagePipeline` and implements only the global stage
-(`_load_global`, `_run_global`), so the face stage is inherited rather than copied
+subclasses `TwoStageZImagePipeline` and implements the global stage
+(`_load_global`, `_run_global`) plus the deterministic verified-text VAE donor, so
+the face stage is inherited rather than copied
 and cannot drift between the profiles; a test asserts the shared methods are the
 same objects.

@@ -1985,13 +2040,213 @@ Four things are architecture-bound and swap with the model: the ControlNet
 AuraFlow shift), and the latent grid (8 px against Qwen's 16).

 **Strength is architecture-bound too, and that is the easy mistake.** An SDXL global
-pass leaves SynthID at the strength Qwen needs: verified through the Gemini app on a
-native 2816x1536 original, 0.154 is FOUND while 0.20, 0.25 and 0.30 are clean. So this
+pass can leave SynthID at the strength Qwen needs. The source-fresh 2026-09-15
+matrix found its limiting Google CJK carrier detected through 0.45 and clean at
+0.50, with independent flat-text and face-plus-text replication at 0.50. So this
 profile takes a vendor policy (`SDXL_ZIMAGE_OPENAI_STRENGTH` 0.15,
-`SDXL_ZIMAGE_GEMINI_STRENGTH` 0.25, unknown following Gemini) rather than
+`SDXL_ZIMAGE_GEMINI_STRENGTH` 0.50, unknown following Google) rather than
 `resolution_adaptive_denoise`. Flat values are what was measured; no size dependence
 has been established for this stage, so none is asserted.

+#### Cross-profile strength sweep, 2026-09-14
+
+A fresh single-source, seed-0 sweep compared qwen-zimage, chroma-zimage, and
+sdxl-zimage on the same provider-specific carriers. The original content matrix covered
+photorealistic scenes, flat low-detail art, single and multiple faces, dense Latin,
+CJK and Cyrillic text, multilingual posters, and faces plus text. Meta completed
+48/48 real variants, Microsoft 12/12, the original OpenAI sweep 36/36, and Google
+44/48. Thirteen unique adaptive OpenAI extension candidates later closed the two
+missing content quadrants. The minimum
+tested global strengths that cleared every completed stratum were:
+
+| Provider | qwen-zimage | chroma-zimage | sdxl-zimage |
+| --- | ---: | ---: | ---: |
+| OpenAI | 0.06 | 0.12 | 0.06 |
+| Microsoft | 0.10 | <= 0.04 | <= 0.08 |
+| Meta | 0.06 | 0.10 | <= 0.08 |
+| Google | 0.35 | 0.50 | 0.50 |
+
+`<=` means the first tested rung already cleared, not that its lower boundary is
+known. Google qwen cleared both CJK and Cyrillic only at 0.35; its face series was
+non-monotonic. Google SDXL cleared Cyrillic at 0.35 after detections through 0.28,
+but CJK remained detected at 0.35, 0.40, and 0.45 before clearing at 0.50. Google
+Chroma CJK remained detected through 0.40 and cleared at 0.50; Cyrillic cleared at
+0.40. Chroma face+text first cleared at
+0.22 after detections at 0.08 and 0.16; the face carrier cleared at 0.22 after a
+detection at 0.08. A recovered-quota recheck on 2026-09-15 also detected the face
+carrier at 0.16, closing that tested bracket at `0.16 detected -> 0.22 clean` and
+showing that the shipped 0.125 face arm is too low for this carrier. These are
+candidate rungs at this stage of the study, not replacement defaults: there is one
+source and one seed per case. The independent three-carrier replication below later
+closed the production seed-0 decision; size, crop/resize, JPEG, local-edit, and broader
+multi-seed robustness remain separate holdouts.
+
+The original OpenAI set had faces-only and multilingual text-only carriers but no
+explicit no-face/no-text carrier. A fresh 1536x1024 photorealistic rocky-coast scene
+closed that stratum on 2026-09-15. The official OpenAI verifier found Qwen 0.04 and
+Chroma 0.10, then cleared Qwen 0.06 and Chroma 0.12. It found SDXL 0.04 and cleared
+0.06. These are 0.02-grid brackets, not mathematical minima. The public Web verifier
+settled the earlier points and exposed SynthID separately from C2PA. A later Chroma
+0.12 Web request remained in `Verifying` for more than 90 seconds, so it was recorded
+as indeterminate; the preferred official Content Provenance API returned
+`not_detected` for that same processed source hash. Three independent official API
+requests returned HTTP 200 for Chroma 0.12, SDXL 0.04, and SDXL 0.06 respectively. No source control was
+submitted.
+
+The other missing OpenAI stratum, one face plus multilingual text, was added with a
+second 1536x1024 carrier. The official API found Qwen 0.04 and cleared 0.06, while
+Chroma 0.10 and SDXL 0.04 already cleared. All responses were HTTP 200 with C2PA
+`not_present`, and original-size review found the face identity, Latin, CJK, Cyrillic,
+and date text preserved at every generated rung. This stratum does not raise any
+global OpenAI candidate; the no-face/no-text scene remains the limiting case for
+Chroma and SDXL.
+
+At the original non-Google candidate rungs, Chroma had the best mean quality in all
+three provider cohorts. For Meta, mean LPIPS/SSIM was 0.0320/0.8882 for Chroma 0.10,
+0.0357/0.8595 for Qwen 0.06, and 0.0531/0.8503 for SDXL 0.08. For OpenAI it was
+0.0183/0.8763, 0.0192/0.8595, and 0.0304/0.8459 respectively over the original three
+carriers. Recomputed over all five OpenAI carriers at the freshly verified candidates,
+Qwen 0.06 has the best LPIPS, SSIM, and edge F1 at 0.03073, 0.82461, and 0.89141.
+Chroma 0.12 has the best MAE by a narrow margin at 6.42, but its LPIPS/SSIM is worse
+at 0.05057/0.80940 and Qwen narrowly leads PSNR 27.99 to 27.97 dB; Chroma's
+no-face/no-text scene is the main loss.
+SDXL 0.06 is third on the aggregate perceptual and structural metrics at
+0.03766/0.77202, while keeping the mean Laplacian ratio closest to one. Original-size
+review agrees that there is no single winner on every criterion, but Qwen is the best
+overall OpenAI candidate when perceptual similarity, structure, edges, faces, and text
+are weighted together. Microsoft's
+single flat carrier likewise favored Chroma 0.04 at 0.00265/0.99621 over Qwen
+0.10 at 0.00347/0.99426 and SDXL 0.08 at 0.00324/0.99566; this is a carrier
+comparison, not a multi-source mean. At Google 0.35, SDXL preserved substantially
+more quality than Qwen (mean LPIPS 0.0655 versus 0.1576, SSIM 0.7285 versus
+0.5457), although SDXL needed 0.50 rather than 0.35 on the limiting CJK carrier.
+On that carrier, SDXL 0.50 measured LPIPS/SSIM/PSNR 0.1041/0.7437/23.93 dB,
+substantially better than Chroma 0.50 at 0.2129/0.5853/20.74 dB and Qwen 0.35 at
+0.1834/0.5550/18.93 dB. The newly cleared Chroma 0.22 face and face+text points also lost to
+SDXL at its corresponding tested rungs on every reported metric: LPIPS/SSIM was
+0.0476/0.6657 versus 0.0318/0.7036 for faces, and 0.0506/0.6359 versus
+0.0426/0.7016 for face+text. Original-resolution visual review agreed with that
+direction: Qwen
+0.35 changed the last glyph of the Cyrillic `УЛЫБКОЙ` into a Latin-looking `N` and
+altered substantially more wood grain and background structure; Chroma 0.40 and
+SDXL 0.35 retained the word, with SDXL visually closest to the source. Chroma 0.22
+preserved foreground face identity and text legibility but visibly changed
+background people, room structure, lighting, and confetti. The
+official-oracle access and rolling-quota caveats are recorded in
+[`provider-oracles.md`](provider-oracles.md).
+
+The original Google carriers omitted the explicit no-face/no-text stratum. A
+2026-09-14 extension added two real Gemini-generated 2752x1536 carriers: a
+photorealistic alpine lake and a low-detail geometric desert. At the current
+candidate rungs, the flat carrier remained close under all three profiles;
+LPIPS/SSIM was 0.0237/0.9654 for Qwen 0.35, 0.0120/0.9706 for Chroma 0.40,
+and 0.0134/0.9655 for SDXL 0.35. On the photographic carrier, every profile
+visibly regenerated scene detail. SDXL 0.35 was least destructive at
+0.1174/0.4933, ahead of Qwen 0.35 at 0.2070/0.4120 and Chroma 0.40 at
+0.2216/0.4229. Original-resolution review found major mountain, shoreline,
+vegetation, rock, and reflection changes in Qwen and Chroma; SDXL retained the
+main composition and geometry but still changed fine detail. Full four-rung
+ladders produced 24 variants across the two carriers and three profiles. After the
+independent session allowances recovered on 2026-09-15, the official verifier cleared
+both Qwen carriers at the first 0.15 rung. Chroma's flat carrier cleared at 0.12,
+while its photographic carrier closed at `0.20 detected -> 0.30 clean`. SDXL's
+photographic carrier cleared at 0.15 and its quality-qualified flat 0.35 carrier
+cleared directly. These scene cases therefore do not raise the global Google
+candidate above the text-bound values in the table.
+The SDXL flat ladder was quality-non-monotonic: 0.15 introduced severe texture
+and scored LPIPS 0.1283, while 0.35 looked cleaner and scored 0.0134. Repeating
+0.15 and 0.22 with adaptive polish disabled changed the polished outputs by only
+0.51 and 0.57 mean byte levels per channel and left LPIPS effectively unchanged.
+The artifact therefore comes from the low-step SDXL reconstruction, not polish.
+The 0.15 flat variant is rejected before oracle use; 0.35 is the
+quality-qualified SynthID candidate for that one path.
+
+A Chroma extension was run on an isolated CUDA research worker from the same
+`c4edc282` checkout. Strength 0.50 retained readable
+CJK and Cyrillic but was already borderline: LPIPS/SSIM/PSNR was
+0.2129/0.5853/20.74 dB and 0.2268/0.5490/20.83 dB respectively. At 0.60 both
+scripts were visibly destroyed and LPIPS exceeded 0.42. The official verifier
+subsequently detected CJK 0.40 and cleared 0.50, making the borderline 0.50 rung the
+fresh global Chroma candidate. Strength 0.60 remains rejected on fidelity.
+
+The corresponding SDXL CJK extension used the same checkout and an isolated CUDA
+research worker. The official verifier detected 0.40 and 0.45 and cleared 0.50.
+At 0.50, LPIPS/SSIM/PSNR was 0.1041/0.7437/23.93 dB; original-resolution review
+found the CJK text fully legible and the scene recognizably preserved. SDXL is the
+best-quality Google candidate on the limiting carrier.
+
+#### Google boundary replication, 2026-09-15
+
+Three new Gemini-generated controls replicated the hard boundary on independent
+content: a busy night market with dense CJK and many faces, a flat CJK poster without
+faces, and a portrait reading a dense CJK newspaper. An isolated CUDA run generated
+both seed 0 and seed 1 at the candidate and neighboring lower rungs, 36 variants total. Production
+uses the fixed seed 0, and only processed seed-0 variants were sent to the official
+verifier; source controls were not uploaded.
+
+At the candidate rungs, the official verifier returned `not_detected` for all nine
+profile-by-carrier cells: Qwen 0.35, Chroma 0.50, and SDXL 0.50. A separate lower-rung
+night-market check also cleared Chroma 0.40 and SDXL 0.45, but those two results do
+not lower the policy because the original fresh CJK carrier remained detected there.
+Three quota failures and one duplicate post-success tool-call failure were recorded
+as `unreachable` in separate immutable batches and excluded from the watermark
+count. Every settled batch passed source, upload, manifest, and result hash
+verification.
+
+Quality over all three new carriers and both seeds selected SDXL for Google auto
+routing. Mean LPIPS/SSIM/PSNR was 0.0609/0.8178/25.15 dB for SDXL 0.50,
+0.0980/0.7055/21.85 dB for Qwen 0.35, and 0.1269/0.6909/22.53 dB for Chroma 0.50.
+Original-resolution review agreed: SDXL best preserved the night-market headline,
+newspaper layout, faces, and the flat poster. Therefore the production Google floors
+are Qwen 0.35, Chroma 0.50, and SDXL 0.50, and `auto` routes Google to SDXL.
+
+A localized follow-up checked that aggregate image metrics were not hiding face or
+text failures. InsightFace `buffalo_l` matched three faces in the newspaper scene and
+15 in the night market one-to-one. Across the four scene-seed pairs, SDXL beat Qwen
+on ArcFace cosine in three and lost one, and beat Chroma in all four. That identity
+result is directional but small: the two-sided sign-test p-values are 0.625 and
+0.125, respectively. Face-crop LPIPS favored SDXL in all four pairs against both
+other profiles, also p=0.125 with only four non-tied pairs. This supports, but does
+not by itself prove, an identity advantage over Qwen.
+
+Text was scored separately in seven manually verified fixed regions whose originals
+all produced an exact OCR line: the newspaper masthead, headline, two halves of its
+subhead, the night-market headline, and two poster lines. The metric was the closest
+OCR-line normalized edit distance to that fixed target, across both seeds. SDXL's
+mean was 0.0119 with 13/14 exact region-seed results, versus 0.3770 and 7/14 for Qwen
+and 0.5374 and 4/14 for Chroma. In paired sign tests SDXL was better than Qwen in
+seven non-tied cells and worse in none (two-sided p=0.015625), and better than Chroma
+in ten and worse in none (p=0.001953125). These results cover the selected prominent
+lines, not arbitrary text preservation: fine newspaper body copy was regenerated as
+pseudotext by every profile. The safe claim is therefore that SDXL was the relative
+quality winner on these replicated Google carriers, not that it preserves faces or
+text losslessly.
+
+Two current-floor extensions closed the remaining content quadrants. On the
+face-only carrier, rerun at Qwen 0.35, Chroma 0.50, and SDXL 0.50 for seeds 0 and
+1, InsightFace matched all 18 faces. SDXL had the highest mean ArcFace cosine in
+both seeds, 0.828 and 0.842, versus 0.780/0.794 for Chroma and 0.744/0.773 for
+Qwen. Chroma narrowly won face-crop LPIPS at 0.059/0.058, versus 0.060/0.063 for
+SDXL, while SDXL was much closer over the whole frame. Its face Laplacian ratios
+of 1.152/1.176 also show slight texture amplification; Qwen and Chroma instead
+smoothed face detail to roughly half the source level. Original-resolution review
+therefore favors SDXL for overall composition and identity, but not every local
+face metric.
+
+The no-face/no-text photographic and flat-graphic carriers were also rerun at the
+same current floors and two seeds. On the photographic scene, SDXL had the best
+LPIPS/SSIM in both seeds: 0.345/0.457 and 0.339/0.461, ahead of Qwen at
+0.413/0.412 and 0.404/0.414 and Chroma at 0.568/0.395 and 0.579/0.401.
+Original-resolution review likewise found SDXL closest, although every profile
+regenerated major landscape detail. The flat graphic reversed the LPIPS result:
+Qwen scored 0.045/0.042, Chroma 0.068/0.065, and SDXL 0.099/0.088; all SSIM values
+were 0.962-0.965. Visual review found model-specific gradient, object, or sun-halo
+artifacts in every output. One carrier and two seeds do not justify a new
+content-adaptive route. Together with the text-only and face-plus-text replication,
+these extensions cover all four primary face/text quadrants at the shipped Google
+floors, while also showing why `auto` is a relative default rather than a universal
+per-image optimum.
+
 `requested_steps` is shared by SDXL and Chroma in `two_stage_pipeline.py`,
 which also owns `_target_size` with each profile's grid passed explicitly.
 The step helper exists because the two runtimes truncate differently. DiffSynth sets
@@ -2041,6 +2296,17 @@ resolution case -- every verdict so far comes from one fixture and one seed.
 Both ends of the shipped curve now have oracle verdicts, and the shipped curve clears
 everything it has been tested at:

+A source-fresh GPT Image 2.5 check on 2026-09-15 exercised the complete shipped
+profile rather than only a strength sweep. One 1024 x 1024 API generation from
+Flare and one from Sunburst both validated as current OpenAI C2PA, routed to the
+flat OpenAI strength, and remained OpenAI SynthID-positive after pixel-identical
+metadata removal. The complete `qwen-zimage` profile at 0.07675, seed 0, cleared
+both in the official OpenAI Content Provenance API. The Flare output measured
+37.48 dB PSNR / 0.9329 SSIM and the Sunburst output 37.01 dB / 0.9179 against
+their sources. Those fidelity numbers describe two neutral studio-photo
+carriers, not a content-wide quality estimate, and the source-positive/output-negative
+oracle pairs do not replace the wider OpenAI calibration that set the floor.
+
 | oracle | fixture size | detected at | clean from |
 |---|---|---|---|
 | openai.com/verify | 1.57 MP | 0.06 | 0.08 |
@@ -2078,7 +2344,7 @@ spread): 0.06 + (0.0525 - 0.015) = 0.0975, rounded up to **0.1**. Shipped as
 `QWEN_ZIMAGE_META_STRENGTH`; `--vendor meta` / `InvisibleOptions.vendor` names
 the cohort explicitly, implying the scrub runs (naming the cohort asserts the
 watermark is present). sdxl-zimage has no measured Meta rung and an
-explicit meta vendor there falls to the conservative unknown 0.25. The default
+explicit meta vendor there falls to the conservative unknown 0.50. The default
 resolution-adaptive curve (~0.1305 at 2.56 MP) also clears every measured
 source, so default behavior needed no change. Oracle verdicts carry a
 generation ID and creation timestamp embedded in the watermark payload; both
@@ -42,6 +42,112 @@ The old Google SynthID Detector portal does not work. On 2026-09-07, the
 official `labs.google/synthid` entry point redirected a signed-in browser to a
 Google 404. Historical corpus rows may still name `synthid-portal`; that value
 describes how an old result was obtained and is not an operational surface.
+Google's newer Search verification entry point did not provide an independent
+oracle allowance in the session tested on 2026-09-13. Lens accepted the
+metadata-stripped, previously Gemini-positive image
+`sha256:c86feaf788df2c423eeb53fcaee52b74f12a8e83829f94414e6dec2966e6e62e`,
+but ordinary Lens results only described and matched the visible content. An
+explicit `Is this made with AI? Check whether it contains SynthID.` request in
+Search AI Mode invoked the technical verifier and returned `The verification
+tool's daily limit has been exceeded.` after that session's Gemini image
+verification allowance was already exhausted. Record this as a quota refusal,
+not a watermark verdict and not a second quota pool.
+
+A second Search check on 2026-09-15 used a real Qwen 0.15 no-face/no-text
+variant rather than a control. Lens and Search AI Mode accepted the upload, but
+both answered from visible-content reasoning and Web matches instead of
+returning a SynthID result. The exact Google-AI verification question even
+misattributed the file to an Imagen 4 Fast demonstration and invented a matching
+prompt. Search is therefore neither a usable independent oracle nor a substitute
+verdict when its response does not explicitly report the verification tool's
+SynthID outcome.
+
+Fresh independent-session rechecks on 2026-09-14 split: one returned `detected`
+for the same control while another returned `unable to process your request due
+to a tool error or quota limit`. A later session returned five usable real-variant
+verdicts and refused the sixth, while the other tested sessions still refused real
+variants. Recovery is therefore not a global calendar event. Google [documents approximately ten
+image checks in a rolling 24-hour window](https://support.google.com/gemini/answer/16722517?hl=en-GB),
+but that approximation is orientation rather than a per-session reset schedule.
+When a campaign is conserving that allowance, probe each explicitly selected
+session with its next real pending variant and record a refusal without assigning
+a watermark verdict. Reserve a positive-control probe for surface diagnosis, not
+ordinary batch allocation.
+
+A second independent-session sweep used one real pending variant per session and
+again obtained no watermark verdict: three responses explicitly reported a temporary
+verification quota, while one reported a verification tool-call error. Four completed immutable result batches preserve
+the exact uploads and settled responses. Twenty-four newly generated
+no-face/no-text scene variants form six four-rung ladders; the six lowest rungs
+were prepared in immutable batches, and none had been submitted at that point.
+The later completed sweep followed the planned lowest-rung-first escalation. Do
+not treat batch preparation or visual quality scores as SynthID results.
+
+A session recovered earlier than the estimate derived from the oldest usable call
+preserved locally. Its next real pending
+variant, Chroma 0.16 on the face carrier, returned `detected`; the immutable batch
+is complete and hash-verified. This demonstrates why estimated rolling-window
+times are retry guidance rather than reset timestamps, and closes the tested
+face bracket at `0.16 detected -> 0.22 not_detected` without spending a duplicate
+request on the already settled 0.22 file.
+
+The same recovered session returned `not_detected` for Chroma 0.40 on the Cyrillic
+carrier, then explicitly quota-refused SDXL CJK 0.35. The other selected sessions
+were checked with real pending CJK variants, not controls, and returned quota or
+tool limitations. One settled response exposed a request-specific instruction to
+wait 18 hours. Treat that as the retry point for that session and request class, not as a replacement
+for Google's documented approximate rolling 24-hour allowance or a universal
+reset time. All four sessions were unavailable again at their latest checks.
+Because one response did not distinguish service failure from quota, its real
+Chroma CJK 0.40 candidate was retried. This time the
+verifier explicitly named a temporary image-analysis quota. The refusal was
+recorded and the same bytes were prepared in a new immutable retry batch; no
+watermark verdict was inferred. One session was probed once more with a
+real Qwen 0.15 no-face/no-text candidate. It returned a verification-tool
+technical error rather than a SynthID result. That failure was recorded as
+`refused`, the batch was hash-verified complete, and a fresh immutable retry
+batch was prepared for the evidence-based session window.
+
+All four sessions returned usable results again on 2026-09-15, earlier than the
+previous retry guidance. The remaining real
+variants were submitted adaptively without controls. SDXL CJK remained detected
+at 0.35, 0.40, and 0.45 and cleared at 0.50. Chroma CJK remained detected at
+0.40 and cleared at 0.50. The two no-face/no-text carriers cleared Qwen at 0.15;
+Chroma cleared the flat carrier at 0.12 and the photographic carrier at 0.30
+after detections at 0.12 and 0.20; SDXL cleared the photographic carrier at 0.15
+and the quality-qualified flat carrier at 0.35. One quota refusal and one tool-call
+failure encountered while routing SDXL CJK 0.40 were kept as separate completed
+batches; another session then returned the usable `detected` result for the same
+upload hash. Every settled batch passed immutable-source, upload, manifest, and
+result verification. The observed early recovery reinforces that a retry time is
+guidance for the next real request, not a promised reset.
+
+A same-day production-boundary replication then used three new independent
+Gemini-generated controls: busy CJK plus many faces, flat CJK without faces, and a
+portrait plus dense CJK. Only processed variants were submitted. Across Qwen 0.35,
+Chroma 0.50, and SDXL 0.50, all nine profile-by-carrier cells returned
+`not_detected`. The night-market lower rungs Chroma 0.40 and SDXL 0.45 also returned
+`not_detected`; they do not supersede the earlier harder-source detections at those
+rungs. Two sessions returned explicit quota failures during this pass, one
+candidate retry also hit quota, and a byte-normalized duplicate after the usable
+SDXL result returned a tool-call error. Those four calls were recorded as
+`unreachable`, not spliced into the nine usable verdicts. All result batches were
+hash-verified.
+Google did not expose a reset timestamp: the returned quota text supports only
+waiting and retrying, not a universal 24-hour claim.
+
+The documented Vertex AI Media Studio verification flow was also unavailable
+in the live console on 2026-09-14. Its old console URL redirected to Agent Media
+Studio after the Vertex-to-Agent Platform rename. The current Image playground
+was inspected in eligible configurations; its task menu exposed generation and image
+editing but no watermark upload or validation action. The deprecated Python SDK
+can still resolve `imageverification@001` to its publisher-model endpoint. A
+2026-09-15 request reached `predict`, but the real Qwen 0.15 variant returned
+HTTP 403 `The caller does not have permission` in both tested configurations. The current
+Python reference still documents `WatermarkVerificationModel.verify_image`, but
+the backing model was access-restricted in both tested configurations and the SDK path is
+past its announced removal date. Recheck the live console and a real pending
+variant before adding Search or Agent Media Studio to the surface table.
 Adobe Inspect and local C2PA, DWT-DCT, TrustMark, and research detectors are
 useful evidence readers, not interchangeable provider pixel-watermark oracles.

@@ -306,6 +412,17 @@ response settles as `indeterminate`, not clean. Meta's live positive wording
 changed from `AI signatures from Meta were found` to `AI signatures were found`
 by 2026-09-07; both exact forms remain accepted.

+Isolation is a transport property, not a verdict requirement. On 2026-09-15,
+the isolated OpenAI runner returned one generic error and two Cloudflare
+refusals for a prepared three-file batch. Uploading those exact prepared hashes
+through an authenticated first-party browser session reached the official verifier
+and returned explicit SynthID results. A later Chroma 0.12 upload remained in
+`Verifying` for more than 90 seconds and was recorded as indeterminate rather
+than clean. The official API then returned HTTP 200 and SynthID `not_detected`
+for the same processed source hash. Keep each route in its own immutable batch;
+do not splice an isolated-runner failure and a first-party browser or API verdict into
+one response.
+
 ## OpenAI API

 The one automated path calls the hardened metadata-independent adapter already
@@ -332,6 +449,20 @@ OpenAI documents `not_detected` as absence of a supported signal, not proof that
 content was not AI-generated. The API must not become an adaptive removal loss
 or automatic search loop. See [`synthid.md`](synthid.md) for the oracle boundary.

+The fresh no-face/no-text calibration on 2026-09-15 used three independent
+official API requests after the public Web surface stalled. All three returned HTTP 200
+with C2PA `not_present`: Chroma 0.12 and SDXL 0.06 were SynthID `not_detected`,
+while SDXL 0.04 was `detected`. Together with settled Web results for the same
+carrier, the tested 0.02-grid brackets are Qwen `0.04 detected -> 0.06
+not_detected`, Chroma `0.10 detected -> 0.12 not_detected`, and SDXL `0.04
+detected -> 0.06 not_detected`. The generated source control was not uploaded.
+
+A separate face-plus-multilingual-text carrier then returned Qwen `0.04
+detected -> 0.06 not_detected`, Chroma 0.10 `not_detected`, and SDXL 0.04
+`not_detected`. The lower Chroma and SDXL results did not change the global
+candidate because the no-face/no-text carrier remained the limiting OpenAI
+stratum.
+
 ## Microsoft API

 The direct Microsoft path uses the Azure Content Provenance Detection preview
@@ -376,8 +507,8 @@ error rate estimates:

 | Surface | Route and repetitions | Recorded result |
 | --- | --- | --- |
-| Gemini Web | User's real Chrome, three authenticated account slots | 3/3 `not_detected`; Gemini reported no reliable Google AI signal. |
-| OpenAI API | Three named API-key slots, direct transport | 3/3 SynthID `not_detected`; C2PA `not_present`; all responses were HTTP 200. |
+| Gemini Web | Official browser verifier, three independent authenticated sessions | 3/3 `not_detected`; Gemini reported no reliable Google AI signal. |
+| OpenAI API | Official API, three independent requests | 3/3 SynthID `not_detected`; C2PA `not_present`; all responses were HTTP 200. |
 | OpenAI Web | ThorData GB, isolated Playwright | `indeterminate`; after accepting the prepared upload, the provider returned `Something went wrong. Please try again.` |
 | Microsoft API | Direct API, private hash-matched Azure blob | HTTP 202, operation `Succeeded`; Watermark `not_detected`, C2PA `absent`, provider outcome `NoProvenanceDetected`. |
 | Microsoft Web | ThorData US, isolated Playwright | `not_detected`; the page's exact settled label was `Inconclusive`. |
@@ -6,12 +6,29 @@ and pipeline modules are intended for maintainers and specialized workflows.
 Dependency groups are identical for the CLI and Python API. The default install
 covers metadata extraction, normalization, verdict logic, and stripping.
 Array/pixel APIs use `pixels`; visible removal uses `visible`; DWT-DCT detection
-uses `detect`; pixel photo classification uses `classify`; invisible image removal uses `qwen-zimage` and an NVIDIA GPU; and
+uses `detect`; pixel photo classification uses `classify`; lightweight source-pipeline classification uses `source-classify`; invisible image removal uses `qwen-zimage` and an NVIDIA GPU; and
 visible video processing uses `video`. The video-pixel SynthID-removal profile is
 a separate VAE path that still runs on CPU and combines `video` and `diffusion`. Add `heif`
 independently when path-based pixel APIs must decode HEIC, HEIF, or AVIF. See
 the complete [feature-extra matrix](installation.md#feature-extras).

+## Classify a source/export pipeline
+
+Install `remove-ai-watermarks[source-classify]`, then call the explicit API:
+
+```python
+import remove_ai_watermarks as raiw
+
+result = raiw.classify_source("metadata-stripped.png")
+print(result.label)   # "openai" | "google" | "unknown"
+print(result.reason)  # "classified" | "abstained" | "feature_unavailable"
+```
+
+This signal classifies a complete image-generation and export pipeline. It is
+not a SynthID detector and does not establish watermark presence or absence.
+See [source-pipeline classification](source-classify.md) for the result fields,
+model pin, evaluation boundary, and offline weights override.
+
 ## Remove visible marks

 Install `remove-ai-watermarks[visible]` before using the visible-removal API.
@@ -153,6 +170,32 @@ engine's own parameter names and defaults. `force`, which decides whether the
 engine runs at all, is a parameter of `remove_all` and `remove_batch` alongside
 `backend` and `sensitivity`.

+Specialized callers can inspect or apply the clean-room local Paint pixel
+operation directly. This is deliberately separate from the high-level diffusion
+route. Passing the UUID signed in pristine C2PA makes agreement with that UUID a
+write precondition:
+
+```python
+from remove_ai_watermarks.microsoft_invismark import (
+    detect_local_invismark,
+    disrupt_local_invismark,
+    invismark_id_from_c2pa,
+)
+
+expected = invismark_id_from_c2pa("paint.png")
+detection = detect_local_invismark("paint.png")
+if expected is not None and detection is not None:
+    result = disrupt_local_invismark(
+        "paint.png",
+        "clean.png",
+        expected_watermark_id=expected,
+    )
+```
+
+The direct call needs the `pixels` extra. It writes atomically only after the
+encoded output no longer decodes as a valid local payload. The result does not
+claim acceptance by a proprietary Microsoft detector.
+
 The complete field set, with the shipped defaults:

 | Field | Default | Effect |
@@ -739,8 +782,8 @@ engine = InvisibleEngine(pipeline="sdxl-zimage")

 The `qwen-zimage` extra is required for every profile, including `auto`: each
 concrete engine runs the same DiffSynth Z-Image face stage. `pipeline="auto"`
-selects chroma-zimage for Microsoft provenance and qwen-zimage
-otherwise, after the vendor is known and before strength resolution.
+selects sdxl-zimage for Google provenance, chroma-zimage for Microsoft, and
+qwen-zimage otherwise, after the vendor is known and before strength resolution.

 The opt-in verified-text stage uses the same `text_manifest` argument as the CLI:

@@ -754,10 +797,11 @@ engine.remove_watermark(

 Install `remove-ai-watermarks[text-restoration]`. The manifest schema and safety
 constraints are documented in the CLI guide. The engine verifies its decoded RGB
-hash before loading the diffusion models. Qwen and Chroma reconstruct the donor with
-the VAE already loaded for the one profile selected by `auto`; no second generative
-profile runs. The engine rejects SDXL, downscaling, and postprocessing combinations
-that were not evaluated. Tiling is also rejected because the combined
+hash before loading the diffusion models. Qwen, SDXL, and Chroma reconstruct the
+donor with the VAE already loaded for the selected profile; no second generative
+profile runs. The engine turns an unset `adaptive_polish` off in verified-text mode
+and rejects explicit postprocessing or downscaling combinations that were not
+evaluated. Tiling is also rejected because the combined
 tiled-restoration path has no provider-oracle calibration. `InvisibleOptions` exposes
 the same field for `remove_all`; after a visible-stage edit, the manifest must be built
 against the staged pixels rather than the pristine source.
@@ -92,7 +92,17 @@ alone does not upload new heads -- bump the pinned Hub revision when a new
 head must reach installed clients. Pin a Hub revision in production so a
 later freeze cannot silently change installed clients.

-The workflow can also be started manually with an optional version input.
+The separate `wiltodelta/raiw-source-classify` model is published with
+`.github/workflows/publish-source-classify-hf.yml`. The workflow downloads the
+derivative `source-pipeline-mlp.npz` artifact from the declared GitHub model
+freeze, verifies its SHA-256, and uploads only that file plus the tracked card
+and aggregate metrics under `docs/source-classify-hf/`. The private corpus,
+catalog, paths, hashes, embeddings, and feature caches are outside this flow.
+
+The workflow is manual and requires the model-freeze release tag plus an
+explicit `publish` confirmation. Keep model-freeze GitHub releases marked as
+pre-releases so a model artifact cannot replace the latest package release in
+GitHub's release UI.

 If a distribution job fails because a repository or Hugging Face credential is
 invalid, rotate the corresponding GitHub secret and rerun the failed job. A
@@ -255,13 +265,15 @@ lists it: uv serves its cached pre-release view of the index. Re-run with a
 fresh cache (`UV_CACHE_DIR=$(mktemp -d)`) before treating it as a release
 failure.

-The ComfyUI sync run can fail the same way on its own: it resolves the version
-from the PyPI JSON API (which updates first) but installs the test dependency
-with pip against the simple index, whose CDN edges lag by minutes. The
-`distribute.yml` waiter also has a clock: 0.40.0's node suite finished after
-the 15-minute poll, so the job now waits 30 minutes and then accepts a
-registry row that already requires this library version. Rerun the failed
-`comfyui` job only when the registry still has no such row.
+The ComfyUI sync resolves the version from the PyPI JSON API but installs its
+test dependency through pip's simple index, whose CDN edges can lag by
+minutes. `distribute.yml` gates every downstream job on both APIs listing the
+release, so it cannot dispatch that race again. The ComfyUI waiter itself has
+a clock: 0.40.0's node suite finished after the 15-minute poll, so it now
+waits 30 minutes and then accepts a registry row that already requires this
+library version. For an older distribution that started before the simple
+index caught up, rerun its failed `comfyui` job only when the registry still
+has no such row.

 The automatic `verify-release.yml` run after `distribute.yml` checks the
 Hugging Face Space pin. When `RAIW_HF_SPACE_TOKEN` is set, that pin lands
@@ -0,0 +1,65 @@
+---
+license: apache-2.0
+library_name: numpy
+pipeline_tag: image-classification
+tags:
+  - image-forensics
+  - source-attribution
+  - open-set-classification
+---
+
+# RAIW source-pipeline classifier
+
+This is a compact, abstaining pixel-forensics classifier for three labels:
+`openai`, `google`, and `unknown`. It is intended for original-export-looking
+images whose metadata may have been removed without changing the pixels.
+
+This model is **not a SynthID detector**. It does not decode or verify a
+watermark payload. Its labels describe complete generation and export
+pipelines, so renderer, geometry, encoding, post-processing, and watermark
+state may be confounded. `unknown` does not mean that SynthID is absent, and a
+provider label does not prove that SynthID is present.
+
+## Model
+
+- Input: a 124-dimensional residual/spectral forensic feature vector plus a
+  768-dimensional phase-free spectral vector.
+- Head: 892 -> 128 -> 64 -> 3 MLP with GELU activations.
+- Output: `openai`, `google`, or abstain as `unknown` using independent
+  provider margins.
+- Artifact: `source-pipeline-mlp.npz`
+- SHA-256: `6430b0738869ff07b6845b7c6149c2a83bc3f3785878082886c241c6f7ca0c4a`
+
+The artifact contains only learned parameters, normalization constants,
+labels, and thresholds. It does not contain training images, paths, image
+hashes, embeddings, or a training catalog.
+
+## Evaluation
+
+The operating point was selected on an internal calibration split. The locked
+internal test and the public `bigdatamark/synthid-research` dataset were not
+used for architecture, checkpoint, or threshold selection.
+
+| Evaluation | Class | Recall | Precision |
+|---|---|---:|---:|
+| Locked internal test (5,369 images) | OpenAI | 60.4% | 84.9% |
+| Locked internal test (5,369 images) | Google | 57.6% | 87.9% |
+| Locked internal test (5,369 images) | Unknown | 98.6% | 93.7% |
+| Public external originals (900 images) | OpenAI | 68.7% | 100.0% |
+| Public external originals (900 images) | Google | 81.3% | 98.8% |
+| Public external originals (900 images) | Unknown/Flux | 99.0% | 66.4% |
+
+On the public external set, a 95% resize reduced Google recall to 1.0%, and a
+JPEG quality-75 round trip caused the model to abstain on every image. These
+are expected abstentions, not evidence that a watermark was removed.
+
+## Intended use
+
+Use this model as one probabilistic signal in a broader forensic report or to
+suggest an appropriate processing path. Do not use it to claim watermark
+presence, authorship, policy compliance, fraud, or provenance. Applications
+should expose the abstention and the model's narrow source-pipeline claim.
+
+Feature extraction and inference are versioned in the
+`remove-ai-watermarks` Python project rather than in this model repository.
+Use its explicit `classify_source` API; `identify` does not run this model.
@@ -0,0 +1,22 @@
+{
+  "claim": "abstaining OpenAI/Google source-pipeline classification; not SynthID detection",
+  "artifact_sha256": "6430b0738869ff07b6845b7c6149c2a83bc3f3785878082886c241c6f7ca0c4a",
+  "classes": ["openai", "google", "unknown"],
+  "locked_internal_test": {
+    "n": 5369,
+    "confusion_actual_by_predicted": [[230, 6, 145], [0, 217, 160], [41, 24, 4546]],
+    "recall": {"openai": 0.6036745407, "google": 0.5755968170, "unknown": 0.9859032748},
+    "precision": {"openai": 0.8487084871, "google": 0.8785425101, "unknown": 0.9371263657}
+  },
+  "public_external_originals": {
+    "dataset": "bigdatamark/synthid-research",
+    "n": 900,
+    "confusion_actual_by_predicted": [[206, 0, 94], [0, 244, 56], [0, 3, 297]],
+    "recall": {"openai": 0.6866666667, "google": 0.8133333333, "unknown": 0.99},
+    "precision": {"openai": 1.0, "google": 0.9878542510, "unknown": 0.6644295302}
+  },
+  "selection": {
+    "minimum_calibration_precision": {"openai": 0.99, "google": 0.95},
+    "maximum_calibration_unknown_false_positive_rate": 0.01
+  }
+}
@@ -0,0 +1,42 @@
+# Source-pipeline classification
+
+`classify_source` is an opt-in, metadata-free pixel classifier for images that
+still resemble an original provider export. It returns `openai`, `google`, or
+an explicit `unknown` abstention.
+
+This is not a SynthID detector. It does not decode or verify a watermark
+payload, and it must not be used to claim that SynthID is present or absent.
+The model describes the complete source/export pipeline, where renderer,
+geometry, encoding, post-processing, and watermark state can be confounded.
+
+Install the lightweight runtime:
+
+```bash
+uv add "remove-ai-watermarks[source-classify]"
+```
+
+Call the Python API explicitly:
+
+```python
+import remove_ai_watermarks as raiw
+
+result = raiw.classify_source("metadata-stripped.png")
+print(result.label)       # "openai" | "google" | "unknown"
+print(result.reason)      # "classified" | "abstained" | "feature_unavailable"
+print(result.margins)     # provider margins before thresholding
+print(result.thresholds)  # frozen provider thresholds
+```
+
+`feature_unavailable` is returned for images below the required 256x256
+geometry or for a degenerate feature vector. `abstained` means neither provider
+cleared its independent calibrated margin. Both are honest `unknown` results.
+
+The first call downloads the exact pinned revision of
+[`wiltodelta/raiw-source-classify`](https://huggingface.co/wiltodelta/raiw-source-classify).
+Set `RAIW_SOURCE_CLASSIFY_WEIGHTS` to the NPZ itself or a directory containing
+`source-pipeline-mlp.npz` for an offline deployment.
+
+The model is useful as a routing suggestion for a processing workflow or as one
+probabilistic signal in a broader forensic report. Do not use it as evidence of
+authorship, fraud, policy compliance, provenance, or watermark presence. The
+published model card records the evaluation and robustness limits.
@@ -121,6 +121,10 @@ entries enter the watermark inventory, while a fingerprint remains a
 re-linkability warning, never selects pixel regeneration, and does not suppress
 independent SynthID evidence.

+The opt-in `classify_source` API is not an `identify` signal. It can suggest an
+OpenAI or Google source/export pipeline after metadata removal, but it neither
+detects nor decodes a watermark and never changes the provenance verdict.
+
 ## File and container formats

 Pixel based image commands discover these extensions:
@@ -148,16 +152,17 @@ use their format specific paths.

 ## Invisible watermarks

-The `invisible` command uses diffusion regeneration. It targets watermark
-patterns by changing the image rather than decoding and deleting a known
-payload.
+The `invisible` command uses diffusion regeneration. A specialized Python API can
+instead apply decoder-guided disruption to a positively identified local
+Microsoft Paint payload.

-Current pipeline values, all CUDA-only:
+Current diffusion pipeline values, all CUDA-only:

 - `qwen-zimage`, the default;
 - `sdxl-zimage`, the same recipe and the same face stage on an SDXL global pass;
 - `chroma-zimage`, the same face stage on a Chroma1 global pass;
- `auto`, chroma-zimage for Microsoft provenance, otherwise qwen-zimage.
+- `auto`, sdxl-zimage for Google provenance, chroma-zimage for Microsoft,
+  otherwise qwen-zimage.

 The `controlnet`, `sdxl`, `qwen` and `default` values were removed. A retired name
 is rejected at parse time rather than remapped onto a surviving profile.
@@ -170,7 +175,11 @@ and `scripts/synthid_runtime/`.
 The tool recognizes presence from supported provenance: Google AI C2PA
 under Google's all-media watermark policy, and current OpenAI C2PA carrying an
 explicit `c2pa.watermarked.*` action. Legacy OpenAI C2PA without that action
-does not assert SynthID.
+does not assert SynthID. GPT Image 2.5 Flare and Sunburst were verified on
+2026-09-15: both API outputs carried valid OpenAI C2PA with
+`c2pa.watermarked.unbound` and exercised the existing model-independent OpenAI
+route. Automatic routing selected `qwen-zimage` with the OpenAI strength cohort,
+and both outputs cleared the official OpenAI SynthID check.

 Provider verifiers and provenance APIs are reserved for development-time
 calibration. They are not exposed through the installed CLI, the top-level
@@ -180,19 +189,24 @@ Microsoft Paint can name `com.microsoft.invismark.1` in a C2PA soft-binding
 assertion. Inspection reports both that exact algorithm and its signed `value`,
 which Paint uses as the identifier carried by the pixel watermark, and emits an
 additive `invismark` signal so callers can select pixel removal without parsing
-the generic `soft_binding` detail. Photos uses a parallel local writer path.
-Metadata stripping removes only the embedded manifest; `invisible` and `all`
-guarantee the supported InvisMark removal contract by regenerating the pixel
-layer as well. The project has no validated local InvisMark decoder, so local
-inspection cannot independently verify the output. Microsoft's official
-[Content Provenance Detection API](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/how-to/how-to-provenance-detection)
-is the external oracle: it reports pixel `Watermark` and embedded `C2PA` results
-separately; a control-positive, output-negative pair is the available per-file
-verification path. The API needs Azure credentials; the page a human can check
-without an account is <https://ai.azure.com/nextgen/validate>. That page is the
-public provider oracle used for the measured removal certification. Its clean
-outcome is rendered as `Inconclusive`; the authenticated API exposes separate
-watermark and C2PA fields but is not a different class of certification.
+the generic `soft_binding` detail.
+
+For Paint's local `Watermarker.dll` format, the reader recovers the repeated
+144-bit message and requires its prefix, GUID byte order, checksum, and per-bit
+support. A caller can additionally require agreement with the UUID signed in
+pristine C2PA. Only a positive local payload uses decoder-guided bit disruption;
+the encoded output is read back and must no longer contain a valid local payload
+before it is published. A Microsoft
+soft-binding declaration without a matching local payload, including the
+committed cloud Image Creator fixture, is rejected by this decoder. The CLI and
+high-level API retain their Chroma diffusion route. Metadata stripping alone still
+removes only the manifest.
+
+This is local payload invalidation, not external removal certification. The
+available Azure Content Provenance Detection API did not detect a known-positive
+local-format control, so it cannot certify this direct output. Photos and future
+Paint builds may use another pixel format and therefore remain on the regeneration
+fallback unless this decoder positively identifies them.

 Meta Muse Image stamps every output with Content Seal, a proprietary invisible
 pixel watermark, and ships no visible mark (the legacy `Imagined with AI`
@@ -243,7 +257,8 @@ not a universal clean verdict.
 | Google Veo video | Veo diamond and legacy text | Oracle-certified VAE removal for SynthID | C2PA and related source signals |
 | OpenAI image generators | None registered | Diffusion regeneration for supported invisible signals | C2PA and generator provenance |
 | Meta Muse Image | None on Muse output (legacy `Imagined with AI` unregistered) | Diffusion regeneration for Content Seal, oracle-verified on the default profile | XMP IPTC `trainedAlgorithmicMedia` companion tag; no local Content Seal decoder |
-| Microsoft Paint and Photos | None registered | External Microsoft oracle for InvisMark; no validated local decoder | Paint C2PA soft-binding algorithm and identifier |
+| Microsoft Paint local `Watermarker.dll` format | None registered | Local repeated-payload decoder and decoder-guided disruption; no usable external oracle for this format | Paint C2PA soft-binding algorithm and matching UUID |
+| Microsoft cloud and other InvisMark variants | None registered | Chroma regeneration fallback | Microsoft C2PA attribution and soft-binding declaration |
 | Microsoft image outputs (measured variant) | One top-right white AI-badge variant | No registered pixel decoder | C2PA attribution |
 | Stable Diffusion and SDXL | None registered | Diffusion regeneration; optional open decoder | Embedded parameters and text metadata |
 | FLUX | None registered | Diffusion regeneration; optional open decoder | C2PA for supported sources |
@@ -23,6 +23,130 @@ are explicit negatives, a strict rule can abstain, and any watermark claim uses
 an independent oracle. CLIP content embeddings and the 124-d origin-locked
 residual bank are different features for different jobs.

+### Phase-free spectral pipeline probe, 2026-09-15
+
+Mark's public
+[`bigdatamark/synthid-research`](https://huggingface.co/datasets/bigdatamark/synthid-research)
+corpus pairs 300 prompts across Gemini 2048-pixel JPEG, OpenAI 1254-pixel PNG,
+and FLUX 1280-pixel PNG exports. It is useful for testing source-pipeline
+features, but it is not a causal SynthID corpus: renderer, dimensions, encoder,
+and watermark state all change together.
+
+The paper-inspired probe in [`scripts/spectral_pipeline_probe.py`](../scripts/spectral_pipeline_probe.py)
+discards Fourier phase, subtracts each image's radial log-magnitude envelope,
+pools luma plus two opponent-color spectra, and fits a dual ridge classifier.
+Matching prompt ids stay together in a grouped five-fold evaluation. The
+256-pixel / 32x32-grid / ridge-1000 configuration was selected after an
+exploratory comparison on these same folds, so this is a development result,
+not blind validation. Its complete-corpus confusion matrix was:
+
+| Actual pipeline | Gemini | OpenAI | FLUX |
+| --- | ---: | ---: | ---: |
+| Gemini | 220 | 25 | 55 |
+| OpenAI | 33 | 219 | 48 |
+| FLUX | 75 | 48 | 177 |
+
+Accuracy was 616/900 (68.4%), against 33.3% chance. Applying each transform
+only to held-out rows gave 67.3% after a two-pixel top-left crop, 67.8% after a
+95% resize, 67.9% after JPEG quality 75, and 68.4% after Gaussian blur radius
+0.7. This is materially more transform-stable than the origin-locked lattice,
+but it remains a source/export classifier. There are no photographs, foreign
+generators beyond FLUX, independently generated 1K Gemini files, or same-image
+SynthID on/off pairs in the corpus. No runtime detector, threshold, or product
+verdict is justified by this result.
+
+The representation follows the frequency-domain premise of Zhang et al.,
+[*Detecting and Simulating Artifacts in GAN Fake Images*](https://arxiv.org/abs/1907.06515),
+not their ResNet-34 implementation. Their own post-processing results also
+support treating augmentation and architecture shift as first-class tests,
+rather than reading an untransformed source-classification score as watermark
+evidence.
+
+### Abstaining fused source classifier, 2026-09-15
+
+The follow-up experiment combined the existing 124-d forensic descriptor with
+the 16x16 phase-free spectral grid in a class-balanced ridge head. Its three
+outputs are `openai`, `google`, and `unknown`; a provider result is emitted only
+when it beats both the other provider and `unknown` by a margin selected on a
+separate calibration split. The public implementation is
+[`scripts/source_pipeline_classifier.py`](../scripts/source_pipeline_classifier.py).
+It includes training, strict pickle-free model loading, and JSON inference, but
+no fitted weights: the experimental training freeze contains inputs that are
+not redistributable. This keeps the code reproducible without publishing a
+derivative artifact whose data provenance does not permit redistribution.
+
+Penalty 100 and margin 0.6392407698660646 were selected using only the training
+and calibration partitions. The 2,669-image locked internal test and Mark's
+900-image public set were not used for feature, penalty, or threshold selection.
+Feature-extraction failures count as `unknown` in every result below.
+
+| Evaluation | Accuracy | OpenAI recall | Google recall | Unknown recall |
+| --- | ---: | ---: | ---: | ---: |
+| Locked internal, 2,669 images | 84.2% | 60.9% | 48.8% | 95.8% |
+| Mark original exports, 900 images | 90.1% | 76.3% | 94.3% | 99.7% |
+| Mark crop 2 px | 89.8% | 76.7% | 94.3% | 98.3% |
+| Mark resize 95% | 59.7% | 81.7% | 0.0% | 97.3% |
+| Mark JPEG quality 75 | 33.3% | 0.0% | 0.0% | 100.0% |
+| Mark blur radius 0.7 | 68.3% | 42.3% | 63.0% | 99.7% |
+
+The untouched-export result is useful source-attribution evidence, but the
+transforms decide the claim. Resizing eliminated the Google branch, JPEG made
+the classifier abstain on every row, and blur materially reduced both provider
+recalls. The locked unknown false-attribution rate was 4.2%, versus 0.9% on the
+calibration split. This is therefore an inspectable research baseline, not a
+runtime detector or a product watermark verdict.
+
+Training requires hash-disjoint `train` and `calibration` directories, each
+with `openai`, `google`, and `unknown` subdirectories:
+
+```bash
+uv run python scripts/source_pipeline_classifier.py train \
+  /path/to/train /path/to/calibration --model-out .local-eval/source-pipeline.npz
+
+uv run python scripts/source_pipeline_classifier.py predict \
+  .local-eval/source-pipeline.npz image.png
+```
+
+A stronger public release needs a publication-cleared multi-era training set
+and a causal marked/unmarked contrast. Mark's same-prompt triples remain an
+external source/export test, not evidence that the fitted features read the
+SynthID payload.
+
+### Full-corpus source-pipeline model, 2026-09-16
+
+The next experiment trained a compact nonlinear head on the complete available
+private corpus, including target-provider exports, photographs, and foreign
+generator negatives. The corpus itself is not published. The released
+`source-pipeline-mlp.npz` is a derivative artifact containing only learned
+parameters, normalization constants, labels, and abstention thresholds; it
+contains no images, paths, image hashes, embeddings, or training catalog.
+
+The model joins the 124-dimensional forensic descriptor with the
+768-dimensional phase-free spectral descriptor and applies a
+892 -> 128 -> 64 -> 3 GELU MLP. Checkpoint and thresholds were selected on a
+separate calibration partition. The 5,369-image hash-disjoint locked test and
+Mark's public 900-image originals were not used for architecture, checkpoint,
+or threshold selection.
+
+| Evaluation | Class | Recall | Precision |
+| --- | --- | ---: | ---: |
+| Locked internal test (5,369 images) | OpenAI | 60.4% | 84.9% |
+| Locked internal test (5,369 images) | Google | 57.6% | 87.9% |
+| Locked internal test (5,369 images) | Unknown | 98.6% | 93.7% |
+| Mark public originals (900 images) | OpenAI | 68.7% | 100.0% |
+| Mark public originals (900 images) | Google | 81.3% | 98.8% |
+| Mark public originals (900 images) | Unknown/FLUX | 99.0% | 66.4% |
+
+A 95% resize reduced Google recall on the public set to 1%, and a JPEG
+quality-75 round trip made every row abstain. This behavior matches the narrow
+claim: the model classifies original-looking source/export pipelines after
+metadata removal. It neither decodes nor detects SynthID, and its `unknown`
+label does not mean that SynthID is absent. The model card and aggregate
+metrics are tracked under
+[`docs/source-classify-hf/`](source-classify-hf/README.md). Runtime integration
+is provided by the explicit `classify_source` library API and never feeds the
+`identify` watermark verdict.
+
 Krawetz's Gemini-chat TPR critique is a verifier-quality claim, not a
 feature we can ship. [Lead Stories, 2026-07](https://leadstories.com/analysis/2026/07/google-gemini-synthid-detector-confuses-results-within-same-chat.html)
 documented Gemini repeating the first file's SynthID verdict inside a
@@ -95,10 +95,13 @@ Two real but non-local verifier paths surfaced. The historical Vertex SDK
 exposed `WatermarkVerificationModel.from_pretrained("imageverification@001")`.
 The class still exists in `google-cloud-aiplatform` 1.165.1, but warns that it
 was deprecated 2025-06-24 with removal scheduled for 2026-06-24. Calls in two
-projects with Vertex enabled now reject the publisher model id as invalid. A
-repository using `response.watermark_detected` is also stale: the actual SDK
-property was `watermark_verification_result`, backed by the prediction's
-`decision` field.
+eligible Vertex configurations were rechecked on 2026-09-15 with a real processed
+variant and correctly attributed user credentials. The publisher
+model resolved, but both `predict` calls returned HTTP 403 `The caller does not
+have permission`. The surviving class reference therefore does not establish
+public access to the backing model. A repository using
+`response.watermark_detected` is also stale: the actual SDK property was
+`watermark_verification_result`, backed by the prediction's `decision` field.

 A newer unlisted upload route,
 `POST https://synthid.googleapis.com/upload/v1:verifyContent`, is live enough
@@ -775,6 +775,20 @@ SynthID. Current OpenAI manifests distinguish the watermarked output with the
 signed `c2pa.watermarked.unbound` action. The detector requires that action, so
 old OpenAI C2PA remains an origin signal without becoming a SynthID claim.

+**GPT Image 2.5 compatibility, verified 2026-09-15.** One fresh low-quality
+1024 x 1024 API generation from each released model, `gpt-image-2.5-flare` and
+`gpt-image-2.5-sunburst`, carried a valid bound OpenAI C2PA claim generated by
+`OpenAI Media Service API`, with `trainedAlgorithmicMedia` and the action set
+`created, converted, watermarked.unbound`. `identify` reported high-confidence
+OpenAI provenance plus SynthID, `has_invisible_target` returned true, and the
+automatic strength cohort resolved to OpenAI for both files. After C2PA was
+removed without changing decoded pixels, the official OpenAI Content Provenance
+API still returned SynthID `detected` for both sources. The complete shipped
+`qwen-zimage` profile at strength 0.07675 and seed 0 then produced two outputs
+for which the same metadata-independent API returned SynthID `not_detected`.
+This is one controlled source per model and proves the current routing and
+operating point on those two sources; it is not a provider-wide removal rate.
+
 ### 4.3 Other vendors

 - **Kakao** (South Korea): SynthID adopter as of May 2026 (Google announcement)
@@ -501,6 +501,12 @@ invisible marks, detection/removal parity, or historical crossfire.
 The scan runs files in parallel and checkpoints each completed row beside the
 requested CSV as `.progress.jsonl`; repeat the same command to resume an
 interrupted run, and use a distinct report path for every code/corpus snapshot.
+The visible counterpart, `scripts/visible_positives.py`, records both strict and
+metadata-corroborated registry results in a resumable JSONL. With `--sheets` it
+adds blinded, native-resolution top and bottom bands from quiet metadata cohorts
+and from an unbiased quiet sample. The latter keeps a provider with no mapped
+metadata from being excluded by construction; neither stratum is ground truth
+until the pixels are adjudicated.

 Run a fresh-delta discovery review weekly after the retained corpus is
 refreshed, using `corpus_gap_scan.py --since` with an overlap from the preceding
@@ -31,6 +31,15 @@ Exact claim-generator mappings also cover `Grok Imagine` as xAI Grok Imagine
 and `Suno` as Suno. Both can carry a valid AI source claim without a separately
 readable issuer, so issuer-only attribution leaves the platform blank.

+Some signers expose the product in `signature_info` while retaining a generic
+claim generator. xAI Grok Imagine is recognized from that signed identity even
+when the generator is only `c2pa-rs`; Producer.ai is handled the same way.
+FastVid has an exact claim-generator mapping. SPRING (SG) PTE. LTD. is retained
+as the signer on AI-source claims without guessing which of its products created
+the file. TikTok, ByteDance, CapCut, Anthropic Claude file provenance, and vivo
+Camera are neutral signer or capture attributions: without an AI source assertion,
+none of them changes the AI verdict.
+
 **OpenRouter raster metadata survey (2026-09-02):** a paid Recraft V3 API
 output carried a valid C2PA claim with `claim_generator` `recraft.ai`, a
 `c2pa.created` action, and `trainedAlgorithmicMedia`; that exact generator now
@@ -39,7 +48,7 @@ Riverflow V2 Fast API outputs carried no locally readable C2PA, XMP, EXIF, or
 container metadata. Their absence is evidence about those exact files and API
 routes, not a claim that every Krea or Riverflow export is metadata-free.

-**ByteDance Volcano Engine (Volcengine)** — the cloud behind Doubao / Jimeng — signs its AI image output with a cert from `certificate_center@volcengine.com` + `trainedAlgorithmicMedia` (issuer `b"volcengine"` → "ByteDance (Volcano Engine)", platform "ByteDance Volcano Engine"); note this is the C2PA-signed surface, distinct from the XMP/PNG TC260 `AIGC` label Doubao also uses. ByteDance's **international brand (BytePlus / Seedream / Seededit)** signs the same content as **"Byteplus Pte. Ltd."**. The bare `volcengine` needle missed it, so BytePlus output was mis-attributed to "Adobe Firefly" through an incidental "Adobe XMP" toolkit string. Issuer `b"Byteplus"` maps directly to "BytePlus (ByteDance)". ByteDance's consumer app **Dreamina** (the international Jimeng brand) signs as **"Bytedance Pte. Ltd."** with a `Dreamina/x.y` claim generator but, unlike the Volcano Engine surface, ships **no `trainedAlgorithmicMedia`**. Issuer `b"Dreamina"` maps to "ByteDance Dreamina" with **`asserts_ai=True`**. Registering the broader **issuer** `b"Bytedance Pte"` was deliberately avoided because that same entity also signs non-AI CapCut edits; keying on the `Dreamina` generator token is precise.
+**ByteDance Volcano Engine (Volcengine)** — the cloud behind Doubao / Jimeng — signs its AI image output with a cert from `certificate_center@volcengine.com` + `trainedAlgorithmicMedia` (issuer `b"volcengine"` → "ByteDance (Volcano Engine)", platform "ByteDance Volcano Engine"); note this is the C2PA-signed surface, distinct from the XMP/PNG TC260 `AIGC` label Doubao also uses. ByteDance's **international brand (BytePlus / Seedream / Seededit)** signs the same content as **"Byteplus Pte. Ltd."**. The bare `volcengine` needle missed it, so BytePlus output was mis-attributed to "Adobe Firefly" through an incidental "Adobe XMP" toolkit string. Issuer `b"Byteplus"` maps directly to "BytePlus (ByteDance)". ByteDance's consumer app **Dreamina** (the international Jimeng brand) signs as **"Bytedance Pte. Ltd."** with a `Dreamina/x.y` claim generator but, unlike the Volcano Engine surface, ships **no `trainedAlgorithmicMedia`**. Issuer `b"Dreamina"` maps to "ByteDance Dreamina" with **`asserts_ai=True`**. The broader **issuer** `b"Bytedance Pte"` is registered only as a neutral signer because that same entity also signs non-AI CapCut edits; it has no AI platform and cannot assert an AI verdict. Keying the product attribution on the `Dreamina` generator token remains the precise path.
 - **EXIF/XMP/PNG-text generator tag (caught by `exif_generator`):** **Ideogram** writes EXIF `Make="Ideogram AI"` (collected 2026-05-24 — no C2PA, no SynthID, no imwatermark; the Make tag is the only signal). Additional verified generator stamps include **NovelAI** (`Software`, `Source`, and `Title` PNG text chunks), **Reve** (`Software` or XMP `CreatorTool` = `reve.com`), and **Aphrodite AI** (`Make` or `Software` = `Aphrodite AI`). A JSON `generationParams.modelName` value is parsed structurally and reduced to a recognized model name, rather than exposing its surrounding prompt JSON as a platform label.
 - **App-export provenance and AIGC JSON:** supported ByteDance-family exports can place a JSON object in EXIF `ImageDescription` or `UserComment`, independently of C2PA or TC260. Exact `product` values for Doubao, Xinghui, and Dreamina are removable product provenance, but do not alone prove that the pixels were generated. Dreamina additionally requires `exportType=generation` for that verdict. A nested Aweme `aigc_type=1` or private ByteDance `aigc_label_type=1` / `2` is an AIGC disclosure; `0` is inconclusive and can occur on a Dreamina generation export. Plain Aweme, retouch, and `lv` exports are preserved. The lower-case private field is deliberately not interpreted as the normative TC260 `Label`, whose values `1` / `2` / `3` mean generated / possibly generated / suspected generated under [GB 45438-2025](https://www.tc260.org.cn/upload/2025-03-15/1742009439794081593.pdf).
 - **xAI / Grok signature scheme (DETECTED by `metadata.xai_signature`, built 2026-05-26, expanded 2026-09-13).**
@@ -4,7 +4,7 @@
 # on it, including the ComfyUI node package. The console script below carries
 # the same weight, since users have it on PATH.
 name = "remove-ai-watermarks"
-version = "0.40.2"
+version = "0.41.0"
 description = "AI watermark remover for visible, invisible, and provenance marks in images and video"
 readme = "README.md"
 requires-python = ">=3.11,<3.15"
@@ -160,6 +160,13 @@ classify-onnx = [
    "remove-ai-watermarks[classify]",
    "onnxruntime>=1.24.0",
 ]
+# Compact OpenAI/Google/unknown source-export classifier. This is intentionally
+# separate from the heavier photo AI-versus-camera classifier and is not a
+# SynthID detector. Weights download on first use.
+source-classify = [
+    "remove-ai-watermarks[pixels]",
+    "huggingface-hub>=0.20.0",
+]
 # Universal region eraser backend -- big-LaMa via onnxruntime (Carve/LaMa-ONNX,
 # Apache-2.0). CPU, no torch. Model (~200 MB) is downloaded on first use and
 # cached by huggingface_hub; it is never bundled in this repo. The default cv2
@@ -205,7 +212,7 @@ dev = [
 ]
 # ``qwen-zimage`` already pulls ``diffusion``; naming both would suggest diffusion is
 # independently sufficient for a removal, which it is not.
-all = ["remove-ai-watermarks[video,heif,detect,trustmark,text-restoration,migan,classify]"]
+all = ["remove-ai-watermarks[video,heif,detect,trustmark,text-restoration,migan,classify,source-classify]"]

 [project.scripts]
 remove-ai-watermarks = "remove_ai_watermarks.cli:main"
@@ -9,7 +9,7 @@ explicitly names a tracked canonical result.

 | Script | Purpose |
 | --- | --- |
-| `corpus_gap_scan.py` | Inventory `identify`, metadata markers, unattributed outcomes, errors, and gap candidates over a full or date-bounded corpus. |
+| `corpus_gap_scan.py` | Inventory `identify`, metadata markers, normalized C2PA review fields, unattributed outcomes, errors, and gap candidates over a full or date-bounded corpus. Its schema-versioned checkpoint resumes safely after interruption. |
 | `detection_timing.py` | Record per-method metadata and verdict timings. |
 | `detection_timing_report.py` | Aggregate timing records by method and segment. |
 | `fidelity_metrics.py` | Compute objective image-fidelity metrics for paired outputs. |
@@ -26,7 +26,7 @@ explicitly names a tracked canonical result.
 | `sidecar_regression.py` | Compare current identification with recorded sidecars. |
 | `smoke_matrix.py` | Exercise CLI parameter choices on real local data. |
 | `video_fidelity_probe.py` | Compare delivered video fidelity with its source. |
-| `visible_eval.py` | Benchmark registered visible-mark detectors. |
+| `visible_eval.py` | Benchmark the adjudicated registry subset and expose unmeasured marks as scope-zero rows. |
 | `visible_removal_audit.py` | Audit visible-removal results over a local corpus. |
 | `watermark_benchmark.py` | Run hash-pinned image cases while keeping detection, removal observation, and fidelity separate. |
 | `watermark_benchmark_report.py` | Aggregate repeated benchmark results with artifact-aware counts and separate cold/warm timing. |
@@ -46,13 +46,16 @@ explicitly names a tracked canonical result.
 | `vendor_mark_calibrate.py` | Calibrate a candidate vendor text detector. |
 | `visible_alpha_solve.py` | Rebuild visible-watermark alpha assets from controlled captures and the cleared Qwen symbol fixture. |
 | `visible_groundtruth.py` | Consolidate blinded contact-sheet labels into ground truth. |
-| `visible_positives.py` | List corpus images carrying a registered visible mark. |
+| `visible_positives.py` | Resumably scan registered marks plus metadata cohorts, and build blinded native-detail bands for uncovered and unbiased quiet samples. |
 | `visible_recall_sample.py` | Build an unbiased blinded sample for recall measurement. |
 | `visible_sheets.py` | Build blinded contact sheets for relaxation candidates. |
 | `watermark_benchmark_cohort.py` | Build the deterministic synthetic DWT-DCT and TrustMark cohort consumed by `watermark_benchmark.py`. |
+| `spectral_pipeline_probe.py` | Measure a prompt-grouped, phase-free spectral source-pipeline hypothesis without presenting it as SynthID detection. |
+| `source_pipeline_classifier.py` | Train or run an abstaining OpenAI/Google/unknown source-export classifier; research-only and not SynthID detection. |
 | `watermark_benchmark_real_cohort.py` | Build a publication-cleared, provider-paired and content-stratified real-image cohort. |
 | `retrain_photo_classify.py` | CPU-retrain the 2026-08-31 photo heads from a sha256-keyed cache pack, no images. |
 | `publish_photo_classify_hf.py` | Upload the photo-classify card and freeze weights to Hugging Face `wiltodelta/raiw-photo-classify`. Manual; the Action `publish-photo-classify-hf.yml` is the write-token path. |
+| `publish_source_classify_hf.py` | Upload the derivative source-pipeline model and its public card to Hugging Face `wiltodelta/raiw-source-classify`. Manual; the Action `publish-source-classify-hf.yml` is the write-token path. |
 | `verify_engine_selection_fixtures.py` | Verify hashes, dimensions, and prompt pairing in the tracked auto-engine content matrix. |

 ## Research and diagnostic prototypes
@@ -85,5 +88,7 @@ running the sync command without `--check` is the explicit snapshot update.

 `_plain_console.py` provides plain-text fallbacks for Rich output, and
 `_text_eval.py` contains normalization helpers shared by text-evaluation scripts.
+`hf_publish.py` owns the common authenticated, allowlisted Hugging Face model
+repository upload used by the two model publishers.
 `engine_selection_manifest.py` owns the content-matrix schema and validation
 shared by its integrity check and the real-image benchmark builder.
@@ -0,0 +1,44 @@
+"""Shared deterministic corpus walks and append-safe JSONL checkpoints."""
+
+from __future__ import annotations
+
+import datetime
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from collections.abc import Collection
+    from pathlib import Path
+
+
+def corpus_day(path: Path, corpus: Path) -> datetime.date | None:
+    """Read the leading YYYY-MM-DD corpus segment, if present."""
+    try:
+        return datetime.date.fromisoformat(path.relative_to(corpus).parts[0])
+    except (ValueError, IndexError):
+        return None
+
+
+def corpus_files(corpus: Path, since: datetime.date | None, *, suffixes: Collection[str] | None = None) -> list[Path]:
+    """Return a deterministic recursive walk, optionally date- and suffix-bounded."""
+    if since is None:
+        candidates = corpus.rglob("*")
+    else:
+        roots = [
+            path
+            for path in corpus.iterdir()
+            if path.is_dir() and (day := corpus_day(path, corpus)) is not None and day >= since
+        ]
+        candidates = (path for root in roots for path in root.rglob("*"))
+    return sorted(
+        path for path in candidates if path.is_file() and (suffixes is None or path.suffix.lower() in suffixes)
+    )
+
+
+def repair_jsonl_tail(path: Path) -> None:
+    """Truncate an interrupted final JSONL fragment before another append."""
+    if not path.exists():
+        return
+    with path.open("rb+") as stream:
+        data = stream.read()
+        if data and not data.endswith(b"\n"):
+            stream.truncate(data.rfind(b"\n") + 1)
@@ -26,9 +26,13 @@ Usage:
        --since 2026-09-01 --report .local-eval/detector-report-weekly.csv

 Rows stream into ``<report>.progress.jsonl`` and resume by relative path after an
-interruption. Use a new report name for a new code/corpus snapshot, or pass
-``--restart`` when intentionally replacing the checkpoint. The final CSV is
-written atomically after every selected row is present.
+interruption. A schema version prevents an older checkpoint from silently
+producing blank fields after the report grows. Use a new report name for a new
+code/corpus snapshot, or pass ``--restart`` when intentionally replacing the
+checkpoint. The final CSV is written atomically after every selected row is
+present. C2PA candidates include normalized signer, claim-generator, action,
+source-type, and validation fields for aggregate review without reopening every
+manifest.
 """

 from __future__ import annotations
@@ -40,16 +44,19 @@ import logging
 import os
 from collections import Counter
 from concurrent.futures import FIRST_COMPLETED, Future, ProcessPoolExecutor, wait
-from datetime import date
 from importlib.metadata import version
 from pathlib import Path
-from typing import Any
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from datetime import date

 import click
+from _corpus_scan import corpus_day, corpus_files, repair_jsonl_tail
 from _plain_console import Console, Table

 from remove_ai_watermarks.identify import _metadata_region as identify_metadata_region
-from remove_ai_watermarks.identify import identify
+from remove_ai_watermarks.identify import extract_provenance_evidence, identify_from_evidence
 from remove_ai_watermarks.metadata import (
    IPTC_AI_FIELD_MARKERS,
    IPTC_AI_MARKERS,
@@ -90,6 +97,7 @@ MARKERS: tuple[bytes, ...] = (
    b"Stability AI",
 )
 REPORT_FIELDS: tuple[str, ...] = (
+    "report_schema_version",
    "path",
    "suffix",
    "lib_version",
@@ -98,18 +106,24 @@ REPORT_FIELDS: tuple[str, ...] = (
    "confidence",
    "watermarks",
    "signals",
+    "c2pa_issuer",
+    "c2pa_claim_generator",
+    "c2pa_actions",
+    "c2pa_source_type",
+    "c2pa_validation",
    "integrity_clashes",
    "markers",
    "candidate_classes",
    "error",
 )
+_REPORT_SCHEMA_VERSION = "2"
 _DISPLAY_CANDIDATE_LIMIT = 50
 _WORKER_BATCH_SIZE = 16


 def _base_row(path: str, suffix: str, lib_version: str) -> dict[str, str]:
    row = dict.fromkeys(REPORT_FIELDS, "")
-    row.update(path=path, suffix=suffix, lib_version=lib_version)
+    row.update(report_schema_version=_REPORT_SCHEMA_VERSION, path=path, suffix=suffix, lib_version=lib_version)
    return row


@@ -166,23 +180,12 @@ def _candidate_classes(rep: Any, hits: list[str]) -> list[str]:

 def _day_of(path: Path, corpus: Path) -> date | None:
    """Read the leading YYYY-MM-DD corpus segment, if this layout has one."""
-    try:
-        segment = path.relative_to(corpus).parts[0]
-        return date.fromisoformat(segment)
-    except (ValueError, IndexError):
-        return None
+    return corpus_day(path, corpus)


 def _files(corpus: Path, since: date | None) -> list[Path]:
    """Return the deterministic corpus walk, optionally bounded by its date segment."""
-    if since is None:
-        return sorted(path for path in corpus.rglob("*") if path.is_file())
-    roots = [
-        path
-        for path in corpus.iterdir()
-        if path.is_dir() and (day := _day_of(path, corpus)) is not None and day >= since
-    ]
-    return sorted(path for root in roots for path in root.rglob("*") if path.is_file())
+    return corpus_files(corpus, since)


 def _scan_one(args: tuple[str, str, str]) -> dict[str, str]:
@@ -191,7 +194,8 @@ def _scan_one(args: tuple[str, str, str]) -> dict[str, str]:
    path = Path(path_str)
    suffix = path.suffix.lower()
    try:
-        rep = identify(path, check_visible=False, check_invisible=False)
+        evidence = extract_provenance_evidence(path)
+        rep = identify_from_evidence(evidence)
    except Exception as exc:
        log.warning("identify failed on %s: %s", relative_path, exc)
        row = _base_row(relative_path, suffix, lib_version)
@@ -199,7 +203,15 @@ def _scan_one(args: tuple[str, str, str]) -> dict[str, str]:
        row["error"] = f"{type(exc).__name__}: {exc}"[:300].replace("\n", " ")
    else:
        row = _row(rep, path=relative_path, suffix=suffix, lib_version=lib_version)
-        hits = _marker_hits_for_path(path)
+        c2pa = evidence.c2pa_info
+        row.update(
+            c2pa_issuer=str(c2pa.get("issuer", "")),
+            c2pa_claim_generator=str(c2pa.get("claim_generator", "")),
+            c2pa_actions=str(c2pa.get("actions", "")),
+            c2pa_source_type=str(c2pa.get("source_type", "")),
+            c2pa_validation=str(c2pa.get("c2pa_validation_state", "")),
+        )
+        hits = _marker_hits(identify_metadata_region(evidence.scan))
        row["markers"] = "|".join(hits)
        row["candidate_classes"] = "|".join(_candidate_classes(rep, hits))
        return row
@@ -223,20 +235,18 @@ def _read_checkpoint(path: Path) -> dict[str, dict[str, str]]:
                row = json.loads(line)
            except (json.JSONDecodeError, UnicodeDecodeError):
                continue
-            if isinstance(row, dict) and isinstance(row.get("path"), str):
+            if (
+                isinstance(row, dict)
+                and row.get("report_schema_version") == _REPORT_SCHEMA_VERSION
+                and isinstance(row.get("path"), str)
+            ):
                rows[row["path"]] = {field: str(row.get(field, "")) for field in REPORT_FIELDS}
    return rows


 def _repair_checkpoint(path: Path) -> None:
    """Remove an interrupted final JSONL fragment before another append."""
-    if not path.exists():
-        return
-    with path.open("rb+") as stream:
-        data = stream.read()
-        if data and not data.endswith(b"\n"):
-            last_newline = data.rfind(b"\n")
-            stream.truncate(last_newline + 1)
+    repair_jsonl_tail(path)


 def _summarize(rows: list[dict[str, str]]) -> None:
@@ -0,0 +1,31 @@
+"""Shared Hugging Face model-repository publication helper."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from collections.abc import Iterable
+    from pathlib import Path
+
+
+def publish_model_folder(
+    stage: Path,
+    *,
+    repo_id: str,
+    token: str,
+    message: str,
+    allow_patterns: Iterable[str],
+) -> str:
+    """Upload an explicit file allowlist and return the Hub revision."""
+    from huggingface_hub import HfApi, create_repo
+
+    create_repo(repo_id, repo_type="model", exist_ok=True, private=False, token=token)
+    commit = HfApi(token=token).upload_folder(
+        folder_path=str(stage),
+        repo_id=repo_id,
+        repo_type="model",
+        commit_message=message,
+        allow_patterns=list(allow_patterns),
+    )
+    return getattr(commit, "oid", "") or ""
@@ -24,6 +24,8 @@ import shutil
 import tempfile
 from pathlib import Path

+from hf_publish import publish_model_folder
+
 log = logging.getLogger(__name__)

 HUB_REPO = "wiltodelta/raiw-photo-classify"
@@ -92,18 +94,13 @@ def stage_weights(dest: Path, src: Path) -> None:


 def publish(stage: Path, *, token: str, message: str) -> str:
-    from huggingface_hub import HfApi, create_repo
-
-    create_repo(HUB_REPO, repo_type="model", exist_ok=True, private=False, token=token)
-    api = HfApi(token=token)
-    commit = api.upload_folder(
-        folder_path=str(stage),
+    return publish_model_folder(
+        stage,
        repo_id=HUB_REPO,
-        repo_type="model",
-        commit_message=message,
+        token=token,
+        message=message,
        allow_patterns=["README.md", "operating-point.json", *WEIGHT_FILES, GATE_FILE, GATE_STABLE_FILE],
    )
-    return getattr(commit, "oid", "") or ""


 def main() -> int:
@@ -0,0 +1,73 @@
+#!/usr/bin/env python3
+"""Publish the source-pipeline model to Hugging Face.
+
+The derivative weight file is downloaded by the GitHub Action from a model
+freeze release. The private corpus, feature caches, and training catalog never
+enter this repository or the Hub staging directory.
+"""
+
+from __future__ import annotations
+
+import argparse
+import logging
+import os
+import shutil
+import tempfile
+from pathlib import Path
+
+from hf_publish import publish_model_folder
+
+log = logging.getLogger(__name__)
+
+HUB_REPO = "wiltodelta/raiw-source-classify"
+MODEL_FILE = "source-pipeline-mlp.npz"
+REPO_ROOT = Path(__file__).resolve().parents[1]
+CARD_DIR = REPO_ROOT / "docs" / "source-classify-hf"
+PUBLIC_FILES = ("README.md", "metrics.json", MODEL_FILE)
+
+
+def stage_release(dest: Path, src: Path) -> None:
+    """Stage the two tracked documents and one derivative model artifact."""
+    sources = {
+        "README.md": CARD_DIR / "README.md",
+        "metrics.json": CARD_DIR / "metrics.json",
+        MODEL_FILE: src.expanduser().resolve() / MODEL_FILE,
+    }
+    missing = [str(path) for path in sources.values() if not path.is_file()]
+    if missing:
+        raise SystemExit(f"missing publication files: {', '.join(missing)}")
+    dest.mkdir(parents=True, exist_ok=True)
+    for name, path in sources.items():
+        shutil.copy2(path, dest / name)
+
+
+def publish(stage: Path, *, token: str, message: str) -> str:
+    """Upload the exact staged allowlist and return the Hub revision."""
+    return publish_model_folder(
+        stage,
+        repo_id=HUB_REPO,
+        token=token,
+        message=message,
+        allow_patterns=PUBLIC_FILES,
+    )
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--src", type=Path, required=True)
+    parser.add_argument("--message", default="Publish source-pipeline classifier")
+    args = parser.parse_args()
+    logging.basicConfig(level=logging.INFO, format="%(message)s")
+    token = os.environ.get("HF_TOKEN", "").strip()
+    if not token:
+        raise SystemExit("HF_TOKEN is not set (write role required)")
+    with tempfile.TemporaryDirectory(prefix="raiw-source-hf-") as tmp:
+        stage = Path(tmp)
+        stage_release(stage, args.src)
+        revision = publish(stage, token=token, message=args.message)
+    log.info("published %s revision=%s", HUB_REPO, revision or "(unknown)")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
@@ -0,0 +1,371 @@
+"""Train or run an abstaining source/export-pipeline classifier.
+
+This research tool combines the library's 124-d forensic descriptor with the
+phase-free spectral descriptor from ``spectral_pipeline_probe.py``. Its labels
+describe complete generation and export pipelines. It does not decode SynthID
+or establish that a watermark is present.
+
+Expected training layout::
+
+    train/openai/*.png
+    train/google/*.png
+    train/unknown/*.png
+    calibration/openai/*.png
+    calibration/google/*.png
+    calibration/unknown/*.png
+
+Examples::
+
+    uv run python scripts/source_pipeline_classifier.py train \
+        train calibration --model-out model.npz
+    uv run python scripts/source_pipeline_classifier.py predict \
+        model.npz image.png
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import dataclass, replace
+from pathlib import Path
+from typing import cast
+
+import click
+import numpy as np
+import spectral_pipeline_probe as spectral
+from numpy.typing import NDArray
+from PIL import Image
+from watermark_benchmark import sha256_file
+
+from remove_ai_watermarks._internal.forensic_124d import FEATURE_WIDTH, PATCH, image_features
+
+log = logging.getLogger(__name__)
+
+CLASSES = ("openai", "google", "unknown")
+UNKNOWN_INDEX = CLASSES.index("unknown")
+DEFAULT_GRID = 16
+DEFAULT_PENALTY = 100.0
+DEFAULT_UNKNOWN_FPR = 0.01
+CLAIM = "source/export pipeline classification; not SynthID detection"
+
+FloatArray = NDArray[np.float64]
+IntArray = NDArray[np.int64]
+
+
+@dataclass(frozen=True)
+class RidgeModel:
+    """One standardized three-class ridge head with an abstention margin."""
+
+    mean: FloatArray
+    scale: FloatArray
+    weights: FloatArray
+    margin: float
+    grid: int
+    penalty: float
+
+
+def discover_labeled_paths(root: Path) -> tuple[list[Path], IntArray, list[str]]:
+    """Find images under the three fixed label directories."""
+    paths: list[Path] = []
+    labels: list[int] = []
+    hashes: list[str] = []
+    for class_index, class_name in enumerate(CLASSES):
+        directory = root / class_name
+        if not directory.is_dir():
+            raise ValueError(f"class directory does not exist: {directory}")
+        class_paths = sorted(
+            path for path in directory.rglob("*") if path.is_file() and path.suffix.lower() in spectral.IMAGE_SUFFIXES
+        )
+        if not class_paths:
+            raise ValueError(f"class directory has no supported images: {directory}")
+        for path in class_paths:
+            paths.append(path)
+            labels.append(class_index)
+            hashes.append(sha256_file(path))
+    if len(hashes) != len(set(hashes)):
+        raise ValueError(f"duplicate image bytes found inside {root}")
+    return paths, np.asarray(labels, dtype=np.int64), hashes
+
+
+def _feature_from_path(path: Path, geometry: spectral.SpectralGeometry) -> FloatArray | None:
+    """Return the fused representation for one path and prepared geometry."""
+    with Image.open(path) as source:
+        image = source.convert("RGB")
+    forensic = image_features(np.asarray(image, dtype=np.uint8))
+    if forensic is None:
+        return None
+    phase_free = spectral.feature_from_image(image, geometry, "original")
+    return np.concatenate(
+        (
+            np.asarray(forensic, dtype=np.float64),
+            np.asarray(phase_free, dtype=np.float64),
+        )
+    )
+
+
+def feature_from_path(path: Path, grid: int) -> FloatArray | None:
+    """Return the fused forensic and phase-free spectral representation."""
+    return _feature_from_path(path, spectral.make_geometry(size=PATCH, grid=grid))
+
+
+def extract_features(paths: list[Path], labels: IntArray, grid: int) -> tuple[FloatArray, IntArray, list[Path]]:
+    """Extract valid rows and report files whose forensic feature abstains."""
+    geometry = spectral.make_geometry(size=PATCH, grid=grid)
+    features: list[FloatArray] = []
+    kept_labels: list[int] = []
+    skipped: list[Path] = []
+    for index, (path, label) in enumerate(zip(paths, labels, strict=True), start=1):
+        try:
+            feature = _feature_from_path(path, geometry)
+        except (OSError, ValueError) as error:
+            log.warning("Could not extract %s: %s", path, error)
+            feature = None
+        if feature is None:
+            skipped.append(path)
+            continue
+        features.append(feature)
+        kept_labels.append(int(label))
+        if index % 100 == 0:
+            log.info("Extracted %s/%s images", index, len(paths))
+    if not features:
+        raise ValueError("no image produced a valid feature vector")
+    return np.asarray(features), np.asarray(kept_labels, dtype=np.int64), skipped
+
+
+def fit_ridge(
+    features: FloatArray,
+    labels: IntArray,
+    penalty: float,
+    grid: int = DEFAULT_GRID,
+) -> RidgeModel:
+    """Fit a class-balanced standardized ridge head before calibration."""
+    if features.ndim != 2 or len(features) != len(labels):
+        raise ValueError("features and labels must describe the same rows")
+    counts = np.bincount(labels, minlength=len(CLASSES)).astype(np.float64)
+    if np.any(counts == 0):
+        raise ValueError("training data must contain all three classes")
+    if penalty <= 0:
+        raise ValueError("penalty must be positive")
+    mean = features.mean(axis=0)
+    raw_scale = features.std(axis=0)
+    active = raw_scale > 1e-10
+    scale = np.where(active, raw_scale, 1.0)
+    normalized = (features[:, active] - mean[active]) / scale[active]
+    sample_weight = len(labels) / (len(CLASSES) * counts[labels])
+    root_weight = np.sqrt(sample_weight)[:, None]
+    weighted = normalized * root_weight
+    targets = np.eye(len(CLASSES), dtype=np.float64)[labels] * root_weight
+    gram = weighted.T @ weighted + penalty * np.eye(weighted.shape[1])
+    active_weights = cast("FloatArray", np.linalg.solve(gram, weighted.T @ targets))
+    weights = np.zeros((features.shape[1], len(CLASSES)), dtype=np.float64)
+    weights[active] = active_weights
+    return RidgeModel(mean, scale, weights, 0.0, grid, penalty)
+
+
+def scores(features: FloatArray, model: RidgeModel) -> FloatArray:
+    """Return the uncalibrated ridge scores."""
+    return ((features - model.mean) / model.scale) @ model.weights
+
+
+def decision_margins(raw_scores: FloatArray) -> FloatArray:
+    """Return each row's best-provider advantage over every alternative."""
+    target_scores = raw_scores[:, :UNKNOWN_INDEX]
+    best = np.argmax(target_scores, axis=1)
+    other = 1 - best
+    rows = np.arange(len(raw_scores))
+    reference = np.maximum(raw_scores[rows, other], raw_scores[:, UNKNOWN_INDEX])
+    return target_scores[rows, best] - reference
+
+
+def calibrate_margin(
+    calibration_scores: FloatArray,
+    calibration_labels: IntArray,
+    unknown_fpr: float,
+) -> float:
+    """Choose the smallest empirical cut meeting the unknown false-positive target."""
+    if not 0 <= unknown_fpr < 1:
+        raise ValueError("unknown_fpr must be in [0, 1)")
+    unknown_scores = calibration_scores[calibration_labels == UNKNOWN_INDEX]
+    if not len(unknown_scores):
+        raise ValueError("calibration data must contain unknown rows")
+    return max(
+        0.0,
+        float(
+            np.quantile(
+                decision_margins(unknown_scores),
+                1.0 - unknown_fpr,
+                method="higher",
+            )
+            + 1e-12
+        ),
+    )
+
+
+def predict_scores(raw_scores: FloatArray, margin: float) -> IntArray:
+    """Return provider labels only when the calibrated margin is met."""
+    best = np.argmax(raw_scores[:, :UNKNOWN_INDEX], axis=1)
+    accepted = decision_margins(raw_scores) >= margin
+    return np.where(accepted, best, UNKNOWN_INDEX).astype(np.int64)
+
+
+def save_model(path: Path, model: RidgeModel) -> None:
+    """Write a pickle-free model artifact."""
+    path.parent.mkdir(parents=True, exist_ok=True)
+    np.savez_compressed(
+        path,
+        classes=np.asarray(CLASSES),
+        mean=model.mean,
+        scale=model.scale,
+        weights=model.weights,
+        margin=np.asarray(model.margin),
+        spectral_grid=np.asarray(model.grid),
+        penalty=np.asarray(model.penalty),
+    )
+
+
+def load_model(path: Path) -> RidgeModel:
+    """Load and validate a pickle-free model artifact."""
+    with np.load(path, allow_pickle=False) as payload:
+        required = {
+            "classes",
+            "mean",
+            "scale",
+            "weights",
+            "margin",
+            "spectral_grid",
+            "penalty",
+        }
+        if set(payload.files) != required:
+            raise ValueError(f"model fields must be exactly {sorted(required)}")
+        classes = tuple(str(value) for value in payload["classes"])
+        mean = np.asarray(payload["mean"], dtype=np.float64)
+        scale = np.asarray(payload["scale"], dtype=np.float64)
+        weights = np.asarray(payload["weights"], dtype=np.float64)
+        margin = float(payload["margin"])
+        grid = int(payload["spectral_grid"])
+        penalty = float(payload["penalty"])
+    expected_features = FEATURE_WIDTH + 3 * grid * grid
+    if classes != CLASSES:
+        raise ValueError(f"model classes must be {CLASSES}")
+    if mean.shape != (expected_features,) or scale.shape != mean.shape:
+        raise ValueError("model normalization vectors have the wrong shape")
+    if weights.shape != (expected_features, len(CLASSES)):
+        raise ValueError("model weights have the wrong shape")
+    if not np.all(np.isfinite(mean)) or not np.all(np.isfinite(scale)):
+        raise ValueError("model normalization vectors must be finite")
+    if not np.all(np.isfinite(weights)) or np.any(scale <= 0):
+        raise ValueError("model weights must be finite and scales positive")
+    if not np.isfinite(margin) or margin < 0:
+        raise ValueError("model margin must be finite and non-negative")
+    if grid < 2 or PATCH % grid or not np.isfinite(penalty) or penalty <= 0:
+        raise ValueError("model calibration fields are invalid")
+    return RidgeModel(mean, scale, weights, margin, grid, penalty)
+
+
+@click.group()
+def main() -> None:
+    """Train or run the research-only source-pipeline classifier."""
+    logging.basicConfig(level=logging.INFO, format="%(levelname)s %(message)s")
+
+
+@main.command("train")
+@click.argument("train_root", type=click.Path(exists=True, file_okay=False, path_type=Path))
+@click.argument("calibration_root", type=click.Path(exists=True, file_okay=False, path_type=Path))
+@click.option("--model-out", type=click.Path(dir_okay=False, path_type=Path), required=True)
+@click.option("--grid", type=click.IntRange(min=2), default=DEFAULT_GRID, show_default=True)
+@click.option("--penalty", type=click.FloatRange(min=0.0, min_open=True), default=DEFAULT_PENALTY, show_default=True)
+@click.option(
+    "--unknown-fpr",
+    type=click.FloatRange(min=0.0, max=1.0, max_open=True),
+    default=DEFAULT_UNKNOWN_FPR,
+    show_default=True,
+)
+def train_command(
+    train_root: Path,
+    calibration_root: Path,
+    model_out: Path,
+    grid: int,
+    penalty: float,
+    unknown_fpr: float,
+) -> None:
+    """Fit on TRAIN_ROOT and calibrate abstention on CALIBRATION_ROOT."""
+    if PATCH % grid:
+        raise click.UsageError(f"--grid must divide {PATCH}")
+    try:
+        train_paths, train_labels, train_hashes = discover_labeled_paths(train_root)
+        calibration_paths, calibration_labels, calibration_hashes = discover_labeled_paths(calibration_root)
+        overlap = set(train_hashes) & set(calibration_hashes)
+        if overlap:
+            raise ValueError(f"train/calibration byte overlap: {len(overlap)} images")
+        train_features, kept_train_labels, train_skipped = extract_features(train_paths, train_labels, grid)
+        calibration_features, kept_calibration_labels, calibration_skipped = extract_features(
+            calibration_paths, calibration_labels, grid
+        )
+        provisional = fit_ridge(train_features, kept_train_labels, penalty, grid)
+        margin = calibrate_margin(
+            scores(calibration_features, provisional),
+            kept_calibration_labels,
+            unknown_fpr,
+        )
+        model = replace(provisional, margin=margin)
+        save_model(model_out, model)
+    except ValueError as error:
+        raise click.ClickException(str(error)) from error
+    click.echo(
+        json.dumps(
+            {
+                "model": str(model_out),
+                "claim": CLAIM,
+                "train_rows": len(train_features),
+                "calibration_rows": len(calibration_features),
+                "feature_abstentions": len(train_skipped) + len(calibration_skipped),
+                "margin": margin,
+            },
+            indent=2,
+        )
+    )
+
+
+@main.command("predict")
+@click.argument("model_path", type=click.Path(exists=True, dir_okay=False, path_type=Path))
+@click.argument("images", nargs=-1, required=True, type=click.Path(exists=True, dir_okay=False, path_type=Path))
+def predict_command(model_path: Path, images: tuple[Path, ...]) -> None:
+    """Classify original-export IMAGES with MODEL_PATH."""
+    try:
+        model = load_model(model_path)
+    except (OSError, ValueError) as error:
+        raise click.ClickException(str(error)) from error
+    results: list[dict[str, object]] = []
+    geometry = spectral.make_geometry(size=PATCH, grid=model.grid)
+    for path in images:
+        try:
+            feature = _feature_from_path(path, geometry)
+        except (OSError, ValueError) as error:
+            raise click.ClickException(f"could not extract {path}: {error}") from error
+        if feature is None:
+            results.append(
+                {
+                    "path": str(path),
+                    "label": "unknown",
+                    "reason": "124-d feature unavailable",
+                    "claim": CLAIM,
+                }
+            )
+            continue
+        raw_scores = scores(feature[None, :], model)
+        prediction = int(predict_scores(raw_scores, model.margin)[0])
+        results.append(
+            {
+                "path": str(path),
+                "label": CLASSES[prediction],
+                "decision_margin": float(decision_margins(raw_scores)[0]),
+                "required_margin": model.margin,
+                "scores": {name: float(raw_scores[0, index]) for index, name in enumerate(CLASSES)},
+                "claim": CLAIM,
+            }
+        )
+    click.echo(json.dumps(results, indent=2))
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,296 @@
+"""Measure a phase-free spectral source-pipeline hypothesis.
+
+This is a research harness, not a SynthID detector. It classifies complete
+generation/export pipelines whose renderer, output geometry, encoding, and
+watermark state may all be confounded. A watermark claim requires a causal
+marked/clean contrast and an independent oracle.
+
+The input directory contains one subdirectory per class. Matching filename
+stems are treated as one prompt group and always stay in the same fold.
+
+Example:
+    uv run python scripts/spectral_pipeline_probe.py DATASET \
+        --class-dir gemini --class-dir openai --class-dir flux \
+        --attack original --attack crop2 --attack resize95 \
+        --attack jpeg75 --attack blur \
+        --report-out .local-eval/spectral-pipeline-report.json
+"""
+
+from __future__ import annotations
+
+import io
+import json
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+from typing import cast
+
+import click
+import numpy as np
+from numpy.typing import NDArray
+from PIL import Image, ImageFilter
+
+from remove_ai_watermarks._internal import source_spectral as _spectral
+
+SpectralGeometry = _spectral.SpectralGeometry
+make_geometry = _spectral.make_geometry
+phase_free_log_spectrum = _spectral.phase_free_log_spectrum
+spectral_feature = _spectral.spectral_feature
+_source_feature_from_image = _spectral.feature_from_image
+
+log = logging.getLogger(__name__)
+
+IMAGE_SUFFIXES = frozenset({".jpg", ".jpeg", ".png", ".webp"})
+ATTACKS = ("original", "crop2", "resize95", "jpeg75", "blur")
+
+FloatArray = NDArray[np.float64]
+IntArray = NDArray[np.int64]
+BoolArray = NDArray[np.bool_]
+
+
+@dataclass(frozen=True)
+class FoldModel:
+    """One standardized dual-ridge model and its held-out rows."""
+
+    test_rows: BoolArray
+    mean: FloatArray
+    scale: FloatArray
+    active: BoolArray
+    weights: FloatArray
+
+
+def _apply_attack(image: Image.Image, attack: str) -> Image.Image:
+    """Apply one declared robustness transform before canonical resizing."""
+    if attack == "original":
+        return image
+    if attack == "crop2":
+        if image.width <= 4 or image.height <= 4:
+            raise ValueError("crop2 requires both image dimensions to exceed four pixels")
+        return image.crop((2, 2, image.width, image.height))
+    if attack == "resize95":
+        return image.resize(
+            (max(1, round(image.width * 0.95)), max(1, round(image.height * 0.95))),
+            Image.Resampling.LANCZOS,
+        )
+    if attack == "jpeg75":
+        encoded = io.BytesIO()
+        image.save(encoded, format="JPEG", quality=75)
+        encoded.seek(0)
+        with Image.open(encoded) as decoded:
+            return decoded.convert("RGB")
+    if attack == "blur":
+        return image.filter(ImageFilter.GaussianBlur(0.7))
+    raise ValueError(f"unsupported attack: {attack}")
+
+
+def feature_from_path(path: Path, geometry: SpectralGeometry, attack: str) -> FloatArray:
+    """Decode one path, apply ATTACK, and return its phase-free feature."""
+    with Image.open(path) as source:
+        return feature_from_image(source.convert("RGB"), geometry, attack)
+
+
+def feature_from_image(image: Image.Image, geometry: SpectralGeometry, attack: str) -> FloatArray:
+    """Extract one attacked feature from an already decoded RGB image."""
+    attacked = _apply_attack(image, attack)
+    return _source_feature_from_image(attacked, geometry)
+
+
+def discover_grouped_paths(
+    root: Path,
+    class_dirs: tuple[str, ...],
+) -> tuple[list[Path], IntArray, IntArray, list[str]]:
+    """Return class-major paths over filename stems shared by every class."""
+    if len(class_dirs) < 2 or len(set(class_dirs)) != len(class_dirs):
+        raise ValueError("at least two unique class directories are required")
+    per_class: list[dict[str, Path]] = []
+    for class_dir in class_dirs:
+        directory = root / class_dir
+        if not directory.is_dir():
+            raise ValueError(f"class directory does not exist: {directory}")
+        mapping: dict[str, Path] = {}
+        for path in sorted(directory.iterdir()):
+            if path.is_file() and path.suffix.lower() in IMAGE_SUFFIXES:
+                if path.stem in mapping:
+                    raise ValueError(f"duplicate filename stem in {directory}: {path.stem}")
+                mapping[path.stem] = path
+        per_class.append(mapping)
+    common_names: set[str] = set(per_class[0])
+    for mapping in per_class[1:]:
+        common_names.intersection_update(mapping)
+    group_names = sorted(common_names)
+    if not group_names:
+        raise ValueError("class directories have no shared filename stems")
+    paths: list[Path] = []
+    labels: list[int] = []
+    groups: list[int] = []
+    for class_index, mapping in enumerate(per_class):
+        for group_index, group_name in enumerate(group_names):
+            paths.append(mapping[group_name])
+            labels.append(class_index)
+            groups.append(group_index)
+    return paths, np.asarray(labels), np.asarray(groups), group_names
+
+
+def fit_grouped_models(
+    features: FloatArray,
+    labels: IntArray,
+    groups: IntArray,
+    *,
+    class_count: int,
+    folds: int = 5,
+    ridge: float = 1000.0,
+) -> list[FoldModel]:
+    """Fit prompt-grouped dual ridge models without test-fold leakage."""
+    if features.ndim != 2 or len(features) != len(labels) or len(labels) != len(groups):
+        raise ValueError("features, labels, and groups must describe the same rows")
+    if folds < 2 or len(np.unique(groups)) < folds:
+        raise ValueError("fold count must be between two and the number of groups")
+    if class_count < 2 or ridge <= 0:
+        raise ValueError("class_count must be at least two and ridge must be positive")
+    targets = np.eye(class_count, dtype=np.float64)[labels]
+    models: list[FoldModel] = []
+    for fold in range(folds):
+        test_rows = groups % folds == fold
+        train_rows = ~test_rows
+        train = features[train_rows]
+        mean = np.mean(train, axis=0)
+        scale = np.std(train, axis=0)
+        active = scale > 1e-8
+        train_features = (train[:, active] - mean[active]) / scale[active]
+        kernel = train_features @ train_features.T
+        dual_weights = cast(
+            "FloatArray",
+            np.linalg.solve(
+                kernel + ridge * np.eye(len(kernel)),
+                targets[train_rows],
+            ),
+        )
+        weights = train_features.T @ dual_weights
+        models.append(FoldModel(test_rows, mean, scale, active, weights))
+    return models
+
+
+def predict_grouped(models: list[FoldModel], features: FloatArray) -> IntArray:
+    """Predict every row only with the model that held out its group."""
+    predictions = np.full(len(features), -1, dtype=np.int64)
+    for model in models:
+        test_features = (features[model.test_rows][:, model.active] - model.mean[model.active]) / model.scale[
+            model.active
+        ]
+        scores = test_features @ model.weights
+        predictions[model.test_rows] = np.argmax(scores, axis=1)
+    if np.any(predictions < 0):
+        raise ValueError("fold models did not cover every row")
+    return predictions
+
+
+def confusion_matrix(labels: IntArray, predictions: IntArray, class_count: int) -> IntArray:
+    """Count actual rows by predicted class."""
+    matrix = np.zeros((class_count, class_count), dtype=np.int64)
+    np.add.at(matrix, (labels, predictions), 1)
+    return matrix
+
+
+def _extract_all_attacks(
+    paths: list[Path],
+    geometry: SpectralGeometry,
+    attacks: tuple[str, ...],
+) -> dict[str, FloatArray]:
+    """Decode each path once and extract every requested attack feature."""
+    log.info("Extracting %s attack views from %s images", len(attacks), len(paths))
+    rows: dict[str, list[FloatArray]] = {attack: [] for attack in attacks}
+    for path in paths:
+        with Image.open(path) as source:
+            image = source.convert("RGB")
+        for attack in attacks:
+            rows[attack].append(feature_from_image(image, geometry, attack))
+    return {attack: np.asarray(features) for attack, features in rows.items()}
+
+
+@click.command()
+@click.argument("dataset", type=click.Path(exists=True, file_okay=False, path_type=Path))
+@click.option(
+    "class_dirs",
+    "--class-dir",
+    multiple=True,
+    required=True,
+    help="Class subdirectory; repeat in label order.",
+)
+@click.option(
+    "attacks",
+    "--attack",
+    multiple=True,
+    type=click.Choice(ATTACKS),
+    default=("original",),
+    show_default=True,
+)
+@click.option("--size", type=click.IntRange(min=16), default=256, show_default=True)
+@click.option("--grid", type=click.IntRange(min=2), default=32, show_default=True)
+@click.option("--folds", type=click.IntRange(min=2), default=5, show_default=True)
+@click.option("--ridge", type=click.FloatRange(min=0.0, min_open=True), default=1000.0, show_default=True)
+@click.option("--report-out", type=click.Path(dir_okay=False, path_type=Path), required=True)
+def main(
+    dataset: Path,
+    class_dirs: tuple[str, ...],
+    attacks: tuple[str, ...],
+    size: int,
+    grid: int,
+    folds: int,
+    ridge: float,
+    report_out: Path,
+) -> None:
+    """Run grouped spectral pipeline classification over DATASET."""
+    logging.basicConfig(level=logging.INFO, format="%(levelname)s %(message)s")
+    if "original" not in attacks:
+        raise click.UsageError("--attack original is required because every model trains on original exports")
+    unique_attacks = tuple(dict.fromkeys(attacks))
+    try:
+        geometry = make_geometry(size, grid)
+        paths, labels, groups, group_names = discover_grouped_paths(dataset, class_dirs)
+        features_by_attack = _extract_all_attacks(paths, geometry, unique_attacks)
+        original = features_by_attack["original"]
+        models = fit_grouped_models(
+            original,
+            labels,
+            groups,
+            class_count=len(class_dirs),
+            folds=folds,
+            ridge=ridge,
+        )
+    except ValueError as error:
+        raise click.ClickException(str(error)) from error
+
+    results: dict[str, object] = {}
+    for attack in unique_attacks:
+        features = features_by_attack[attack]
+        predictions = predict_grouped(models, features)
+        matrix = confusion_matrix(labels, predictions, len(class_dirs))
+        results[attack] = {
+            "accuracy": float(np.trace(matrix) / np.sum(matrix)),
+            "confusion_matrix": matrix.tolist(),
+        }
+
+    report = {
+        "claim": "phase-free generation/export pipeline classification; not SynthID detection",
+        "causal_limit": (
+            "Renderer, output geometry, encoding, and watermark state vary together. "
+            "Only a marked/clean causal contrast plus an independent oracle can isolate the watermark."
+        ),
+        "evaluation_limit": (
+            "Grouped cross-validation estimates this corpus only. Hyperparameters selected on these folds "
+            "need an independent locked corpus before they support a generalization claim."
+        ),
+        "dataset": str(dataset),
+        "classes": list(class_dirs),
+        "group_count": len(group_names),
+        "image_count": len(paths),
+        "configuration": {"size": size, "grid": grid, "folds": folds, "ridge": ridge},
+        "results": results,
+    }
+    report_out.parent.mkdir(parents=True, exist_ok=True)
+    report_out.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
+    log.info("Wrote spectral pipeline report: %s", report_out)
+
+
+if __name__ == "__main__":
+    main()
@@ -41,7 +41,10 @@ from remove_ai_watermarks.image_io import imread

 GT = Path(".local-eval/textmark-relaxation/groundtruth.jsonl")
 SNAP = Path(".local-eval/textmark-relaxation/snapshots")
-MARKS = ("gemini", "doubao", "jimeng", "samsung", "jimeng_pill")
+# Ground truth currently adjudicates only a subset, but inventory follows the
+# registry. A newly registered mark therefore appears with scope zero instead of
+# silently vanishing from the evaluation harness.
+MARKS = tuple(wr.mark_keys())


 def wilson(k: int, n: int) -> tuple[float, float]:
@@ -122,7 +125,7 @@ def report(res: dict, prev: dict | None = None) -> None:
        lo, hi = wilson(tp, n)
        delta = ""
        if prev:
-            pc = prev["per"][m]
+            pc = prev["per"].get(m, {})
            pn = pc.get("tp", 0) + pc.get("fp", 0)
            if pn:
                d = tp / n - pc.get("tp", 0) / pn
@@ -1,4 +1,4 @@
-"""Parallel detection pass: list every corpus image carrying a known visible mark.
+"""Scan a retained corpus for registered and still-uncovered visible marks.

 Why this exists separately from `visible_removal_audit.py`: that audit is single-process,
 so repeated full-dataset sweeps waste time. Its
@@ -18,22 +18,43 @@ CRASH TOLERANCE IS NOT OPTIONAL AT THIS SCALE

 Treat input datasets as sensitive and read-only, and keep output gitignored.

-    uv run python scripts/visible_positives.py --jobs 6
+The JSONL is also the input to visible-mark discovery. Each row records strict and
+metadata-corroborated detector results plus the metadata cohort independently of
+the pixels. ``--sheets`` then makes blinded, native-detail top/bottom bands from:
+
+* quiet TC260 producer cohorts, including producer identities absent from the
+  visible-mark registry;
+* quiet named-platform cohorts;
+* an unbiased sample of every other quiet image, so a completely novel mark with
+  no useful metadata is not excluded by construction.
+
+The sheets are candidates for human adjudication, not evidence that a watermark is
+present. Their manifest is deliberately separate and must stay closed until the
+image review finishes.
+
+    uv run python scripts/visible_positives.py --corpus .local-eval/originals --jobs 6
+    uv run python scripts/visible_positives.py --since 2026-09-01 --sheets .local-eval/visible-fresh-sheets
 """

 from __future__ import annotations

 import argparse
 import collections
-import glob
 import json
 import os
+import random
+import shutil
+import subprocess
 import sys
+from datetime import date
+from importlib.metadata import version
 from pathlib import Path
+from typing import Any

 sys.path.insert(0, str(Path(__file__).parent.parent))
 sys.path.insert(0, str(Path(__file__).parent))

+from _corpus_scan import corpus_files, repair_jsonl_tail
 from _isolated_image_workers import run_batch

 # The package's own format set. An inlined copy here silently skipped .heif, which
@@ -43,20 +64,66 @@ from remove_ai_watermarks._internal.constants import SUPPORTED_FORMATS as _EXTS
 REPO = Path(__file__).resolve().parents[1]
 CORPUS = REPO / ".local-eval" / "originals"
 OUT = REPO / ".local-eval" / "visible-positives.jsonl"
-PATHS = REPO / ".local-eval" / "visible-positives.txt"
+_SHEET_BAND_FRACTION = 0.12
+_SCHEMA_VERSION = 2


 def _one(path: str) -> dict[str, object]:
+    from remove_ai_watermarks.api import _provenance_from_report
+    from remove_ai_watermarks.identify import identify
    from remove_ai_watermarks.image_io import imread
-    from remove_ai_watermarks.watermark_registry import detect_marks
+    from remove_ai_watermarks.metadata import aigc_label, uscc_of
+    from remove_ai_watermarks.watermark_registry import _provenance_confirms_product, known_marks

    try:
-        img = imread(path)
+        source = Path(path)
+        img = imread(source)
        if img is None:
-            return {"path": path, "keys": [], "status": "unreadable"}
-        return {"path": path, "keys": [d.key for d in detect_marks(img) if d.detected], "status": "ok"}
+            return _error_record(path, "unreadable")
+        report = identify(source, check_visible=False, check_invisible=False)
+        provenance = _provenance_from_report(report, source)
+        strict = []
+        accepted = []
+        for mark in known_marks():
+            strict_detection, relaxed_detection = mark.detect_both(img)
+            if strict_detection.detected:
+                strict.append(mark.key)
+            detection = (
+                relaxed_detection if _provenance_confirms_product(mark.product, provenance) else strict_detection
+            )
+            if detection.detected:
+                accepted.append(mark.key)
+        label = aigc_label(source) or {}
+        producer = str(label.get("ContentProducer") or "")
+        return {
+            "schema_version": _SCHEMA_VERSION,
+            "path": path,
+            "keys": accepted,
+            "strict_keys": strict,
+            "provenance": sorted(provenance),
+            "platform": report.platform or "",
+            "producer": producer,
+            "uscc": uscc_of(producer) if producer else "",
+            "status": "ok",
+            "error": "",
+        }
    except Exception as e:
-        return {"path": path, "keys": [], "status": f"error:{type(e).__name__}"}
+        return _error_record(path, f"error:{type(e).__name__}", str(e))
+
+
+def _error_record(path: str, status: str, error: str = "") -> dict[str, object]:
+    return {
+        "schema_version": _SCHEMA_VERSION,
+        "path": path,
+        "keys": [],
+        "strict_keys": [],
+        "provenance": [],
+        "platform": "",
+        "producer": "",
+        "uscc": "",
+        "status": status,
+        "error": error[:300].replace("\n", " "),
+    }


 def _run_batch(batch: list[str], jobs: int, timeout: int) -> list[dict[str, object]]:
@@ -64,65 +131,194 @@ def _run_batch(batch: list[str], jobs: int, timeout: int) -> list[dict[str, obje
    return run_batch(Path(__file__), batch, jobs=jobs, timeout=timeout)


+def _files(corpus: Path, since: date | None) -> list[Path]:
+    return corpus_files(corpus, since, suffixes=_EXTS)
+
+
+def _read_records(path: Path) -> list[dict[str, Any]]:
+    rows: dict[str, dict[str, Any]] = {}
+    if not path.exists():
+        return []
+    with path.open(encoding="utf-8") as stream:
+        for line in stream:
+            try:
+                row = json.loads(line)
+            except (json.JSONDecodeError, UnicodeDecodeError):
+                continue
+            if isinstance(row, dict) and isinstance(row.get("path"), str):
+                rows[row["path"]] = row
+    return list(rows.values())
+
+
+def _repair_checkpoint(path: Path) -> None:
+    """Drop a torn tail before append, otherwise the next row joins invalid JSON."""
+    repair_jsonl_tail(path)
+
+
+def _discovery_candidates(
+    rows: list[dict[str, Any]], *, per_cohort: int, random_quiet: int, seed: int
+) -> list[dict[str, Any]]:
+    """Stratify quiet rows without treating provenance as a pixel-level label."""
+    rng = random.Random(seed)  # noqa: S311 -- reproducible review sampling
+    quiet = [row for row in rows if row.get("status") == "ok" and not row.get("keys")]
+    strata: dict[str, list[dict[str, Any]]] = collections.defaultdict(list)
+    for row in quiet:
+        if row.get("uscc"):
+            strata[f"tc260:{row['uscc']}"].append(row)
+        if row.get("platform"):
+            strata[f"platform:{row['platform']}"].append(row)
+
+    selected: dict[str, dict[str, Any]] = {}
+    labels: dict[str, set[str]] = collections.defaultdict(set)
+    for label, members in sorted(strata.items()):
+        for row in rng.sample(members, k=min(per_cohort, len(members))):
+            selected[row["path"]] = row
+            labels[row["path"]].add(label)
+
+    for row in rng.sample(quiet, k=min(random_quiet, len(quiet))):
+        selected[row["path"]] = row
+        labels[row["path"]].add("unbiased-quiet")
+
+    return [{**row, "discovery_strata": sorted(labels[path])} for path, row in sorted(selected.items())]
+
+
+def _bands(path: str) -> tuple[Any, Any] | None:
+    from remove_ai_watermarks.image_io import imread
+
+    image = imread(path)
+    if image is None:
+        return None
+    height = image.shape[0]
+    band_height = max(24, int(height * _SHEET_BAND_FRACTION))
+    return image[:band_height], image[height - band_height :]
+
+
+def _write_discovery_sheets(rows: list[dict[str, Any]], output: Path) -> int:
+    import csv
+
+    import cv2
+    import numpy as np
+
+    if output.exists() and any(output.iterdir()):
+        raise ValueError(f"discovery output must be empty: {output}")
+    output.mkdir(parents=True, exist_ok=True)
+    manifest: list[dict[str, Any]] = []
+    written = 0
+    for row in rows:
+        bands = _bands(row["path"])
+        if bands is None:
+            continue
+        top, bottom = bands
+        width = max(top.shape[1], bottom.shape[1])
+        divider = np.full((6, width, 3), 90, np.uint8)
+        header = np.full((28, width, 3), 40, np.uint8)
+        cv2.putText(header, f"{written:05d}", (8, 20), cv2.FONT_HERSHEY_SIMPLEX, 0.55, (255, 255, 255), 1)
+        cv2.imwrite(str(output / f"{written:05d}.png"), np.vstack((header, top, divider, bottom)))
+        manifest.append({"idx": written, **row})
+        written += 1
+    if manifest:
+        fields = list(manifest[0])
+        with (output / "MANIFEST_DO_NOT_OPEN.csv").open("w", newline="", encoding="utf-8") as stream:
+            writer = csv.DictWriter(stream, fieldnames=fields, extrasaction="ignore")
+            writer.writeheader()
+            writer.writerows(manifest)
+    return written
+
+
 def main() -> None:
    ap = argparse.ArgumentParser()
+    ap.add_argument("--corpus", type=Path, default=CORPUS)
+    ap.add_argument("--since", type=date.fromisoformat, help="include dated corpus directories on or after YYYY-MM-DD")
    ap.add_argument("--jobs", type=int, default=max(1, (os.cpu_count() or 4) - 2))
    ap.add_argument("--limit", type=int, default=0)
    ap.add_argument("--batch", type=int, default=200)
    ap.add_argument("--timeout", type=int, default=600, help="seconds per image before terminating its isolated worker")
    ap.add_argument("--out", type=Path, default=OUT)
    ap.add_argument("--restart", action="store_true")
+    ap.add_argument("--sheets", type=Path, help="write blinded discovery bands after the scan")
+    ap.add_argument("--per-cohort", type=int, default=12, help="quiet images sampled from each metadata cohort")
+    ap.add_argument("--random-quiet", type=int, default=400, help="unbiased quiet images sampled across the corpus")
+    ap.add_argument("--seed", type=int, default=2026)
    a = ap.parse_args()

-    files = sorted(p for p in glob.glob(str(CORPUS / "*" / "*")) if Path(p).suffix.lower() in _EXTS)
+    files = [str(path) for path in _files(a.corpus, a.since)]
    if a.limit:
        files = files[: a.limit]

    done: set[str] = set()
    if a.restart and a.out.exists():
        a.out.unlink()  # --restart must TRUNCATE; the file is reopened in append mode below
+    _repair_checkpoint(a.out)
    if a.out.exists() and not a.restart:
-        with open(a.out, encoding="utf-8") as fh:
-            for line in fh:
-                try:
-                    done.add(json.loads(line)["path"])
-                except Exception:  # noqa: S112 -- tolerate a torn last line
-                    continue
+        done = {row["path"] for row in _read_records(a.out) if row.get("schema_version") == _SCHEMA_VERSION}
    todo = [p for p in files if p not in done]
    print(f"images {len(files)}  already done {len(done)}  to do {len(todo)}  jobs {a.jobs}", flush=True)

-    counts: collections.Counter[str] = collections.Counter()
-    status: collections.Counter[str] = collections.Counter()
    seen = 0
+    a.out.parent.mkdir(parents=True, exist_ok=True)
    with open(a.out, "a", encoding="utf-8") as fh:
        for start in range(0, len(todo), a.batch):
            for rec in _run_batch(todo[start : start + a.batch], a.jobs, a.timeout):
                fh.write(json.dumps(rec) + "\n")
-                status[str(rec["status"])] += 1
-                for k in rec["keys"]:  # type: ignore[union-attr]
-                    counts[str(k)] += 1
                seen += 1
            fh.flush()
-            print(f"  {seen}/{len(todo)}  {dict(counts)}  {dict(status)}", flush=True)
+            print(f"  {seen}/{len(todo)}", flush=True)

-    # Rebuild the paths file from the FULL record set, not just this run's slice.
-    hits = []
-    with open(a.out, encoding="utf-8") as fh:
-        for line in fh:
-            try:
-                r = json.loads(line)
-            except Exception:  # noqa: S112
-                continue
-            if r.get("keys"):
-                hits.append(r["path"])
+    records = _read_records(a.out)
+    selected_paths = set(files)
+    records = [row for row in records if row["path"] in selected_paths]
+    hits = [row["path"] for row in records if row.get("keys")]
+    counts = collections.Counter(str(key) for row in records for key in row.get("keys", []))
+    strict_counts = collections.Counter(str(key) for row in records for key in row.get("strict_keys", []))
+    status = collections.Counter(str(row.get("status")) for row in records)
    # Derive the paths file from --out so a trial run with a scratch --out cannot
    # overwrite the shared list a full sweep produced.
    paths_out = a.out.with_suffix(".txt")
-    paths_out.write_text("\n".join(sorted(set(hits))) + "\n", encoding="utf-8")
+    paths_tmp = paths_out.with_suffix(paths_out.suffix + ".tmp")
+    paths_tmp.write_text("\n".join(sorted(set(hits))) + ("\n" if hits else ""), encoding="utf-8")
+    paths_tmp.replace(paths_out)
    print(f"\npositives: {len(set(hits))} images")
    for k, v in counts.most_common():
-        print(f"   {v:6d}  {k}")
+        print(f"   {v:6d}  {k} ({strict_counts[k]} strict)")
    print(f"statuses: {dict(status)}\npaths -> {paths_out}\nrecords -> {a.out}")
+    if a.sheets:
+        candidates = _discovery_candidates(
+            records,
+            per_cohort=a.per_cohort,
+            random_quiet=a.random_quiet,
+            seed=a.seed,
+        )
+        written = _write_discovery_sheets(candidates, a.sheets)
+        print(f"discovery candidates: {len(candidates)}  sheets written: {written} -> {a.sheets}")
+
+    from remove_ai_watermarks.watermark_registry import mark_keys
+
+    git = shutil.which("git")
+    commit = (
+        subprocess.run(  # noqa: S603 -- resolved git binary and fixed read-only arguments
+            [git, "rev-parse", "HEAD"],
+            cwd=REPO,
+            capture_output=True,
+            text=True,
+            check=False,
+        ).stdout.strip()
+        if git
+        else ""
+    )
+    metadata = {
+        "command": [str(Path(__file__).relative_to(REPO)), *sys.argv[1:]],
+        "library_commit": commit or None,
+        "library_version": version("remove-ai-watermarks"),
+        "corpus": str(a.corpus.resolve()),
+        "since": a.since.isoformat() if a.since else None,
+        "selected_images": len(files),
+        "complete_records": len(records),
+        "registry_keys": mark_keys(),
+        "schema_version": _SCHEMA_VERSION,
+    }
+    a.out.with_suffix(a.out.suffix + ".run.json").write_text(
+        json.dumps(metadata, indent=2, sort_keys=True) + "\n", encoding="utf-8"
+    )


 if __name__ == "__main__":
@@ -1,7 +1,7 @@
 {
  "name": "remove-ai-watermarks",
  "displayName": "Remove AI Watermarks",
-  "version": "1.0.8",
+  "version": "1.0.9",
  "description": "Identify and remove AI provenance marks from images and video the user generated or edited, using the remove-ai-watermarks CLI.",
  "author": {
    "name": "wiltodelta",
@@ -5,7 +5,7 @@ license: Apache-2.0
 compatibility: Python 3.11-3.14. Invisible image removal needs NVIDIA CUDA. Some video writes need ffmpeg. Installer may be uv, pipx, or pip.
 metadata:
  author: wiltodelta
-  version: "1.0.8"
+  version: "1.0.9"
  homepage: https://raiw.cc
  repository: https://github.com/wiltodelta/remove-ai-watermarks
 ---
@@ -89,11 +89,15 @@ run on.

 Pipeline profiles: `qwen-zimage` (the default), `sdxl-zimage`, `chroma-zimage`,
 `auto`. All are CUDA-only and all install from the same `qwen-zimage` extra.
-`auto` picks the engine from the file's own provenance -- chroma-zimage for
-Microsoft, qwen-zimage for OpenAI (since CLI 0.39.0), Google, Meta and unknown
+`auto` picks the engine from the file's own provenance -- sdxl-zimage for Google,
+chroma-zimage for Microsoft, qwen-zimage for OpenAI (since CLI 0.39.0), Meta and unknown
 -- so prefer it over naming an engine when the user has no reason to care. `sdxl-zimage` is the
 heavier alternative and `chroma-zimage` is the Apache-2.0 global stage.

+Verified-text manifests work with every profile (since CLI 0.40.5). Each profile
+uses its own already loaded VAE donor, so `auto` keeps the normal SDXL route for a
+Google image instead of switching engines.
+
 ```bash
 remove-ai-watermarks all image.png -o clean.png
 remove-ai-watermarks batch ./images --mode visible --output-dir ./clean
@@ -32,7 +32,7 @@ _GPU_MARKERS = ("GPU ", "gpu:")
 # told an agent its CLI was current right before it typed a value that build
 # rejects. tests/test_agent_skill.py pins it against the package version and
 # against every "(since CLI X.Y.Z)" annotation in the skill.
-MIN_CLI_VERSION = (0, 40, 0)
+MIN_CLI_VERSION = (0, 40, 5)
 # Markers a build without the pixel stack prints. The first is the current CLI's own
 # install hint; the rest are what older builds emit before the guard existed, and the
 # Homebrew formula ships exactly such a build.
@@ -14,6 +14,7 @@ High-level API (lazy, so ``import remove_ai_watermarks`` stays cheap)::
    raiw.remove_video_metadata("in.mp4", "out.mp4")     # verified metadata strip
    raiw.remove_video_invisible("in.mp4", "out.mp4")    # calibrated video-pixel regeneration
    raiw.remove_video_visible("in.mp4", "out.mp4")      # stable visible video-mark removal
+    raiw.classify_source("in.png")                      # source pipeline, not SynthID detection

 For a provenance verdict use the ``identify`` submodule::

@@ -33,7 +34,7 @@ _os.environ.setdefault("TRANSFORMERS_VERBOSITY", "error")
 _warnings.filterwarnings("ignore", message=r".*ImageProcessorFast.*")


-__version__ = "0.40.2"
+__version__ = "0.41.0"

 __all__ = [
    "BatchItemResult",
@@ -41,8 +42,10 @@ __all__ = [
    "InvisibleOptions",
    "MetadataStripIncomplete",
    "RemoveAllResult",
+    "SourceClassification",
    "VisibleRemovalResult",
    "__version__",
+    "classify_source",
    "identify_video",
    "inspect_video_metadata",
    "remove_all",
@@ -71,6 +74,7 @@ if TYPE_CHECKING:
        remove_visible_detailed,
        visible_provenance,
    )
+    from remove_ai_watermarks.source_classify import SourceClassification, classify_source
    from remove_ai_watermarks.video import (
        identify_video,
        inspect_video_metadata,
@@ -101,6 +105,10 @@ def __getattr__(name: str) -> object:
        from remove_ai_watermarks import api

        return getattr(api, name)
+    if name in ("SourceClassification", "classify_source"):
+        from remove_ai_watermarks import source_classify
+
+        return getattr(source_classify, name)
    if name in (
        "identify_video",
        "inspect_video_metadata",
@@ -18,6 +18,7 @@ from remove_ai_watermarks._internal.constants import (
    C2PA_AI_TOOLS,
    C2PA_AI_VENDORS,
    C2PA_CHUNK_TYPE,
+    C2PA_IDENTITY_AI_ORGS,
    C2PA_ISSUERS,
    C2PA_SIGNATURES,
    C2PA_SOFT_BINDING_REGISTRY,
@@ -529,7 +530,8 @@ def _structured_manifest_fields(store: dict[str, Any]) -> dict[str, Any]:
        info["ai_tool"] = ", ".join(tools)
    if actions:
        info["actions"] = ", ".join(actions)
-    if claim_generator_asserts_ai:
+    signer_identity_asserts_ai = not C2PA_IDENTITY_AI_ORGS.isdisjoint(issuers)
+    if claim_generator_asserts_ai or signer_identity_asserts_ai:
        info["c2pa_identity_ai"] = True

    generated = any(
@@ -36,7 +36,7 @@ from remove_ai_watermarks._internal.two_stage_pipeline import (
    _prompt_cache_path,
    _store_prompt_payload,
    _target_size,
-    edge_pad_to_grid,
+    diffusers_vae_roundtrip,
    requested_steps,
 )

@@ -182,28 +182,12 @@ class ChromaZImagePipeline(TwoStageZImagePipeline):

    def _vae_roundtrip(self, image: Image.Image) -> Image.Image:
        """Reconstruct source pixels through the already loaded Chroma VAE."""
-        import torch
-
        pipe = self._load_global()
-        source_width, source_height = image.size
-        padded = edge_pad_to_grid(image, _LATENT_GRID)
-
-        tensor = pipe.image_processor.preprocess(
-            padded,
-            height=padded.height,
-            width=padded.width,
-        ).to(device=self.device, dtype=self.torch_dtype)
-        with torch.inference_mode():
-            encoded = pipe.vae.encode(tensor)
-            if hasattr(encoded, "latent_dist"):
-                # A restoration donor must be deterministic. Diffusers samples this
-                # distribution for img2img noise initialization; its mode is the
-                # faithful VAE reconstruction needed here.
-                latents = encoded.latent_dist.mode()
-            elif hasattr(encoded, "latents"):
-                latents = encoded.latents
-            else:
-                raise AttributeError("Chroma VAE encoder output contains no latents")
-            decoded = pipe.vae.decode(latents, return_dict=False)[0]
-        reconstructed = pipe.image_processor.postprocess(decoded, output_type="pil")[0]
-        return reconstructed.crop((0, 0, source_width, source_height)).convert("RGB")
+        return diffusers_vae_roundtrip(
+            pipe,
+            image,
+            grid=_LATENT_GRID,
+            device=self.device,
+            dtype=self.torch_dtype,
+            profile_name=self.profile_name,
+        )
@@ -32,7 +32,7 @@ C2PA_SIGNATURES = tuple(

@dataclass(frozen=True, slots=True)
 class C2paAiVendor:
-    """One issuer signature and its normalized product attribution."""
+    """One C2PA identity and its AI-product or neutral-signer attribution."""

    issuer: bytes
    org: str
@@ -41,6 +41,8 @@ class C2paAiVendor:
    synthid: bool = False
    asserts_ai: bool = False
    synthid_requires_watermark_action: bool = False
+    signer_platform: str | None = None
+    raw_signer_platforms: tuple[tuple[bytes, str], ...] = ()


 def _vendor(
@@ -52,6 +54,8 @@ def _vendor(
    synthid: bool = False,
    asserts_ai: bool = False,
    synthid_requires_watermark_action: bool = False,
+    signer_platform: str | None = None,
+    raw_signer_platforms: tuple[tuple[bytes, str], ...] = (),
 ) -> C2paAiVendor:
    token = issuer.encode() if isinstance(issuer, str) else issuer
    return C2paAiVendor(
@@ -62,6 +66,8 @@ def _vendor(
        synthid=synthid,
        asserts_ai=asserts_ai,
        synthid_requires_watermark_action=synthid_requires_watermark_action,
+        signer_platform=signer_platform,
+        raw_signer_platforms=raw_signer_platforms,
    )


@@ -109,11 +115,68 @@ C2PA_AI_VENDORS: tuple[C2paAiVendor, ...] = (
    # "Bria Artificial Intelligence"). Found as an unmapped signer on 4 corpus
    # uploads 2026-08-08 that identify reported as unknown-signer C2PA.
    _vendor(b"Ideogram", "Ideogram", "Ideogram", "Ideogram", asserts_ai=True),
+    _vendor(b"xAI Grok Imagine", "xAI Grok Imagine", "xAI Grok Imagine", "xAI", asserts_ai=True),
+    _vendor(b"Producer.ai", "Producer.ai", "Producer.ai", "Producer.ai", asserts_ai=True),
+    # This signature names only SPRING's legal entity, not a product. Preserve
+    # that signer identity without guessing which product produced a file.
+    _vendor(b"SPRING (SG) PTE. LTD.", "SPRING (SG) PTE. LTD.", "SPRING (SG) PTE. LTD.", "SPRING"),
+    # These companies also sign ordinary edits and publications. Register the
+    # signer identity for provenance display, but provide no AI platform and do
+    # not let the identity assert an AI verdict.
+    _vendor(
+        b"TikTok Inc.",
+        "TikTok",
+        None,
+        None,
+        signer_platform="TikTok (C2PA signer)",
+    ),
+    _vendor(
+        b"Bytedance Pte",
+        "ByteDance",
+        None,
+        None,
+        signer_platform="ByteDance (C2PA signer)",
+        raw_signer_platforms=(
+            (b"Bytedance Pte. Ltd.", "ByteDance (C2PA signer)"),
+            (b"CapCut/", "CapCut (C2PA signer)"),
+            (b"capcut c2pa-rs", "CapCut (C2PA signer)"),
+        ),
+    ),
+    _vendor(
+        b"Anthropic Claude Content Signing",
+        "Anthropic Claude",
+        None,
+        None,
+        signer_platform="Anthropic Claude (C2PA signer)",
+    ),
+    _vendor(
+        b"Samsung Galaxy",
+        "Samsung Galaxy",
+        None,
+        None,
+        signer_platform="Samsung Galaxy (C2PA)",
+    ),
+    _vendor(
+        b"com.asus.gallery",
+        "ASUS Gallery",
+        None,
+        None,
+        signer_platform="ASUS Gallery (C2PA signer)",
+    ),
    _vendor(b"Truepic", "Truepic", None, None),
 )

 C2PA_ISSUERS = {vendor.issuer: vendor.org for vendor in C2PA_AI_VENDORS}
 C2PA_IDENTITY_AI_ORGS = frozenset(vendor.org for vendor in C2PA_AI_VENDORS if vendor.asserts_ai)
+C2PA_SIGNER_PLATFORM_BY_ORG = {
+    vendor.org: vendor.signer_platform for vendor in C2PA_AI_VENDORS if vendor.signer_platform is not None
+}
+C2PA_SIGNER_PLATFORMS = tuple(
+    pair
+    for vendor in C2PA_AI_VENDORS
+    if vendor.signer_platform is not None
+    for pair in (vendor.raw_signer_platforms or ((vendor.issuer, vendor.signer_platform),))
+)

 # Product-specific claim generators can sign through a different upstream issuer.
 # Keep this attribution beside the issuer registry so every C2PA consumer has one
@@ -128,6 +191,8 @@ C2PA_CLAIM_GENERATOR_PLATFORMS: tuple[tuple[str, str], ...] = (
    ("recraft.ai", "Recraft"),
    ("topaz labs image api", "Topaz Labs"),
    ("tiktok ad creative toolbox", "TikTok Ad Creative Toolbox"),
+    ("fastvid", "FastVid"),
+    ("capcut", "CapCut (C2PA signer)"),
 )

 C2PA_AI_TOOLS = {
@@ -0,0 +1,14 @@
+"""Small shared numerical transforms used by independent pixel analyzers."""
+
+from __future__ import annotations
+
+from typing import Any
+
+
+def dct_matrix(np: Any, size: int = 8) -> Any:
+    """Return an orthonormal ``size`` by ``size`` DCT-II basis."""
+    frequencies = np.arange(size)[:, None]
+    samples = np.arange(size)[None, :]
+    matrix = np.cos(np.pi * (2 * samples + 1) * frequencies / (2 * size))
+    matrix[0, :] *= 1 / np.sqrt(2)
+    return matrix * np.sqrt(2 / size)
@@ -28,6 +28,7 @@ from remove_ai_watermarks._internal.two_stage_pipeline import (
    TwoStageZImagePipeline,
    _target_size,
    build_canny_control_image,
+    diffusers_vae_roundtrip,
    requested_steps,
 )
 from remove_ai_watermarks._internal.watermark_profiles import (
@@ -120,3 +121,15 @@ class SdxlZImagePipeline(TwoStageZImagePipeline):
        if result.size != image.size:
            result = result.resize(image.size, Image.Resampling.LANCZOS)
        return result.convert("RGB")
+
+    def _vae_roundtrip(self, image: Image.Image) -> Image.Image:
+        """Reconstruct source pixels through the already loaded SDXL VAE."""
+        pipe = self._load_global()
+        return diffusers_vae_roundtrip(
+            pipe,
+            image,
+            grid=_LATENT_GRID,
+            device=self.device,
+            dtype=self.torch_dtype,
+            profile_name=self.profile_name,
+        )
@@ -0,0 +1,78 @@
+"""Phase-free spectral features for source/export-pipeline classification."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from functools import lru_cache
+from typing import TYPE_CHECKING
+
+import numpy as np
+from PIL import Image
+
+if TYPE_CHECKING:
+    from numpy.typing import NDArray
+
+
+@dataclass(frozen=True)
+class SpectralGeometry:
+    """Precomputed coordinates for one phase-free spectral representation."""
+
+    size: int
+    grid: int
+    radius: NDArray[np.int32]
+    radius_counts: NDArray[np.int64]
+    passband: NDArray[np.bool_]
+
+
+@lru_cache(maxsize=4)
+def make_geometry(size: int = 256, grid: int = 16) -> SpectralGeometry:
+    """Build radial detrending and pooling geometry."""
+    if size < 16 or grid < 2 or size % grid:
+        raise ValueError("size must be at least 16 and divisible by grid")
+    yy, xx = np.mgrid[:size, :size]
+    radius = np.rint(np.hypot(yy - size // 2, xx - size // 2)).astype(np.int32)
+    radius_counts = np.bincount(radius.ravel())
+    passband = (radius >= 4) & (radius <= size * 0.48)
+    return SpectralGeometry(size, grid, radius, radius_counts, passband)
+
+
+def phase_free_log_spectrum(channel: NDArray[np.float64]) -> NDArray[np.float64]:
+    """Return centered log FFT magnitude, deliberately discarding phase."""
+    if channel.ndim != 2:
+        raise ValueError("channel must be a two-dimensional array")
+    centered = channel - float(np.mean(channel))
+    return np.log1p(np.abs(np.fft.fftshift(np.fft.fft2(centered))))
+
+
+def spectral_feature(pixels: NDArray[np.float64], geometry: SpectralGeometry) -> NDArray[np.float64]:
+    """Extract pooled radial-residual spectra from luma and opponent color."""
+    expected = (geometry.size, geometry.size, 3)
+    if pixels.shape != expected:
+        raise ValueError(f"pixels must have shape {expected}, got {pixels.shape}")
+    red, green, blue = np.moveaxis(pixels, 2, 0)
+    channels = (
+        0.299 * red + 0.587 * green + 0.114 * blue,
+        red - green,
+        blue - (red + green) / 2.0,
+    )
+    block = geometry.size // geometry.grid
+    features: list[NDArray[np.float64]] = []
+    for channel in channels:
+        spectrum = phase_free_log_spectrum(channel)
+        radial_sum = np.bincount(geometry.radius.ravel(), weights=spectrum.ravel())
+        radial_mean = radial_sum / geometry.radius_counts
+        residual = np.where(
+            geometry.passband,
+            spectrum - radial_mean[geometry.radius],
+            0.0,
+        )
+        pooled = residual.reshape(geometry.grid, block, geometry.grid, block).mean(axis=(1, 3))
+        features.append(pooled.ravel())
+    return np.concatenate(features)
+
+
+def feature_from_image(image: Image.Image, geometry: SpectralGeometry) -> NDArray[np.float64]:
+    """Resize an RGB image canonically and extract its phase-free feature."""
+    canonical = image.resize((geometry.size, geometry.size), Image.Resampling.LANCZOS)
+    pixels = np.asarray(canonical, dtype=np.float64) / 255.0
+    return spectral_feature(pixels, geometry)
@@ -96,6 +96,41 @@ def edge_pad_to_grid(image: Image.Image, grid: int) -> Image.Image:
    )


+def diffusers_vae_roundtrip(
+    pipe: Any,
+    image: Image.Image,
+    *,
+    grid: int,
+    device: Any,
+    dtype: Any,
+    profile_name: str,
+) -> Image.Image:
+    """Reconstruct pixels deterministically through a Diffusers pipeline VAE."""
+    import torch
+
+    source_width, source_height = image.size
+    padded = edge_pad_to_grid(image, grid)
+    tensor = pipe.image_processor.preprocess(
+        padded,
+        height=padded.height,
+        width=padded.width,
+    ).to(device=device, dtype=dtype)
+    with torch.inference_mode():
+        encoded = pipe.vae.encode(tensor)
+        if hasattr(encoded, "latent_dist"):
+            # A restoration donor must be deterministic. Diffusers samples this
+            # distribution for img2img noise initialization; its mode is the
+            # faithful VAE reconstruction needed here.
+            latents = encoded.latent_dist.mode()
+        elif hasattr(encoded, "latents"):
+            latents = encoded.latents
+        else:
+            raise AttributeError(f"{profile_name} VAE encoder output contains no latents")
+        decoded = pipe.vae.decode(latents, return_dict=False)[0]
+    reconstructed = pipe.image_processor.postprocess(decoded, output_type="pil")[0]
+    return reconstructed.crop((0, 0, source_width, source_height)).convert("RGB")
+
+
 def resolve_face_model_residency(
    requested: bool | None,
    *,
@@ -55,9 +55,8 @@ PROFILE_ADAPTIVE_POLISH = {
    QWEN_ZIMAGE_PROFILE: False,
    SDXL_ZIMAGE_PROFILE: True,
    CHROMA_ZIMAGE_PROFILE: False,
-    # Auto resolves to qwen-zimage or chroma-zimage, both of which keep the
-    # input's detail level. If it ever resolves to an engine that needs polish,
-    # the resolved profile's own default applies.
+    # This fallback applies only before a vendor is known. The engine resolves
+    # auto first so Google-to-SDXL receives SDXL's enabled default.
    AUTO_PROFILE: False,
 }

@@ -80,26 +79,26 @@ SDXL_LIGHTNING_PATTERN = "sdxl_lightning_4step_lora.safetensors"
 PROFILE_SEED = 0

 # sdxl-zimage runs the qwen-zimage recipe on an SDXL global stage, and strength is
-# architecture-bound: at Qwen's 0.154 an SDXL global pass leaves SynthID on a native
-# 2816x1536 Gemini original, while 0.20, 0.25 and 0.30 all read clean in the Gemini
-# app. 0.25 keeps a rung of margin over that boundary, which the historical SDXL
-# certification argues for -- it recorded 0.20 as DETECTED against Gemini on an
-# older SDXL pipeline. OpenAI is the easier oracle: the profile already cleared
-# openai.com/verify at 0.1102, so 0.15 sits above what was verified rather than on
-# it. Unknown follows Gemini, the stricter of the two.
+# architecture-bound. The source-fresh 2026-09-15 calibration found Google CJK
+# detectable through 0.45 and clean at 0.50. The 0.50 candidate was then replicated
+# clean on independent busy, flat-text, and face-plus-text carriers. OpenAI is
+# the easier oracle: the profile cleared openai.com/verify at 0.1102, so 0.15
+# sits above what was verified rather than on it. Unknown follows Google, the
+# stricter measured cohort.
 #
 # Unlike qwen-zimage this is a flat vendor policy rather than a resolution curve,
 # because flat values are what was measured. Every verdict above comes from a fixed
 # strength at one size; no size dependence has been established for this stage.
 SDXL_ZIMAGE_OPENAI_STRENGTH = 0.15
-SDXL_ZIMAGE_GEMINI_STRENGTH = 0.25
+SDXL_ZIMAGE_GEMINI_STRENGTH = 0.50
 SDXL_ZIMAGE_UNKNOWN_STRENGTH = SDXL_ZIMAGE_GEMINI_STRENGTH

 # qwen-zimage keeps its resolution curve for unknown content, but measured vendor
-# cohorts bypass it. Google remained detectable through 0.24375; 0.25 cleared all
-# three valid sources and 0.27 was separately repeated clean across them and three
-# accounts, so the independently checked 0.27 candidate is the operating floor.
-QWEN_ZIMAGE_GOOGLE_STRENGTH = 0.27
+# cohorts bypass it. A source-fresh 2026-09-15 sweep found Latin scenes and faces
+# clean at 0.15 but CJK and Cyrillic text requiring 0.35. The 0.35 candidate was
+# then independently replicated clean on busy, flat-text, and face-plus-text CJK
+# carriers, so 0.35 is the content-agnostic operating floor.
+QWEN_ZIMAGE_GOOGLE_STRENGTH = 0.35

 # The two OpenAI sources first cleared at 0.06225 and 0.0695. Add one full observed
 # cross-source spread (0.00725) to the worst clean boundary: 0.0695 + 0.00725.
@@ -122,7 +121,7 @@ QWEN_ZIMAGE_MICROSOFT_STRENGTH = 0.15
 # mug <= 0.03, text <= 0.015. Worst clean boundary plus one full observed
 # cross-source spread: 0.06 + (0.0525 - 0.015) = 0.0975, rounded up to 0.1.
 # sdxl-zimage has no measured Meta floor; its vendor map stays without a meta
-# entry so an explicit --vendor meta there falls to the unknown 0.25, which is
+# entry so an explicit --vendor meta there falls to the unknown 0.50, which is
 # above this floor and therefore conservative.
 QWEN_ZIMAGE_META_STRENGTH = 0.1

@@ -165,11 +164,12 @@ _SDXL_ZIMAGE_STRENGTH_BY_VENDOR: dict[str, float] = {
 # - Microsoft InvisMark: paint-1 (0.06, 0.08], paint-2 <= 0.04, paint-3
 #   (0.06, 0.08] -> 0.08 + (0.08 - 0.04) = 0.12, rounded up to the measured
 #   rung 0.125, oracle-verified clean on both worst sources. BELOW qwen's 0.15.
-# - Google: 633uuy (0.20, 0.25], akdbei (0.20, 0.25], y48j3c (0.08, 0.12],
-#   3mc4t9 (0.08, 0.12] -> 0.25 + (0.25 - 0.12) = 0.38, rounded up to 0.40,
-#   oracle-verified clean on the worst fixture. ABOVE qwen's 0.27; at 0.40 the
-#   regeneration destroys dense text and collapses face identity, which is why
-#   the matched-strength addendum recommends the adaptive-strength follow-up.
+# - Google: the source-fresh 2026-09-15 sweep superseded the four-fixture 0.40
+#   derivation: CJK remained detected at 0.40 and cleared at 0.50. The 0.50
+#   candidate was independently replicated clean on busy, flat-text, and
+#   face-plus-text CJK carriers. Face detection cannot select the older 0.125
+#   arm safely because fresh face-plus-text carriers require more denoise and the
+#   runtime has no measured text-content gate.
 # - Meta Content Seal: lighthouse (0.08, 0.10], fox (0.06, 0.08], night_city
 #   (0.045, 0.06], studio_mug (0.03, 0.045], text_poster <= 0.03 ->
 #   0.10 + (0.10 - 0.03) = 0.17. ABOVE qwen's 0.1; the floors exist because
@@ -177,23 +177,10 @@ _SDXL_ZIMAGE_STRENGTH_BY_VENDOR: dict[str, float] = {
 #   worse per unit strength.
 CHROMA_ZIMAGE_OPENAI_STRENGTH = 0.20
 CHROMA_ZIMAGE_MICROSOFT_STRENGTH = 0.125
-CHROMA_ZIMAGE_GOOGLE_STRENGTH = 0.40
+CHROMA_ZIMAGE_GOOGLE_STRENGTH = 0.50
 CHROMA_ZIMAGE_META_STRENGTH = 0.17
 CHROMA_ZIMAGE_UNKNOWN_STRENGTH = CHROMA_ZIMAGE_GOOGLE_STRENGTH

-# Content-adaptive Google floor: the four-fixture calibration showed a clean
-# face-count split. Both zero-face fixtures (dense-text cards 633uuy and akdbei)
-# need 0.25 first-clean; both face fixtures (y48j3c with 17 detected faces and
-# 3mc4t9 with 7) clear at 0.12. The flat 0.40 floor is the zero-face policy
-# (worst 0.25 + spread 0.13 = 0.38, rounded to 0.40). Face content can use the
-# measured 0.125 rung instead: both face fixtures first-cleaned at 0.12, and
-# with identical boundaries the cross-source spread is zero, so 0.125 (the next
-# measured rung above 0.12) is the operating point. This split is
-# Google-SynthID-specific: OpenAI's face fixture was HARDER than its text
-# fixture (0.075 vs 0.06), so no other cohort gets an adaptive arm.
-# Oracle-verified 2026-08-30; see docs/chroma1-engine-research.md.
-CHROMA_ZIMAGE_GOOGLE_FACE_STRENGTH = 0.125
-
 _CHROMA_ZIMAGE_STRENGTH_BY_VENDOR: dict[str, float] = {
    "openai": CHROMA_ZIMAGE_OPENAI_STRENGTH,
    "google": CHROMA_ZIMAGE_GOOGLE_STRENGTH,
@@ -209,10 +196,11 @@ _ALIASES = {

 # The deterministic per-cohort selection policy: which profile wins on which vendor, from the
 # 2026-08-29/30 four-cohort calibration (docs/chroma1-engine-research.md).
-# chroma-zimage has the lower measured Microsoft floor. qwen-zimage wins on
-# OpenAI, Google, and Meta. Unknown stays on qwen (the shipped default,
-# conservative).
+# chroma-zimage has the lower measured Microsoft floor. SDXL wins Google on
+# paired quality at the independently clean operating points. Qwen wins OpenAI
+# and Meta; unknown stays on Qwen, the shipped conservative fallback.
 _ENGINE_BY_VENDOR: dict[str, str] = {
+    "google": SDXL_ZIMAGE_PROFILE,
    "microsoft": CHROMA_ZIMAGE_PROFILE,
 }

@@ -228,6 +216,14 @@ def normalize_profile(profile: str) -> str:
    return _ALIASES.get(value, value)


+def resolve_effective_profile(profile: str, vendor: str | None) -> str:
+    """Resolve a configured profile to the concrete engine for one image."""
+    normalized = normalize_profile(profile)
+    if normalized == AUTO_PROFILE:
+        return resolve_auto_profile(vendor)
+    return normalized
+
+
 def resolve_seed(seed: int | None) -> int:
    """Keep every profile reproducible by default."""
    return PROFILE_SEED if seed is None else seed
@@ -247,9 +243,7 @@ def global_offload_supported(pipeline: str) -> bool:
    the only unresolved ``auto`` load, and that path deliberately uses qwen-zimage
    as its fallback, so it has the same offload support as qwen-zimage.
    """
-    profile = normalize_profile(pipeline)
-    if profile == AUTO_PROFILE:
-        profile = resolve_auto_profile(None)
+    profile = resolve_effective_profile(pipeline, None)
    return profile in GLOBAL_OFFLOAD_PROFILES


@@ -273,7 +267,6 @@ def resolve_strength(
    pipeline: str | None = None,
    *,
    size: tuple[int, int] | None = None,
-    face_count: int | None = None,
 ) -> float:
    """Resolve a user override or the calibrated policy for a profile and vendor.

@@ -288,13 +281,11 @@ def resolve_strength(
    """
    if strength is not None:
        return strength
-    normalized = normalize_profile(pipeline or "")
+    normalized = resolve_effective_profile(pipeline or "", vendor)
    if normalized == SDXL_ZIMAGE_PROFILE:
        return _SDXL_ZIMAGE_STRENGTH_BY_VENDOR.get((vendor or "").casefold(), SDXL_ZIMAGE_UNKNOWN_STRENGTH)
    if normalized == CHROMA_ZIMAGE_PROFILE:
        vendor_key = (vendor or "").casefold()
-        if vendor_key == "google" and face_count is not None and face_count > 0:
-            return CHROMA_ZIMAGE_GOOGLE_FACE_STRENGTH
        return _CHROMA_ZIMAGE_STRENGTH_BY_VENDOR.get(vendor_key, CHROMA_ZIMAGE_UNKNOWN_STRENGTH)
    if size is None:
        raise ValueError("qwen-zimage resolves strength from image area, so size is required")
@@ -11,7 +11,6 @@ from typing import TYPE_CHECKING, Any
 from PIL import Image

 from remove_ai_watermarks._internal.watermark_profiles import (
-    AUTO_PROFILE,
    CHROMA_ZIMAGE_PROFILE,
    DEFAULT_PROFILE,
    INVISIBLE_EXTRA,
@@ -21,7 +20,7 @@ from remove_ai_watermarks._internal.watermark_profiles import (
    SDXL_ZIMAGE_PROFILE,
    global_offload_supported,
    normalize_profile,
-    resolve_auto_profile,
+    resolve_effective_profile,
    resolve_seed,
    resolve_strength,
 )
@@ -43,6 +42,14 @@ except ImportError:
    torch = None  # type: ignore[assignment]
    _HAS_TORCH = False

+
+def _dtype_for_profile(profile: str) -> Any:
+    """Return the weight dtype bound to a concrete regeneration profile."""
+    if profile == SDXL_ZIMAGE_PROFILE:
+        return torch.float16  # type: ignore[union-attr]
+    return torch.bfloat16  # type: ignore[union-attr]
+
+
 # Probed once at import. ``torch`` is imported above rather than probed because this
 # module needs the object, not just the answer.
 _HAS_REMOVAL_MODULES = module_available(*(name for name in REMOVAL_MODULES if name != "torch"))
@@ -111,13 +118,14 @@ class WatermarkRemover:
        controlnet_conditioning_scale: float = 1.0,
        cpu_offload: bool = False,
    ) -> None:
-        self.model_profile = normalize_profile(pipeline)
-        if self.model_profile not in PROFILE_CHOICES:
+        self.configured_profile = normalize_profile(pipeline)
+        if self.configured_profile not in PROFILE_CHOICES:
            raise ValueError(f"Unsupported pipeline '{pipeline}'. Use one of: {', '.join(PROFILE_CHOICES)}.")
-        # The auto profile resolves to a concrete engine per-image in
-        # remove_watermark, once the provenance vendor is known. It never
-        # resolves to sdxl-zimage, so the dtype below is correct either way.
-        self._auto = self.model_profile == AUTO_PROFILE
+        # Auto starts on the concrete Qwen fallback for preload, then resolves
+        # from the immutable configured profile for every image. Keeping the
+        # configured and active profiles separate prevents one image's vendor
+        # from becoming the next image's policy input.
+        self.model_profile = resolve_effective_profile(self.configured_profile, None)
        # There is no ``model_id`` parameter and no ``model_id`` attribute: each
        # profile pins a fixed model stack, and the dtype below is bound to that
        # stack's weights. Both used to be constructor overrides that existed only to
@@ -136,13 +144,10 @@ class WatermarkRemover:
                "every identify command still run on CPU."
            )

-        if self.model_profile == SDXL_ZIMAGE_PROFILE:
-            # SDXL ships fp16 weights and an fp16-safe VAE; bf16 would give up the
-            # variant without buying anything on this architecture.
-            self.torch_dtype = torch.float16  # type: ignore[union-attr]
-        else:
-            # qwen-zimage, chroma-zimage, and auto all resolve to bf16 engines.
-            self.torch_dtype = torch.bfloat16  # type: ignore[union-attr]
+        # SDXL ships fp16 weights and an fp16-safe VAE. Qwen and Chroma use
+        # bf16; auto starts on the concrete Qwen fallback and switches together
+        # with its active profile per image.
+        self.torch_dtype = _dtype_for_profile(self.model_profile)

        self.cpu_offload = cpu_offload
        self.controlnet_conditioning_scale = controlnet_conditioning_scale
@@ -212,26 +217,6 @@ class WatermarkRemover:

        remove_ai_metadata(output_path, output_path, keep_standard=True)

-    def _resolve_chroma_strength(
-        self,
-        strength: float | None,
-        vendor: str | None,
-        source: Image.Image,
-    ) -> float:
-        """Resolve strength, running face detection for the chroma Google arm.
-
-        The content-adaptive Google floor needs a face count. YuNet is fast
-        (~50 ms) and the model download is already managed by the shared base.
-        """
-        if self.model_profile != CHROMA_ZIMAGE_PROFILE or (vendor or "").casefold() != "google":
-            return resolve_strength(strength, vendor, self.model_profile, size=source.size)
-        if strength is not None:
-            return strength
-        from remove_ai_watermarks._internal.two_stage_pipeline import detect_faces
-
-        face_count = len(detect_faces(source))
-        return resolve_strength(strength, vendor, self.model_profile, size=source.size, face_count=face_count)
-
    def remove_watermark(
        self,
        image_path: Path,
@@ -257,21 +242,19 @@ class WatermarkRemover:
        with Image.open(image_path) as opened:
            source = opened.convert("RGB")

-        # The auto policy resolves per-image once the provenance vendor is
-        # known, BEFORE the strength resolution so the right engine's floors
-        # are used. If the vendor changed the engine, reset the cached pipeline.
-        if self._auto:
-            resolved = resolve_auto_profile(vendor)
-            if resolved != self.model_profile:
-                self.model_profile = resolved
-                self._qwen_zimage_pipeline = None
-        if text_manifest is not None and self.model_profile == SDXL_ZIMAGE_PROFILE:
-            raise ValueError("Verified text restoration is not supported by the sdxl-zimage profile")
+        # Resolve from the configured policy for every image, before strength
+        # resolution, so a previous image's vendor cannot pin the next one to
+        # its engine. Reset cached weights only when the concrete engine changes.
+        resolved = resolve_effective_profile(self.configured_profile, vendor)
+        if resolved != self.model_profile:
+            self.model_profile = resolved
+            self.torch_dtype = _dtype_for_profile(resolved)
+            self._qwen_zimage_pipeline = None
        if text_manifest is not None and tile:
            raise ValueError("Verified text restoration is not calibrated with tiled diffusion")
        if fidelity_anchor and self.model_profile != QWEN_ZIMAGE_PROFILE:
            raise ValueError("The fidelity anchor is supported only by the qwen-zimage profile")
-        resolved_strength = self._resolve_chroma_strength(strength, vendor, source)
+        resolved_strength = resolve_strength(strength, vendor, self.model_profile, size=source.size)
        if not 0.0 <= resolved_strength <= 1.0:
            raise ValueError(f"Strength must be between 0.0 and 1.0, got {resolved_strength}")

@@ -227,8 +227,8 @@ _PIPELINE_HELP = (
    "followed by SAM-masked Z-Image face repair; sdxl-zimage = the same recipe and the "
    "same face stage on an SDXL global pass, which needs more denoise; chroma-zimage = "
    "the same face stage on an Apache-2.0 Chroma1 global pass; auto = pick the engine "
-    "from the provenance (chroma-zimage for Microsoft, qwen-zimage for "
-    "OpenAI/Google/Meta/unknown). "
+    "from the provenance (sdxl-zimage for Google, chroma-zimage for Microsoft, "
+    "qwen-zimage for OpenAI/Meta/unknown). "
    "All are CUDA-ONLY -- install the qwen-zimage extra. There is no CPU or MPS profile for "
    "invisible-watermark removal."
 )
@@ -318,9 +318,8 @@ _text_manifest_option = click.option(
    type=click.Path(exists=True, dir_okay=False, path_type=Path),
    default=None,
    help=(
-        "Experimental verified-text restoration manifest. Requires qwen-zimage or "
-        "chroma-zimage, the text-restoration extra, native untiled geometry, and no "
-        "postprocessing."
+        "Experimental verified-text restoration manifest. Requires the text-restoration "
+        "extra, native untiled geometry, and no postprocessing."
    ),
 )

@@ -44,6 +44,8 @@ from remove_ai_watermarks._internal.constants import (
    C2PA_CLAIM_GENERATOR_PLATFORMS,
    C2PA_IDENTITY_AI_ORGS,
    C2PA_ISSUERS,
+    C2PA_SIGNER_PLATFORM_BY_ORG,
+    C2PA_SIGNER_PLATFORMS,
 )
 from remove_ai_watermarks._internal.schema import require_schema_version
 from remove_ai_watermarks.metadata import (
@@ -631,7 +633,8 @@ def _ai_tools_in(data: bytes) -> list[str]:
 # Only tokens verified against a real signed file are listed (Leica, Nikon,
 # Sony, Truepic, Google Pixel); add Canon/Bria as real samples are captured.
 # Samsung Galaxy is an AI-capable editing device, not a pure-capture camera, so
-# it lives in `_SIGNER_C2PA_PLATFORM` below (it must not feed the camera clash).
+# it lives in the neutral `C2PA_SIGNER_PLATFORMS` registry (it must not feed the
+# camera clash).
 _DEVICE_C2PA_PLATFORM: tuple[tuple[bytes, str], ...] = (
    (b"lc_c2pa", "Leica (camera, C2PA capture)"),
    (b"Leica Camera", "Leica (camera, C2PA capture)"),
@@ -646,6 +649,8 @@ _DEVICE_C2PA_PLATFORM: tuple[tuple[bytes, str], ...] = (
    # Truepic is a C2PA signing authority whose name appears in the trust chain
    # of unrelated manifests (e.g. OpenAI), so the bare token mis-attributes.
    (b"Truepic_Lens", "Truepic Lens (verified capture)"),
+    (b"vivo Camera", "vivo (camera, C2PA capture)"),
+    (b"vivo X300 Pro", "vivo (camera, C2PA capture)"),
 )


@@ -743,15 +748,14 @@ def _device_platform(head: bytes) -> str | None:
 #     (distinct from the EXIF "SM-xxxx" model string on ordinary Samsung photos).
 #   com.asus.gallery -- ASUS Gallery claim_generator (a C2PA-signed edit, no AI
 #     source type or genAIType on the samples, so it never asserts is_ai).
-_SIGNER_C2PA_PLATFORM: tuple[tuple[bytes, str], ...] = (
-    (b"Samsung Galaxy", "Samsung Galaxy (C2PA)"),
-    (b"com.asus.gallery", "ASUS Gallery (C2PA signer)"),
-)
-
-
-def _signer_platform(head: bytes) -> str | None:
-    """Map a C2PA editing-app / AI-capable-device signer token to a platform."""
-    return _first_token_match(head, _SIGNER_C2PA_PLATFORM)
+def _signer_platform(head: bytes, issuers: list[str]) -> str | None:
+    """Map a raw or normalized C2PA signer identity to a neutral platform."""
+    raw = _first_token_match(head, C2PA_SIGNER_PLATFORMS)
+    normalized = next(
+        (C2PA_SIGNER_PLATFORM_BY_ORG[org] for org in issuers if org in C2PA_SIGNER_PLATFORM_BY_ORG),
+        None,
+    )
+    return raw or normalized


 def _attribute_platform(issuers: list[str], *, is_ai: bool = True) -> str | None:
@@ -1178,7 +1182,6 @@ def _identify_from_evidence(
    # below keeps the full buffer: their markers are long and distinctive.
    region = _metadata_region(head)
    camera_label = _device_platform(region)
-    signer_label = _signer_platform(region)

    # ── C2PA Content Credentials ────────────────────────────────────
    has_c2pa = bool(info) or c2pa_marker_in(head)
@@ -1189,6 +1192,7 @@ def _identify_from_evidence(
    # by substring made the displayed reason a second, looser rule than the verdict.
    failed_c2pa_codes = [str(code) for code in cast("list[object]", info.get("c2pa_failed_codes", []))]
    issuers = [info["issuer"]] if info.get("issuer") else _issuers_in(region)
+    signer_label = _signer_platform(region, issuers)
    # Full AI generation (trainedAlgorithmicMedia) vs an AI-enhanced real photo
    # (compositeWithTrainedAlgorithmicMedia). The structured kind is parsed once in
    # _internal.c2pa._structured_manifest_fields (covers PNG + any container the c2pa-python
@@ -1216,15 +1220,18 @@ def _identify_from_evidence(
        or (", ".join(tools) if (tools := _ai_tools_in(region)) else None)
    )
    # Platform: a distinctive device/camera token in the manifest wins (it is the
-    # signer/producer), then an editing-app/AI-device signer (Samsung Galaxy,
-    # ASUS Gallery), with the issuer byte-scan only as fallback. The issuer scan
-    # alone mis-attributed real samples (Leica->Truepic timestamp authority,
-    # Nikon->Adobe namespace, Pixel->Google Gemini) -- the token scans fix that.
+    # signer/producer), then an exact product generator, then an editing-app or
+    # AI-device signer (Samsung Galaxy, ASUS Gallery), with the issuer byte-scan
+    # only as fallback. The issuer scan alone mis-attributed real samples
+    # (Leica->Truepic timestamp authority, Nikon->Adobe namespace, Pixel->Google
+    # Gemini) -- the token scans fix that.
    platform = (
        (
            camera_label
+            # Exact product generators are useful provenance even when the
+            # signed operation is a non-AI edit (for example, CapCut).
+            or _claim_generator_platform(generator)
            or signer_label
-            or (_claim_generator_platform(generator) if c2pa_is_ai else None)
            or (_claim_generator_platform(str(info.get("ai_tool"))) if c2pa_is_ai and info.get("ai_tool") else None)
            or _attribute_platform(issuers, is_ai=c2pa_is_ai)
        )
@@ -19,8 +19,8 @@ from typing import TYPE_CHECKING, Any
 from ._internal.watermark_profiles import (
    DEFAULT_PROFILE,
    REMOVAL_MODULES,
-    SDXL_ZIMAGE_PROFILE,
    resolve_adaptive_polish,
+    resolve_effective_profile,
    resolve_seed,
 )

@@ -191,9 +191,9 @@ class InvisibleEngine:
    def preload(self, *, global_only: bool = False) -> None:
        """Eagerly load the pipeline so download progress is visible.

-        For ``qwen-zimage``, ``global_only=True`` loads the mandatory Qwen stage
-        and leaves the optional Z-Image and SAM face stack lazy until a face is
-        detected. Other profiles have no optional stage and ignore the flag.
+        For every concrete profile, ``global_only=True`` loads its mandatory
+        global stage and leaves the shared Z-Image and SAM face stack lazy until
+        a face is detected.
        """
        self._remover.preload(global_only=global_only)

@@ -267,13 +267,19 @@ class InvisibleEngine:
        import tempfile

        seed = resolve_seed(seed)
-        adaptive_polish = resolve_adaptive_polish(adaptive_polish, self._remover.model_profile)
+        effective_profile = resolve_effective_profile(self._remover.configured_profile, vendor)
+        # Verified text must be the final pixel-writing stage. SDXL normally enables
+        # adaptive polish by profile default, so an unset value becomes False here;
+        # an explicit True remains an error below instead of being silently ignored.
+        adaptive_polish = (
+            False
+            if text_manifest is not None and adaptive_polish is None
+            else resolve_adaptive_polish(adaptive_polish, effective_profile)
+        )

        if fidelity_anchor and text_manifest is None:
            raise ValueError("fidelity_anchor requires a text manifest")
        if text_manifest is not None:
-            if self._remover.model_profile == SDXL_ZIMAGE_PROFILE:
-                raise ValueError("--text-manifest is not supported by the sdxl-zimage profile")
            if tile:
                raise ValueError("--text-manifest is not calibrated with --tile")
            if max_resolution != 0:
@@ -0,0 +1,266 @@
+"""Detect and disrupt the pixel payload written by local Microsoft Paint.
+
+This is a clean-room implementation of the observed ``Watermarker.dll`` wire
+format.  It carries no Microsoft binary, symbols, or writer-quality model.  The
+reader validates the repeated 144-bit message before the remover touches pixels;
+callers with pristine C2PA should also pass its signed watermark UUID.
+"""
+
+from __future__ import annotations
+
+import os
+import tempfile
+import uuid
+from dataclasses import dataclass
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from numpy.typing import NDArray
+
+_BLOCK_SIZE = 4
+_MAP_SIZE = 12
+_PAYLOAD_BITS = _MAP_SIZE * _MAP_SIZE
+_QUANTUM = 24.0
+_CENTERS = (6.0, 18.0)
+_MAX_CENTER_ERROR = 2.0
+_MIN_SUPPORT = 3
+
+
+@dataclass(frozen=True, slots=True)
+class LocalInvisMarkDetection:
+    """A validated local Paint payload recovered from repeated pixel carriers."""
+
+    watermark_id: uuid.UUID
+    payload: bytes
+    minimum_support: int
+
+
+@dataclass(frozen=True, slots=True)
+class LocalInvisMarkRemoval(LocalInvisMarkDetection):
+    """A disrupted local Paint payload and the number of rewritten carriers."""
+
+    changed_blocks: int
+
+
+@dataclass(frozen=True, slots=True)
+class _Analysis:
+    ll: NDArray[Any]
+    bits: NDArray[Any]
+    eligible: NDArray[Any]
+    bit_indices: NDArray[Any]
+
+
+def _dct_matrix() -> NDArray[Any]:
+    import numpy as np
+
+    from remove_ai_watermarks._internal.image_math import dct_matrix
+
+    return dct_matrix(np, _BLOCK_SIZE)
+
+
+def _luminance(rgb: NDArray[Any]) -> NDArray[Any]:
+    import numpy as np
+
+    return rgb.astype(np.float64) @ np.array((0.299, 0.587, 0.114), dtype=np.float64)
+
+
+def _ll_plane(luminance: NDArray[Any]) -> NDArray[Any]:
+    height = luminance.shape[0] - luminance.shape[0] % 2
+    width = luminance.shape[1] - luminance.shape[1] % 2
+    plane = luminance[:height, :width]
+    return (plane[0::2, 0::2] + plane[0::2, 1::2] + plane[1::2, 0::2] + plane[1::2, 1::2]) * 0.5
+
+
+def _analyze(rgb: NDArray[Any]) -> _Analysis:
+    import numpy as np
+
+    ll = _ll_plane(_luminance(rgb))
+    block_rows = ll.shape[0] // _BLOCK_SIZE
+    block_columns = ll.shape[1] // _BLOCK_SIZE
+    blocks = (
+        ll[: block_rows * _BLOCK_SIZE, : block_columns * _BLOCK_SIZE]
+        .reshape(block_rows, _BLOCK_SIZE, block_columns, _BLOCK_SIZE)
+        .transpose(0, 2, 1, 3)
+    )
+    transform = _dct_matrix()
+    coefficients = transform @ blocks @ transform.T
+    matrices = coefficients.reshape(block_rows, block_columns, -1)[..., 1:].reshape(block_rows, block_columns, 3, 5)
+    singular_values = np.linalg.svd(matrices, compute_uv=False)[..., 0]
+    residue = singular_values % _QUANTUM
+    center_errors = np.stack(
+        [np.minimum(abs(residue - center), _QUANTUM - abs(residue - center)) for center in _CENTERS],
+        axis=-1,
+    )
+    bits = np.argmin(center_errors, axis=-1).astype(np.uint8)
+    eligible = np.min(center_errors, axis=-1) <= _MAX_CENTER_ERROR
+    block_y, block_x = np.indices((block_rows, block_columns))
+    bit_indices = (block_y % _MAP_SIZE) * _MAP_SIZE + block_x % _MAP_SIZE
+    return _Analysis(ll=ll, bits=bits, eligible=eligible, bit_indices=bit_indices)
+
+
+def _validated_detection(rgb: NDArray[Any]) -> tuple[LocalInvisMarkDetection, _Analysis] | None:
+    import numpy as np
+
+    analysis = _analyze(rgb)
+    support = np.zeros((_PAYLOAD_BITS, 2), dtype=np.int32)
+    np.add.at(support, (analysis.bit_indices[analysis.eligible], analysis.bits[analysis.eligible]), 1)
+    winning_support = support.max(axis=1)
+    if np.any(winning_support < _MIN_SUPPORT) or np.any(support[:, 0] == support[:, 1]):
+        return None
+
+    bits = (support[:, 1] > support[:, 0]).astype(np.uint8)
+    payload = np.packbits(bits).tobytes()
+    if len(payload) != 18 or payload[0] != 0x4C or payload[-1] != sum(payload[1:17]) & 0xFF:
+        return None
+    watermark_id = uuid.UUID(bytes_le=payload[1:17])
+    if watermark_id.int == 0:
+        return None
+    detection = LocalInvisMarkDetection(
+        watermark_id=watermark_id,
+        payload=payload,
+        minimum_support=int(winning_support.min()),
+    )
+    return detection, analysis
+
+
+def _read_rgb(source: str | Path) -> tuple[NDArray[Any], NDArray[Any] | None]:
+    from remove_ai_watermarks import image_io
+
+    bgr, alpha = image_io.read_bgr_and_alpha(source)
+    if bgr is None:
+        raise ValueError(f"Could not read image: {source}")
+    return bgr[..., ::-1], alpha
+
+
+def detect_local_invismark(source: str | Path) -> LocalInvisMarkDetection | None:
+    """Return a validated local Paint payload, or ``None`` for an uncertain image."""
+    rgb, _alpha = _read_rgb(source)
+    recovered = _validated_detection(rgb)
+    return recovered[0] if recovered is not None else None
+
+
+def invismark_id_from_c2pa_info(info: dict[str, Any]) -> uuid.UUID | None:
+    """Return the one signed UUID for a Microsoft InvisMark soft binding."""
+    algorithms = {value.strip() for value in str(info.get("soft_binding_algorithm") or "").split(",") if value.strip()}
+    if "com.microsoft.invismark.1" not in algorithms:
+        return None
+    values = [value.strip() for value in str(info.get("soft_binding_value") or "").split(",") if value.strip()]
+    parsed: list[uuid.UUID] = []
+    for value in values:
+        try:
+            parsed.append(uuid.UUID(value))
+        except ValueError:
+            continue
+    unique = tuple(dict.fromkeys(parsed))
+    return unique[0] if len(unique) == 1 else None
+
+
+def invismark_id_from_c2pa(source: str | Path) -> uuid.UUID | None:
+    """Read the signed Microsoft InvisMark UUID from a pristine image."""
+    from remove_ai_watermarks._internal.c2pa import extract_c2pa_info
+
+    try:
+        return invismark_id_from_c2pa_info(extract_c2pa_info(Path(source)))
+    except Exception:
+        return None
+
+
+def disrupt_local_invismark(
+    source: str | Path,
+    output: str | Path,
+    *,
+    expected_watermark_id: str | uuid.UUID | None = None,
+) -> LocalInvisMarkRemoval | None:
+    """Invert confirmed local Paint carriers and atomically write the result.
+
+    No output is written unless the payload validates, an optional signed UUID
+    agrees, and the encoded output no longer contains a valid local payload.
+    This invalidates the observed message; it is not a claim about an unavailable
+    external Microsoft detector.
+    """
+    import numpy as np
+
+    from remove_ai_watermarks import image_io
+
+    source_path, output_path = Path(source), Path(output)
+    rgb, alpha = _read_rgb(source_path)
+    recovered = _validated_detection(rgb)
+    if recovered is None:
+        return None
+    detection, analysis = recovered
+    if expected_watermark_id is not None:
+        try:
+            expected = uuid.UUID(str(expected_watermark_id))
+        except ValueError:
+            return None
+        if detection.watermark_id != expected:
+            return None
+
+    luminance = _luminance(rgb)
+    ll = analysis.ll
+    original_ll = ll.copy()
+    transform = _dct_matrix()
+    payload_bits = np.unpackbits(np.frombuffer(detection.payload, dtype=np.uint8))
+    selected = analysis.eligible & (analysis.bits == payload_bits[analysis.bit_indices])
+    selected_y, selected_x = np.nonzero(selected)
+    changed_blocks = len(selected_y)
+    block_rows, block_columns = analysis.bits.shape
+    block_grid = (
+        ll[: block_rows * _BLOCK_SIZE, : block_columns * _BLOCK_SIZE]
+        .reshape(block_rows, _BLOCK_SIZE, block_columns, _BLOCK_SIZE)
+        .transpose(0, 2, 1, 3)
+        .copy()
+    )
+    for start in range(0, changed_blocks, 4096):
+        batch_y = selected_y[start : start + 4096]
+        batch_x = selected_x[start : start + 4096]
+        coefficients = transform @ block_grid[batch_y, batch_x] @ transform.T
+        matrices = coefficients.reshape(-1, 16)[:, 1:].reshape(-1, 3, 5)
+        target_bits = 1 - analysis.bits[batch_y, batch_x]
+        u, singular, vh = np.linalg.svd(matrices, full_matrices=False)
+        singular[:, 0] = (np.floor(singular[:, 0] / _QUANTUM) + 0.25 + 0.5 * target_bits) * _QUANTUM
+        rewritten = coefficients.reshape(-1, 16).copy()
+        rewritten[:, 1:] = ((u * singular[:, None, :]) @ vh).reshape(-1, 15)
+        block_grid[batch_y, batch_x] = transform.T @ rewritten.reshape(-1, 4, 4) @ transform
+    ll[: block_rows * _BLOCK_SIZE, : block_columns * _BLOCK_SIZE] = block_grid.transpose(0, 2, 1, 3).reshape(
+        block_rows * _BLOCK_SIZE, block_columns * _BLOCK_SIZE
+    )
+
+    height, width = original_ll.shape
+    delta = (ll - original_ll) * 0.5
+    pixel_delta = np.zeros(luminance.shape, dtype=np.float64)
+    pixel_delta[: height * 2 : 2, : width * 2 : 2] = delta
+    pixel_delta[: height * 2 : 2, 1 : width * 2 : 2] = delta
+    pixel_delta[1 : height * 2 : 2, : width * 2 : 2] = delta
+    pixel_delta[1 : height * 2 : 2, 1 : width * 2 : 2] = delta
+    cleaned_rgb = np.clip(np.rint(rgb.astype(np.float64) + pixel_delta[..., None]), 0, 255).astype(np.uint8)
+
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    file_descriptor, temporary_name = tempfile.mkstemp(
+        prefix=f".{output_path.stem}-", suffix=output_path.suffix, dir=output_path.parent
+    )
+    os.close(file_descriptor)
+    temporary_path = Path(temporary_name)
+    try:
+        if not image_io.write_bgr_with_alpha(
+            temporary_path,
+            cleaned_rgb[..., ::-1],
+            alpha,
+            display_tags_from=source_path,
+            orientation_applied=False,
+        ):
+            raise OSError(f"failed to write output (is the destination writable?): {output_path}")
+        if detect_local_invismark(temporary_path) is not None:
+            return None
+        os.replace(temporary_path, output_path)
+    finally:
+        if temporary_path.exists():
+            temporary_path.unlink()
+
+    return LocalInvisMarkRemoval(
+        watermark_id=detection.watermark_id,
+        payload=detection.payload,
+        minimum_support=detection.minimum_support,
+        changed_blocks=changed_blocks,
+    )
@@ -48,6 +48,7 @@ import time
 from dataclasses import dataclass, field
 from typing import TYPE_CHECKING, Any

+from remove_ai_watermarks._internal.image_math import dct_matrix as _dct_matrix
 from remove_ai_watermarks._internal.schema import require_schema_version

 if TYPE_CHECKING:
@@ -148,15 +149,6 @@ def _numpy() -> Any:
    return np


-def _dct_matrix(np: Any, n: int = 8) -> Any:
-    """Orthonormal n x n DCT-II basis: M[i, j] = cos(pi (2j + 1) i / 2n)."""
-    i = np.arange(n)[:, None]
-    j = np.arange(n)[None, :]
-    m = np.cos(np.pi * (2 * j + 1) * i / (2 * n))
-    m[0, :] *= 1 / np.sqrt(2)
-    return m * np.sqrt(2 / n)
-
-
 def read_gray(image_path: Path) -> tuple[Any, Any, dict[str, Any]]:
    """Decode to float32 grayscale (and RGB for color stats), downscaled.

@@ -0,0 +1,238 @@
+"""Abstaining source/export-pipeline classification from image pixels.
+
+This module does not detect or decode SynthID. It labels original-export-looking
+pixel pipelines as OpenAI, Google, or unknown after metadata may have been
+removed. Call it explicitly; ``identify`` never imports or runs it.
+"""
+
+from __future__ import annotations
+
+import math
+import os
+from dataclasses import dataclass
+from functools import lru_cache
+from hashlib import sha256
+from importlib import import_module
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, Literal
+
+if TYPE_CHECKING:
+    from numpy.typing import NDArray
+
+SOURCE_CLASSIFY_EXTRA = "'remove-ai-watermarks[source-classify]'"
+WEIGHTS_ENV = "RAIW_SOURCE_CLASSIFY_WEIGHTS"
+WEIGHTS_REPO = "wiltodelta/raiw-source-classify"
+WEIGHTS_REVISION = "4b8b408d03dac20231d46217977fd164414b6b9e"
+MODEL_FILE = "source-pipeline-mlp.npz"
+MODEL_SHA256 = "6430b0738869ff07b6845b7c6149c2a83bc3f3785878082886c241c6f7ca0c4a"
+CLASSES = ("openai", "google", "unknown")
+PROVIDERS = CLASSES[:2]
+FEATURE_WIDTH = 892
+SPECTRAL_GRID = 16
+CLAIM = "source/export pipeline classification; not SynthID detection"
+# NumPy and the training runtime can differ by a few float32 ULPs at a matrix
+# boundary. Treat values within this tolerance as the stored threshold so the
+# exported runtime preserves the training runtime's >= decision.
+THRESHOLD_TOLERANCE = 1e-5
+
+SourceLabel = Literal["openai", "google", "unknown"]
+SourceReason = Literal["classified", "abstained", "feature_unavailable"]
+
+
+@dataclass(frozen=True)
+class SourceClassification:
+    """One narrow source-pipeline result with an explicit abstention."""
+
+    label: SourceLabel
+    reason: SourceReason
+    scores: dict[str, float]
+    margins: dict[str, float]
+    thresholds: dict[str, float]
+    claim: str = CLAIM
+    model_revision: str = WEIGHTS_REVISION
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "label": self.label,
+            "reason": self.reason,
+            "scores": self.scores,
+            "margins": self.margins,
+            "thresholds": self.thresholds,
+            "claim": self.claim,
+            "model_revision": self.model_revision,
+        }
+
+
+@dataclass(frozen=True)
+class _Model:
+    mean: NDArray[Any]
+    scale: NDArray[Any]
+    margins: NDArray[Any]
+    layer0_weight: NDArray[Any]
+    layer0_bias: NDArray[Any]
+    layer1_weight: NDArray[Any]
+    layer1_bias: NDArray[Any]
+    output_weight: NDArray[Any]
+    output_bias: NDArray[Any]
+
+
+def is_available() -> bool:
+    """Return whether the lightweight source-classify runtime is installed."""
+    from remove_ai_watermarks.optional_deps import module_available
+
+    return module_available("numpy", "cv2", "huggingface_hub")
+
+
+def _model_path() -> Path:
+    override = os.environ.get(WEIGHTS_ENV, "").strip()
+    if override:
+        path = Path(override).expanduser()
+        candidate = path / MODEL_FILE if path.is_dir() else path
+        if not candidate.is_file():
+            raise RuntimeError(f"{WEIGHTS_ENV} does not contain {MODEL_FILE}: {path}")
+        return candidate
+    try:
+        hub: Any = import_module("huggingface_hub")
+    except ModuleNotFoundError as error:
+        raise RuntimeError(f"Install {SOURCE_CLASSIFY_EXTRA} to download the source classifier") from error
+    return Path(
+        hub.hf_hub_download(
+            repo_id=WEIGHTS_REPO,
+            filename=MODEL_FILE,
+            revision=WEIGHTS_REVISION,
+        )
+    )
+
+
+@lru_cache(maxsize=2)
+def _load_model(path: Path, expected_sha256: str | None = MODEL_SHA256) -> _Model:
+    import numpy as np
+
+    if expected_sha256 is not None:
+        digest = sha256(path.read_bytes()).hexdigest()
+        if digest != expected_sha256:
+            raise ValueError(f"model SHA-256 mismatch: expected {expected_sha256}, got {digest}")
+    required = {
+        "classes",
+        "kind",
+        "mean",
+        "scale",
+        "margins",
+        "layer0_weight",
+        "layer0_bias",
+        "layer1_weight",
+        "layer1_bias",
+        "output_weight",
+        "output_bias",
+        "activation",
+    }
+    with np.load(path, allow_pickle=False) as payload:
+        if set(payload.files) != required:
+            raise ValueError(f"model fields must be exactly {sorted(required)}")
+        if tuple(str(value) for value in payload["classes"]) != CLASSES:
+            raise ValueError(f"model classes must be {CLASSES}")
+        if str(payload["kind"]) != "fused":
+            raise ValueError("model kind must be fused")
+        if str(payload["activation"]) != "gelu_exact":
+            raise ValueError("model activation must be gelu_exact")
+        arrays = {
+            name: np.asarray(
+                payload[name],
+                dtype=np.float64 if name in {"mean", "scale", "margins"} else np.float32,
+            )
+            for name in required - {"classes", "kind", "activation"}
+        }
+    shapes = {
+        "mean": (FEATURE_WIDTH,),
+        "scale": (FEATURE_WIDTH,),
+        "margins": (len(PROVIDERS),),
+        "layer0_weight": (128, FEATURE_WIDTH),
+        "layer0_bias": (128,),
+        "layer1_weight": (64, 128),
+        "layer1_bias": (64,),
+        "output_weight": (len(CLASSES), 64),
+        "output_bias": (len(CLASSES),),
+    }
+    for name, expected in shapes.items():
+        if arrays[name].shape != expected:
+            raise ValueError(f"model field {name} must have shape {expected}")
+        if not np.all(np.isfinite(arrays[name])):
+            raise ValueError(f"model field {name} must be finite")
+    if np.any(arrays["scale"] <= 0) or np.any(arrays["margins"] < 0):
+        raise ValueError("model scales must be positive and margins non-negative")
+    return _Model(**arrays)
+
+
+def _gelu_exact(values: NDArray[Any]) -> NDArray[Any]:
+    import numpy as np
+
+    flat = values.ravel()
+    errors = np.fromiter((math.erf(float(value) / math.sqrt(2.0)) for value in flat), dtype=np.float32)
+    return (0.5 * flat * (1.0 + errors)).reshape(values.shape).astype(np.float32)
+
+
+def _scores(features: NDArray[Any], model: _Model) -> NDArray[Any]:
+    import numpy as np
+
+    values = np.asarray((features - model.mean) / model.scale, dtype=np.float32)
+    hidden0 = _gelu_exact(model.layer0_weight @ values + model.layer0_bias)
+    hidden1 = _gelu_exact(model.layer1_weight @ hidden0 + model.layer1_bias)
+    return np.asarray(model.output_weight @ hidden1 + model.output_bias, dtype=np.float64)
+
+
+def _result(raw_scores: NDArray[Any], thresholds: NDArray[Any]) -> SourceClassification:
+    import numpy as np
+
+    margins = np.empty(len(PROVIDERS), dtype=np.float64)
+    for index in range(len(PROVIDERS)):
+        competitors = np.delete(raw_scores, index)
+        margins[index] = raw_scores[index] - float(np.max(competitors))
+    excess = margins - thresholds
+    best = int(np.argmax(excess))
+    accepted = bool(excess[best] >= -THRESHOLD_TOLERANCE)
+    label: SourceLabel = PROVIDERS[best] if accepted else "unknown"
+    return SourceClassification(
+        label=label,
+        reason="classified" if accepted else "abstained",
+        scores={name: float(raw_scores[index]) for index, name in enumerate(CLASSES)},
+        margins={name: float(margins[index]) for index, name in enumerate(PROVIDERS)},
+        thresholds={name: float(thresholds[index]) for index, name in enumerate(PROVIDERS)},
+    )
+
+
+def _unavailable(model: _Model) -> SourceClassification:
+    return SourceClassification(
+        label="unknown",
+        reason="feature_unavailable",
+        scores={},
+        margins={},
+        thresholds={name: float(model.margins[index]) for index, name in enumerate(PROVIDERS)},
+    )
+
+
+def classify_source(source: str | os.PathLike[str]) -> SourceClassification:
+    """Classify an original-looking image export, abstaining as ``unknown``.
+
+    The result describes a complete source/export pipeline. It is not evidence
+    that SynthID is present or absent.
+    """
+    import numpy as np
+    from PIL import Image
+
+    from remove_ai_watermarks import image_io
+    from remove_ai_watermarks._internal.forensic_124d import image_features
+    from remove_ai_watermarks._internal.source_spectral import feature_from_image, make_geometry
+
+    model = _load_model(_model_path(), expected_sha256=MODEL_SHA256)
+    image_io._register_heif()  # pyright: ignore[reportPrivateUsage]
+    with Image.open(source) as opened:
+        image = opened.convert("RGB")
+    pixels = np.asarray(image, dtype=np.uint8)
+    forensic = image_features(pixels)
+    if forensic is None:
+        return _unavailable(model)
+    spectral = feature_from_image(image, make_geometry(grid=SPECTRAL_GRID))
+    features = np.concatenate((forensic, spectral))
+    if features.shape != (FEATURE_WIDTH,) or not np.all(np.isfinite(features)):
+        return _unavailable(model)
+    return _result(_scores(features, model), model.margins)
@@ -23,6 +23,10 @@ class TestTopLevelExports:
        assert raiw.remove_visible_detailed is api.remove_visible_detailed
        assert raiw.VisibleRemovalResult is api.VisibleRemovalResult
        assert raiw.visible_provenance is api.visible_provenance
+        from remove_ai_watermarks import source_classify
+
+        assert raiw.classify_source is source_classify.classify_source
+        assert raiw.SourceClassification is source_classify.SourceClassification

    def test_unknown_attribute_raises(self):
        with pytest.raises(AttributeError):
@@ -32,8 +32,8 @@ def _engine(run=None):
    engine = object.__new__(invisible_engine.InvisibleEngine)
    engine._progress_callback = None
    remover = object.__new__(WatermarkRemover)
+    remover.configured_profile = "qwen-zimage"
    remover.model_profile = "qwen-zimage"
-    remover._auto = False
    remover._qwen_zimage_pipeline = SimpleNamespace(run=run or (lambda source, **kwargs: source))
    engine._remover = remover
    return engine
@@ -30,14 +30,15 @@ OFFLOAD_SUPPORT_CASES = (
    (QWEN_ZIMAGE_PROFILE, True),
    (CHROMA_ZIMAGE_PROFILE, False),
    (SDXL_ZIMAGE_PROFILE, False),
-    # Unresolved auto appears only in preload, which builds qwen-zimage.
+    # Auto preloads its concrete Qwen fallback before an image vendor is known.
    (AUTO_PROFILE, True),
 )


 def _remover(profile: str, cpu_offload: bool) -> WatermarkRemover:
    remover = WatermarkRemover.__new__(WatermarkRemover)
-    remover.model_profile = profile
+    remover.configured_profile = profile
+    remover.model_profile = QWEN_ZIMAGE_PROFILE if profile == AUTO_PROFILE else profile
    remover.cpu_offload = cpu_offload
    remover.device = "cuda"
    remover.torch_dtype = None
@@ -46,6 +46,12 @@ _SPARKLE_TARGET = "remove_ai_watermarks.gemini_engine.detect_sparkle_confidence"
 SAMPLES_DIR = Path(__file__).resolve().parent.parent / "data" / "fixtures" / "provenance"


+def _write_c2pa_jpeg(tmp_path: Path, name: str, blob: bytes) -> Path:
+    path = tmp_path / name
+    path.write_bytes(b"\xff\xd8\xff\xe1jumbc2pa" + blob + b"\xff\xd9")
+    return path
+
+
 class TestProvenanceEvidence:
    def test_exact_ai_claim_generator_can_assert_ai_without_source_type(self, tmp_path: Path):
        path = tmp_path / "firefly.png"
@@ -646,13 +652,8 @@ class TestIdentifySamsungGalaxy:
    so the signer attribution must NOT trip the camera-vs-AI integrity clash.
    """

-    def _jpeg(self, tmp_path: Path, name: str, blob: bytes) -> Path:
-        path = tmp_path / name
-        path.write_bytes(b"\xff\xd8\xff\xe1jumbc2pa" + blob + b"\xff\xd9")
-        return path
-
    def test_galaxy_trained_source_is_unverified_ai(self, tmp_path: Path):
-        path = self._jpeg(tmp_path, "s25.jpg", b"Samsung Galaxy Galaxy S25 c2pa-rs trainedAlgorithmicMedia")
+        path = _write_c2pa_jpeg(tmp_path, "s25.jpg", b"Samsung Galaxy Galaxy S25 c2pa-rs trainedAlgorithmicMedia")
        r = identify(path, check_visible=False, check_invisible=False)
        assert r.is_ai_generated is True
        assert r.confidence == "medium"
@@ -664,7 +665,7 @@ class TestIdentifySamsungGalaxy:
    def test_galaxy_genai_only_is_medium_ai(self, tmp_path: Path):
        # The Galaxy S24 case: no trainedAlgorithmicMedia, genAIType is the only
        # AI marker -- previously missed, now a medium-confidence verdict.
-        path = self._jpeg(
+        path = _write_c2pa_jpeg(
            tmp_path, "s24.jpg", b'Samsung Galaxy Galaxy S24 c2pa-rs PhotoEditor_Re_Edit_Data{"genAIType":1}'
        )
        r = identify(path, check_visible=False, check_invisible=False)
@@ -677,7 +678,7 @@ class TestIdentifySamsungGalaxy:
    def test_asus_gallery_signer_not_ai(self, tmp_path: Path):
        # ASUS Gallery signs edited photos; no AI source-type or genAIType, so the
        # platform is attributed but the verdict stays unknown.
-        path = self._jpeg(tmp_path, "asus.jpg", b"/com.asus.gallery/3.8.0.98 c2pa-rs no ai marker")
+        path = _write_c2pa_jpeg(tmp_path, "asus.jpg", b"/com.asus.gallery/3.8.0.98 c2pa-rs no ai marker")
        r = identify(path, check_visible=False, check_invisible=False)
        assert r.is_ai_generated is None
        assert r.platform == "ASUS Gallery (C2PA signer)"
@@ -687,13 +688,88 @@ class TestIdentifySamsungGalaxy:
        # A genuine Galaxy phone capture carries Samsung Galaxy C2PA provenance but
        # NO AI source-type / genAIType. It must stay is_ai=None -- the device cert
        # is authenticity provenance of a real photo, not an AI-generation signal.
-        path = self._jpeg(tmp_path, "s25_capture.jpg", b"Samsung Galaxy Galaxy S25 c2pa-rs no ai marker")
+        path = _write_c2pa_jpeg(tmp_path, "s25_capture.jpg", b"Samsung Galaxy Galaxy S25 c2pa-rs no ai marker")
        r = identify(path, check_visible=False, check_invisible=False)
        assert r.is_ai_generated is None
        assert r.platform == "Samsung Galaxy (C2PA)"
        assert any("C2PA" in w for w in r.watermarks)


+class TestIdentifyDiscoveredC2paSigners:
+    @pytest.mark.parametrize(
+        ("blob", "platform"),
+        [
+            (b"Anthropic Claude Content Signing", "Anthropic Claude (C2PA signer)"),
+            (b"CapCut/ c2pa-rs", "CapCut (C2PA signer)"),
+            (b"TikTok Inc.", "TikTok (C2PA signer)"),
+            (b"vivo Camera vivo X300 Pro", "vivo (camera, C2PA capture)"),
+        ],
+    )
+    def test_non_ai_signer_is_attributed_without_asserting_ai(self, tmp_path: Path, blob: bytes, platform: str):
+        path = _write_c2pa_jpeg(tmp_path, "signed.jpg", blob)
+
+        report = identify(path, check_visible=False, check_invisible=False)
+
+        assert report.is_ai_generated is None
+        assert report.platform == platform
+
+    def test_structured_non_ai_capcut_generator_is_attributed(self, tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+        path = tmp_path / "edited.png"
+        from PIL import Image
+
+        Image.new("RGB", (32, 32)).save(path)
+        monkeypatch.setattr(
+            "remove_ai_watermarks.identify.extract_c2pa_info",
+            lambda _path: {
+                "has_c2pa": True,
+                "issuer": "ByteDance",
+                "claim_generator": "CapCut/ c2pa-tool/0.1.0 c2pa-rs/0.31.3",
+            },
+        )
+
+        report = identify(path, check_visible=False, check_invisible=False)
+
+        assert report.is_ai_generated is None
+        assert report.platform == "CapCut (C2PA signer)"
+
+    @pytest.mark.parametrize(
+        ("issuer", "claim_generator", "platform"),
+        [
+            ("xAI Grok Imagine", "c2pa-rs", "xAI Grok Imagine"),
+            ("Producer.ai", "producer_ai", "Producer.ai"),
+            ("", "FastVid", "FastVid"),
+            ("SPRING (SG) PTE. LTD.", "c2pa-rs", "SPRING (SG) PTE. LTD."),
+        ],
+    )
+    def test_ai_signer_or_generator_is_attributed(
+        self,
+        tmp_path: Path,
+        monkeypatch: pytest.MonkeyPatch,
+        issuer: str,
+        claim_generator: str,
+        platform: str,
+    ):
+        path = tmp_path / "generated.png"
+        from PIL import Image
+
+        Image.new("RGB", (32, 32)).save(path)
+        monkeypatch.setattr(
+            "remove_ai_watermarks.identify.extract_c2pa_info",
+            lambda _path: {
+                "has_c2pa": True,
+                "issuer": issuer,
+                "claim_generator": claim_generator,
+                "source_type": "trainedAlgorithmicMedia (AI-generated)",
+                "ai_source_kind": "generated",
+            },
+        )
+
+        report = identify(path, check_visible=False, check_invisible=False)
+
+        assert report.is_ai_generated is True
+        assert report.platform == platform
+
+
 # ── End-to-end verdicts on real fixtures ────────────────────────────


@@ -55,7 +55,7 @@ class TestVerifiedTextMode:
    def _engine(profile: str = "qwen-zimage") -> InvisibleEngine:
        engine = object.__new__(InvisibleEngine)
        engine._progress_callback = None
-        engine._remover = SimpleNamespace(model_profile=profile)
+        engine._remover = SimpleNamespace(configured_profile=profile, model_profile=profile)
        return engine

    def test_rejects_incompatible_pipeline_options(self, tmp_path):
@@ -64,7 +64,6 @@ class TestVerifiedTextMode:
        manifest = tmp_path / "manifest.json"
        manifest.write_text("{}", encoding="utf-8")
        cases = (
-            ("sdxl-zimage", {}, "not supported by the sdxl-zimage profile"),
            ("qwen-zimage", {"tile": True}, "not calibrated with --tile"),
            ("qwen-zimage", {"max_resolution": 1024}, "max-resolution 0"),
            ("qwen-zimage", {"humanize": 1.0}, "humanize=0"),
@@ -124,9 +123,10 @@ class TestVerifiedTextMode:
        engine._remover.remove_watermark = fake_remove
        monkeypatch.setattr(region_eraser, "lama_available", lambda: True)

-        engine.remove_watermark(source, output, text_manifest=manifest)
+        engine.remove_watermark(source, output, text_manifest=manifest, vendor="google")

        assert seen["text_manifest"].lines[0].text == "Exact"
+        assert seen["vendor"] == "google"
        # Leak-safe default since 0.27.1: the global 15% donor blend is OFF unless
        # explicitly requested (measured to return detector-visible OpenAI SynthID
        # on poster-scale manifests; see docs/text-protection-research.md).
@@ -149,7 +149,11 @@ class TestNativeOutputSize:
            Image.open(image_path).crop((0, 0, 24, 16)).save(out)
            return out

-        engine._remover = SimpleNamespace(remove_watermark=_remove_watermark, model_profile="qwen-zimage")
+        engine._remover = SimpleNamespace(
+            remove_watermark=_remove_watermark,
+            configured_profile="qwen-zimage",
+            model_profile="qwen-zimage",
+        )
        engine._progress_callback = None
        src = tmp_path / "src.png"
        out = tmp_path / "out.png"
@@ -262,10 +266,10 @@ class TestEngineResolvesThePolishPerProfile:

        engine = object.__new__(InvisibleEngine)
        engine._progress_callback = None
-        engine._remover = MagicMock(model_profile=profile)
+        engine._remover = MagicMock(configured_profile=profile, model_profile=profile)
        return engine

-    def _polish_used(self, profile: str, requested, tmp_path, monkeypatch) -> bool:
+    def _polish_used(self, profile: str, requested, tmp_path, monkeypatch, *, vendor: str | None = None) -> bool:
        seen: list[bool] = []
        monkeypatch.setattr(
            "remove_ai_watermarks.humanizer.adaptive_polish",
@@ -278,13 +282,21 @@ class TestEngineResolvesThePolishPerProfile:
            Image.open(kw["image_path"]).save(kw["output_path"]),
            kw["output_path"],
        )[1]
-        engine.remove_watermark(src, tmp_path / f"out_{profile}_{requested}.png", adaptive_polish=requested)
+        engine.remove_watermark(
+            src,
+            tmp_path / f"out_{profile}_{requested}.png",
+            adaptive_polish=requested,
+            vendor=vendor,
+        )
        return bool(seen)

    def test_unset_follows_the_profile_not_the_signature_default(self, tmp_path, monkeypatch):
        assert self._polish_used("qwen-zimage", None, tmp_path, monkeypatch) is False
        assert self._polish_used("sdxl-zimage", None, tmp_path, monkeypatch) is True

+    def test_auto_google_uses_the_resolved_sdxl_default(self, tmp_path, monkeypatch):
+        assert self._polish_used("auto", None, tmp_path, monkeypatch, vendor="google") is True
+
    def test_an_explicit_value_still_wins_on_both_profiles(self, tmp_path, monkeypatch):
        assert self._polish_used("qwen-zimage", True, tmp_path, monkeypatch) is True
        assert self._polish_used("sdxl-zimage", False, tmp_path, monkeypatch) is False
@@ -211,6 +211,43 @@ class TestC2PA:
        assert info["ai_tool"] == "Dreamina"
        assert info["c2pa_identity_ai"] is True

+    def test_structured_xai_signer_identity_is_normalized(self):
+        store = {
+            "active_manifest": "active",
+            "manifests": {
+                "active": {
+                    "signature_info": {
+                        "issuer": "Self-signed ephemeral certificate (Content Authenticity SDK) -- LOCAL USE ONLY",
+                        "common_name": "xAI Grok Imagine",
+                    },
+                    "assertions": [],
+                }
+            },
+        }
+
+        info = c2pa_info_from_manifest_store(store)
+
+        assert info["issuer"] == "xAI Grok Imagine"
+        assert info["c2pa_identity_ai"] is True
+
+    @pytest.mark.parametrize(
+        ("signature", "issuer"),
+        [
+            ({"issuer": "Bytedance Pte. Ltd."}, "ByteDance"),
+            ({"issuer": "TikTok Inc."}, "TikTok"),
+        ],
+    )
+    def test_structured_publisher_identity_does_not_assert_ai(self, signature, issuer):
+        store = {
+            "active_manifest": "active",
+            "manifests": {"active": {"signature_info": signature, "assertions": []}},
+        }
+
+        info = c2pa_info_from_manifest_store(store)
+
+        assert info["issuer"] == issuer
+        assert "c2pa_identity_ai" not in info
+
    def test_structured_invismark_exposes_algorithm_and_watermark_id(self):
        watermark_id = "83424621-03cb-40e3-9808-a9fae837156d"
        store = {
@@ -0,0 +1,128 @@
+"""Clean-room tests for Paint's local ``Watermarker.dll`` pixel format."""
+
+from __future__ import annotations
+
+import uuid
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+
+from remove_ai_watermarks.microsoft_invismark import (
+    detect_local_invismark,
+    disrupt_local_invismark,
+    invismark_id_from_c2pa_info,
+)
+
+_WATERMARK_ID = uuid.UUID("83424621-03cb-40e3-9808-a9fae837156d")
+
+
+def _dct_matrix() -> np.ndarray:
+    matrix = np.empty((4, 4), dtype=np.float64)
+    for frequency in range(4):
+        scale = np.sqrt(1 / 4) if frequency == 0 else np.sqrt(2 / 4)
+        for sample in range(4):
+            matrix[frequency, sample] = scale * np.cos(np.pi * (2 * sample + 1) * frequency / 8)
+    return matrix
+
+
+def _synthetic_writer(image: np.ndarray, watermark_id: uuid.UUID, *, valid_checksum: bool = True) -> np.ndarray:
+    """Independent format producer; it deliberately does not call production helpers."""
+    guid = watermark_id.bytes_le
+    checksum = sum(guid) & 0xFF
+    if not valid_checksum:
+        checksum ^= 1
+    payload = bytes((0x4C,)) + guid + bytes((checksum,))
+    bits = np.unpackbits(np.frombuffer(payload, dtype=np.uint8))
+
+    rgb = image.astype(np.float64)
+    luminance = rgb @ np.array((0.299, 0.587, 0.114), dtype=np.float64)
+    ll = (luminance[0::2, 0::2] + luminance[0::2, 1::2] + luminance[1::2, 0::2] + luminance[1::2, 1::2]) * 0.5
+    original_ll = ll.copy()
+    dct_matrix = _dct_matrix()
+
+    for block_y in range(ll.shape[0] // 4):
+        for block_x in range(ll.shape[1] // 4):
+            top, left = block_y * 4, block_x * 4
+            block = ll[top : top + 4, left : left + 4]
+            coefficients = dct_matrix @ block @ dct_matrix.T
+            ac = coefficients.reshape(-1)[1:].reshape(3, 5)
+            u, singular, vh = np.linalg.svd(ac, full_matrices=False)
+            bit = int(bits[(block_y % 12) * 12 + (block_x % 12)])
+            singular[0] = (np.floor(singular[0] / 24) + 0.25 + 0.5 * bit) * 24
+            rewritten = coefficients.reshape(-1).copy()
+            rewritten[1:] = (u @ np.diag(singular) @ vh).reshape(-1)
+            ll[top : top + 4, left : left + 4] = dct_matrix.T @ rewritten.reshape(4, 4) @ dct_matrix
+
+    delta = (ll - original_ll) * 0.5
+    pixel_delta = np.repeat(np.repeat(delta, 2, axis=0), 2, axis=1)
+    return np.clip(np.rint(rgb + pixel_delta[..., None]), 0, 255).astype(np.uint8)
+
+
+def _write_carrier(path: Path, *, valid_checksum: bool = True) -> np.ndarray:
+    original = np.random.default_rng(20260913).integers(24, 232, size=(192, 192, 3), dtype=np.uint8)
+    encoded = _synthetic_writer(original, _WATERMARK_ID, valid_checksum=valid_checksum)
+    Image.fromarray(encoded, "RGB").save(path)
+    return original
+
+
+def test_detects_repeated_payload_and_windows_guid_order(tmp_path: Path) -> None:
+    source = tmp_path / "paint.png"
+    _write_carrier(source)
+
+    detection = detect_local_invismark(source)
+
+    assert detection is not None
+    assert detection.watermark_id == _WATERMARK_ID
+    assert detection.payload == b"\x4c" + _WATERMARK_ID.bytes_le + bytes((sum(_WATERMARK_ID.bytes_le) & 0xFF,))
+    assert detection.minimum_support >= 3
+
+
+def test_rejects_clean_and_malformed_payloads(tmp_path: Path) -> None:
+    clean = tmp_path / "clean.png"
+    malformed = tmp_path / "malformed.png"
+    Image.fromarray(
+        np.random.default_rng(20260913).integers(24, 232, size=(192, 192, 3), dtype=np.uint8),
+        "RGB",
+    ).save(clean)
+    _write_carrier(malformed, valid_checksum=False)
+
+    assert detect_local_invismark(clean) is None
+    assert detect_local_invismark(malformed) is None
+
+
+def test_disruption_requires_expected_guid_and_invalidates_payload(tmp_path: Path) -> None:
+    source = tmp_path / "paint.png"
+    wrong_output = tmp_path / "wrong.png"
+    output = tmp_path / "clean.png"
+    original = _write_carrier(source)
+
+    assert disrupt_local_invismark(source, wrong_output, expected_watermark_id=uuid.uuid4()) is None
+    assert not wrong_output.exists()
+
+    result = disrupt_local_invismark(source, output, expected_watermark_id=_WATERMARK_ID)
+
+    assert result is not None
+    assert result.watermark_id == _WATERMARK_ID
+    assert result.changed_blocks >= 432
+    assert output.exists()
+    assert detect_local_invismark(output) is None
+    cleaned = np.asarray(Image.open(output).convert("RGB"))
+    assert np.mean(np.abs(cleaned.astype(np.int16) - original.astype(np.int16))) < 2.0
+
+
+def test_cloud_image_creator_fixture_is_not_the_local_dll_format() -> None:
+    fixture = Path(__file__).parents[1] / "data/fixtures/provenance/microsoft-paint-invismark.png"
+
+    assert detect_local_invismark(fixture) is None
+
+
+def test_c2pa_uuid_requires_exact_algorithm_and_one_valid_value() -> None:
+    info = {
+        "soft_binding_algorithm": "com.microsoft.invismark.1",
+        "soft_binding_value": str(_WATERMARK_ID),
+    }
+
+    assert invismark_id_from_c2pa_info(info) == _WATERMARK_ID
+    assert invismark_id_from_c2pa_info({**info, "soft_binding_algorithm": "io.iscc.v0"}) is None
+    assert invismark_id_from_c2pa_info({**info, "soft_binding_value": f"{_WATERMARK_ID}, {uuid.uuid4()}"}) is None
@@ -240,6 +240,7 @@ class TestResolveStrength:
        )
        assert resolve_strength(None, None, "qwen-zimage", size=(600, 500)) == pytest.approx(0.084)
        assert resolve_strength(None, "google", "qwen-zimage", size=(600, 500)) == QWEN_ZIMAGE_GOOGLE_STRENGTH
+        assert QWEN_ZIMAGE_GOOGLE_STRENGTH == 0.35
        # The floor holds at every size, not only above the curve's top rung.
        assert resolve_strength(None, "google", "qwen-zimage", size=(2000, 1850)) == QWEN_ZIMAGE_GOOGLE_STRENGTH

@@ -278,7 +279,7 @@ class TestResolveStrength:
    def test_sdxl_zimage_uses_its_flat_vendor_ladder(self):

        assert SDXL_ZIMAGE_OPENAI_STRENGTH == 0.15
-        assert SDXL_ZIMAGE_GEMINI_STRENGTH == 0.25
+        assert SDXL_ZIMAGE_GEMINI_STRENGTH == 0.50
        assert SDXL_ZIMAGE_UNKNOWN_STRENGTH == SDXL_ZIMAGE_GEMINI_STRENGTH
        assert resolve_strength(None, "openai", "sdxl-zimage") == SDXL_ZIMAGE_OPENAI_STRENGTH
        assert resolve_strength(None, "google", "sdxl-zimage") == SDXL_ZIMAGE_GEMINI_STRENGTH
@@ -2,29 +2,21 @@

 from __future__ import annotations

-import importlib.util
+import sys
 from pathlib import Path
-from typing import Any

 import pytest

-
-def _publish() -> Any:
-    path = Path(__file__).resolve().parents[1] / "scripts" / "publish_photo_classify_hf.py"
-    spec = importlib.util.spec_from_file_location("publish_photo_classify_hf", path)
-    assert spec is not None
-    assert spec.loader is not None
-    module = importlib.util.module_from_spec(spec)
-    spec.loader.exec_module(module)
-    return module
+SCRIPTS = Path(__file__).resolve().parents[1] / "scripts"
+sys.path.insert(0, str(SCRIPTS))
+import publish_photo_classify_hf as publish  # noqa: E402


 def test_hub_id_is_the_photo_classify_repo() -> None:
-    assert _publish().HUB_REPO == "wiltodelta/raiw-photo-classify"
+    assert publish.HUB_REPO == "wiltodelta/raiw-photo-classify"


 def test_publish_freeze_includes_both_clip_runtimes() -> None:
-    publish = _publish()
    from remove_ai_watermarks.classify import _WEIGHT_FILES, CLIP_ONNX_FILE, WEIGHTS_ALLOW_PATTERNS

    assert publish.CLIP_FILE in publish.WEIGHT_FILES
@@ -41,7 +33,7 @@ def test_weight_paths_accepts_run1_layout(tmp_path: Path) -> None:
    run1.mkdir()
    (run1 / "detector.pt").write_bytes(b"det")
    (run1 / "provider.pt").write_bytes(b"prov")
-    found = _publish().weight_paths(tmp_path)
+    found = publish.weight_paths(tmp_path)
    assert found["detector.pt"] == run1 / "detector.pt"
    assert found["clip-l-ft.pt"] == tmp_path / "clip-l-ft.pt"

@@ -50,12 +42,12 @@ def test_weight_paths_requires_every_file(tmp_path: Path) -> None:
    (tmp_path / "clip-l-ft.pt").write_bytes(b"clip")
    (tmp_path / "clip-l-ft-vision-fp32.onnx").write_bytes(b"onnx")
    with pytest.raises(SystemExit, match=r"missing probe-weights-clip-l-ft\.npz"):
-        _publish().weight_paths(tmp_path)
+        publish.weight_paths(tmp_path)


 def test_stage_card_copies_tracked_files(tmp_path: Path) -> None:
    dest = tmp_path / "stage"
-    _publish().stage_card(dest)
+    publish.stage_card(dest)
    assert (dest / "README.md").is_file()
    assert (dest / "operating-point.json").is_file()
    text = (dest / "operating-point.json").read_text()
@@ -0,0 +1,37 @@
+"""Local staging checks for the source-classifier Hub publisher."""
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+import pytest
+
+SCRIPTS = Path(__file__).resolve().parents[1] / "scripts"
+sys.path.insert(0, str(SCRIPTS))
+import publish_source_classify_hf as publish  # noqa: E402
+
+
+def test_hub_id_is_the_source_classify_repo() -> None:
+    assert publish.HUB_REPO == "wiltodelta/raiw-source-classify"
+
+
+def test_stage_release_copies_only_public_artifacts(tmp_path: Path) -> None:
+    source = tmp_path / "source"
+    source.mkdir()
+    (source / "source-pipeline-mlp.npz").write_bytes(b"weights")
+    destination = tmp_path / "stage"
+    publish.stage_release(destination, source)
+    assert sorted(path.name for path in destination.iterdir()) == sorted(publish.PUBLIC_FILES)
+
+
+def test_stage_release_requires_the_model(tmp_path: Path) -> None:
+    with pytest.raises(SystemExit, match=r"source-pipeline-mlp\.npz"):
+        publish.stage_release(tmp_path / "stage", tmp_path)
+
+
+def test_card_sha_matches_the_metrics_manifest() -> None:
+    metrics = json.loads((publish.CARD_DIR / "metrics.json").read_text())
+    card = (publish.CARD_DIR / "README.md").read_text()
+    assert metrics["artifact_sha256"] in card
@@ -996,6 +996,70 @@ def test_sdxl_zimage_inherits_the_face_stage_rather_than_copying_it():
    assert SdxlZImagePipeline._load_global is not QwenZImagePipeline._load_global


+@pytest.mark.parametrize(
+    ("pipeline_module", "pipeline_class", "grid"),
+    [
+        ("sdxl_zimage_pipeline", "SdxlZImagePipeline", 8),
+        ("chroma_zimage_pipeline", "ChromaZImagePipeline", 16),
+    ],
+)
+def test_diffusers_vae_roundtrip_uses_deterministic_mode_and_restores_native_size(
+    monkeypatch, pipeline_module, pipeline_class, grid
+):
+    """Diffusers donors pad without distorting geometry and never sample."""
+    import importlib
+
+    import torch
+
+    seen: dict[str, object] = {}
+    padded_width = 40 if grid == 8 else 48
+    padded_height = 24 if grid == 8 else 32
+    latent = torch.ones(1, 4, padded_height // 8, padded_width // 8)
+
+    class LatentDistribution:
+        def mode(self):
+            seen["mode_calls"] = int(seen.get("mode_calls", 0)) + 1
+            return latent
+
+        def sample(self, *_args, **_kwargs):
+            raise AssertionError("A verified-text donor must not sample VAE latents")
+
+    class FakeVae:
+        @staticmethod
+        def encode(tensor):
+            seen["encoded_shape"] = tuple(tensor.shape)
+            return type("Encoded", (), {"latent_dist": LatentDistribution()})()
+
+        @staticmethod
+        def decode(received, *, return_dict):
+            assert received is latent
+            assert return_dict is False
+            return (torch.zeros(1, 3, padded_height, padded_width),)
+
+    class FakeProcessor:
+        @staticmethod
+        def preprocess(image, *, height, width):
+            seen["preprocess"] = (image.size, height, width)
+            return torch.zeros(1, 3, height, width)
+
+        @staticmethod
+        def postprocess(decoded, *, output_type):
+            assert output_type == "pil"
+            return [Image.new("RGB", (decoded.shape[3], decoded.shape[2]))]
+
+    pipe = type("FakePipe", (), {"vae": FakeVae(), "image_processor": FakeProcessor()})()
+    module = importlib.import_module(f"remove_ai_watermarks._internal.{pipeline_module}")
+    pipeline = getattr(module, pipeline_class)(device="cpu", torch_dtype=torch.float32)
+    monkeypatch.setattr(pipeline, "_load_global", lambda: pipe)
+
+    result = pipeline._vae_roundtrip(Image.new("RGB", (35, 19)))
+
+    assert seen["preprocess"] == ((padded_width, padded_height), padded_height, padded_width)
+    assert seen["encoded_shape"] == (1, 3, padded_height, padded_width)
+    assert seen["mode_calls"] == 1
+    assert result.size == (35, 19)
+
+
 def test_chroma_zimage_strength_uses_measured_flat_floors():
    """The chroma floors come from the four-cohort oracle calibration; they must not
    drift from their derivation in watermark_profiles."""
@@ -1015,7 +1079,7 @@ def test_chroma_zimage_strength_uses_measured_flat_floors():
    assert resolve_strength(None, "microsoft", "chroma-zimage") == pytest.approx(CHROMA_ZIMAGE_MICROSOFT_STRENGTH)
    assert pytest.approx(0.125) == CHROMA_ZIMAGE_MICROSOFT_STRENGTH
    assert resolve_strength(None, "google", "chroma-zimage") == pytest.approx(CHROMA_ZIMAGE_GOOGLE_STRENGTH)
-    assert pytest.approx(0.40) == CHROMA_ZIMAGE_GOOGLE_STRENGTH
+    assert pytest.approx(0.50) == CHROMA_ZIMAGE_GOOGLE_STRENGTH
    assert resolve_strength(None, "meta", "chroma-zimage") == pytest.approx(CHROMA_ZIMAGE_META_STRENGTH)
    assert pytest.approx(0.17) == CHROMA_ZIMAGE_META_STRENGTH
    # Unknown follows the strictest measured cohort, as with the other profiles.
@@ -1163,8 +1227,9 @@ def test_chroma_zimage_floors_to_its_own_latent_grid():
    assert chroma_target_size(3, 3) == (16, 16)


-def test_watermark_remover_accepts_text_manifest_for_chroma_but_not_sdxl(tmp_path, monkeypatch):
-    """Chroma owns a VAE donor; SDXL still fails at the public boundary."""
+@pytest.mark.parametrize("profile", ["chroma-zimage", "sdxl-zimage"])
+def test_watermark_remover_accepts_text_manifest_for_every_diffusers_profile(tmp_path, monkeypatch, profile):
+    """Both Diffusers profiles forward reviewed text to their VAE donor."""
    from remove_ai_watermarks._internal.text_restoration import VerifiedTextLine, VerifiedTextManifest
    from remove_ai_watermarks._internal.watermark_remover import WatermarkRemover

@@ -1173,22 +1238,18 @@ def test_watermark_remover_accepts_text_manifest_for_chroma_but_not_sdxl(tmp_pat
    Image.new("RGB", (96, 80)).save(source)
    manifest = VerifiedTextManifest("0" * 64, 96, 80, (VerifiedTextLine((4, 4, 20, 16), "x", "alphabetic"),))

-    chroma = WatermarkRemover(device="cuda", pipeline="chroma-zimage")
-    chroma_runtime = MagicMock()
-    chroma_runtime.run.return_value = Image.new("RGB", (96, 80))
-    monkeypatch.setattr(chroma, "_load_qwen_zimage_pipeline", lambda: chroma_runtime)
+    remover = WatermarkRemover(device="cuda", pipeline=profile)
+    runtime = MagicMock()
+    runtime.run.return_value = Image.new("RGB", (96, 80))
+    monkeypatch.setattr(remover, "_load_qwen_zimage_pipeline", lambda: runtime)

-    chroma.remove_watermark(source, text_manifest=manifest)
+    remover.remove_watermark(source, text_manifest=manifest)

-    assert chroma_runtime.run.call_count == 1
-    assert chroma_runtime.run.call_args.kwargs["text_manifest"] is manifest
+    assert runtime.run.call_count == 1
+    assert runtime.run.call_args.kwargs["text_manifest"] is manifest
    with pytest.raises(ValueError, match=r"fidelity anchor.*qwen-zimage"):
-        chroma.remove_watermark(source, text_manifest=manifest, fidelity_anchor=True)
-    assert chroma_runtime.run.call_count == 1
-
-    sdxl = WatermarkRemover(device="cuda", pipeline="sdxl-zimage")
-    with pytest.raises(ValueError, match="not supported by the sdxl-zimage profile"):
-        sdxl.remove_watermark(source, text_manifest=manifest)
+        remover.remove_watermark(source, text_manifest=manifest, fidelity_anchor=True)
+    assert runtime.run.call_count == 1


 def test_watermark_remover_dispatches_to_chroma_pipeline(monkeypatch):
@@ -1229,95 +1290,32 @@ def test_chroma_adaptive_polish_defaults_off():
    assert resolve_adaptive_polish(True, "chroma-zimage") is True


-def test_chroma_vae_roundtrip_uses_deterministic_mode_and_restores_native_size(monkeypatch):
-    """The Chroma donor pads without distorting geometry and never samples."""
-    import torch
+def test_chroma_google_uses_one_content_agnostic_floor():
+    """Face detection alone cannot safely select a lower Google floor.

-    from remove_ai_watermarks._internal.chroma_zimage_pipeline import ChromaZImagePipeline
-
-    seen: dict[str, object] = {}
-    latent = torch.ones(1, 4, 4, 6)
-
-    class LatentDistribution:
-        def mode(self):
-            seen["mode_calls"] = int(seen.get("mode_calls", 0)) + 1
-            return latent
-
-        def sample(self, *_args, **_kwargs):
-            raise AssertionError("A verified-text donor must not sample VAE latents")
-
-    class FakeVae:
-        @staticmethod
-        def encode(tensor):
-            seen["encoded_shape"] = tuple(tensor.shape)
-            return type("Encoded", (), {"latent_dist": LatentDistribution()})()
-
-        @staticmethod
-        def decode(received, *, return_dict):
-            assert received is latent
-            assert return_dict is False
-            return (torch.zeros(1, 3, 32, 48),)
-
-    class FakeProcessor:
-        @staticmethod
-        def preprocess(image, *, height, width):
-            seen["preprocess"] = (image.size, height, width)
-            return torch.zeros(1, 3, height, width)
-
-        @staticmethod
-        def postprocess(decoded, *, output_type):
-            assert output_type == "pil"
-            return [Image.new("RGB", (decoded.shape[3], decoded.shape[2]))]
-
-    pipe = type("FakePipe", (), {"vae": FakeVae(), "image_processor": FakeProcessor()})()
-    pipeline = ChromaZImagePipeline(device="cpu", torch_dtype=torch.float32)
-    monkeypatch.setattr(pipeline, "_load_global", lambda: pipe)
-
-    result = pipeline._vae_roundtrip(Image.new("RGB", (35, 19)))
-
-    assert seen["preprocess"] == ((48, 32), 32, 48)
-    assert seen["encoded_shape"] == (1, 3, 32, 48)
-    assert seen["mode_calls"] == 1
-    assert result.size == (35, 19)
-
-
-def test_chroma_google_face_content_gets_the_lower_adaptive_floor():
-    """Face content on Google SynthID clears at 0.125, not 0.40.
-
-    The four-fixture calibration showed a clean face-count split: both zero-face
-    text cards need 0.25, both face fixtures clear at 0.12. The adaptive arm
-    saves face identity (0.611 vs 0.279 cosine) and 6.4 dB PSNR on face content.
+    Fresh face-plus-text carriers required more denoise than the old face-only
+    fixtures, and the runtime has no measured text classifier. Keep every Google
+    carrier on the replicated global floor until a richer content gate exists.
    """
    from remove_ai_watermarks._internal.watermark_profiles import (
-        CHROMA_ZIMAGE_GOOGLE_FACE_STRENGTH,
        CHROMA_ZIMAGE_GOOGLE_STRENGTH,
        resolve_strength,
    )

-    assert pytest.approx(0.125) == CHROMA_ZIMAGE_GOOGLE_FACE_STRENGTH
-    # Face content gets the lower adaptive floor.
-    assert resolve_strength(None, "google", "chroma-zimage", face_count=17) == pytest.approx(0.125)
-    assert resolve_strength(None, "google", "chroma-zimage", face_count=1) == pytest.approx(0.125)
-    # Zero-face content gets the flat floor (text/graphic cohort).
-    assert resolve_strength(None, "google", "chroma-zimage", face_count=0) == pytest.approx(0.40)
-    assert pytest.approx(0.40) == CHROMA_ZIMAGE_GOOGLE_STRENGTH
-    # No face_count hint: conservative (flat floor).
-    assert resolve_strength(None, "google", "chroma-zimage") == pytest.approx(0.40)
-    # Other vendors ignore face_count: the split is Google-SynthID-specific.
-    assert resolve_strength(None, "openai", "chroma-zimage", face_count=5) == pytest.approx(0.20)
-    assert resolve_strength(None, "meta", "chroma-zimage", face_count=5) == pytest.approx(0.17)
-    # An explicit strength still wins over the adaptive arm.
-    assert resolve_strength(0.20, "google", "chroma-zimage", face_count=17) == pytest.approx(0.20)
-    # Other profiles are untouched by the face_count parameter.
-    assert resolve_strength(None, "google", "qwen-zimage", size=(2000, 1850), face_count=17) == pytest.approx(0.27)
+    assert pytest.approx(0.50) == CHROMA_ZIMAGE_GOOGLE_STRENGTH
+    assert resolve_strength(None, "google", "chroma-zimage") == pytest.approx(0.50)
+    assert resolve_strength(None, "openai", "chroma-zimage") == pytest.approx(0.20)
+    assert resolve_strength(None, "meta", "chroma-zimage") == pytest.approx(0.17)
+    # An explicit strength still wins over the default.
+    assert resolve_strength(0.20, "google", "chroma-zimage") == pytest.approx(0.20)


 def test_auto_profile_routes_to_the_measured_engine_per_vendor():
-    """--pipeline auto picks the engine the four-cohort calibration supports:
-    chroma-zimage for Microsoft, qwen-zimage for OpenAI/Google/Meta/unknown."""
+    """--pipeline auto picks the best quality-qualified measured engine."""
    from remove_ai_watermarks._internal.watermark_profiles import (
        CHROMA_ZIMAGE_PROFILE,
        QWEN_ZIMAGE_PROFILE,
+        SDXL_ZIMAGE_PROFILE,
        normalize_profile,
        resolve_auto_profile,
    )
@@ -1325,12 +1323,21 @@ def test_auto_profile_routes_to_the_measured_engine_per_vendor():
    assert normalize_profile("auto") == "auto"
    assert resolve_auto_profile("openai") == QWEN_ZIMAGE_PROFILE
    assert resolve_auto_profile("microsoft") == CHROMA_ZIMAGE_PROFILE
-    assert resolve_auto_profile("google") == QWEN_ZIMAGE_PROFILE
+    assert resolve_auto_profile("google") == SDXL_ZIMAGE_PROFILE
    assert resolve_auto_profile("meta") == QWEN_ZIMAGE_PROFILE
    assert resolve_auto_profile(None) == QWEN_ZIMAGE_PROFILE
    assert resolve_auto_profile("unknown") == QWEN_ZIMAGE_PROFILE


+def test_auto_strength_uses_the_effective_vendor_profile():
+    """Displayed and executed auto strengths must use the same engine policy."""
+    from remove_ai_watermarks._internal.watermark_profiles import resolve_strength
+
+    assert resolve_strength(None, "google", "auto", size=(64, 48)) == pytest.approx(0.50)
+    assert resolve_strength(None, "microsoft", "auto", size=(64, 48)) == pytest.approx(0.125)
+    assert resolve_strength(None, "openai", "auto", size=(64, 48)) == pytest.approx(0.07675)
+
+
 def test_auto_profile_strength_uses_the_resolved_engine_floors(tmp_path, monkeypatch):
    """Auto resolves OpenAI to Qwen and uses its oracle-verified 0.07675 floor."""
    from remove_ai_watermarks._internal.watermark_remover import WatermarkRemover
@@ -1340,7 +1347,8 @@ def test_auto_profile_strength_uses_the_resolved_engine_floors(tmp_path, monkeyp
    Image.new("RGB", (64, 48)).save(source)

    remover = WatermarkRemover(device="cuda", pipeline="auto")
-    assert remover.model_profile == "auto"
+    assert remover.configured_profile == "auto"
+    assert remover.model_profile == "qwen-zimage"

    runtime = MagicMock()
    runtime.run.return_value = Image.new("RGB", (64, 48), (50, 60, 70))
@@ -1353,6 +1361,27 @@ def test_auto_profile_strength_uses_the_resolved_engine_floors(tmp_path, monkeyp
    assert kwargs["strength"] == pytest.approx(0.07675)


+def test_auto_google_switches_to_the_sdxl_floor_and_dtype(tmp_path, monkeypatch):
+    """Auto must switch both policy and weights dtype when Google selects SDXL."""
+    from remove_ai_watermarks._internal import watermark_remover
+    from remove_ai_watermarks._internal.watermark_remover import WatermarkRemover
+
+    _mock_watermark_runtime_deps(monkeypatch)
+    source = tmp_path / "source.png"
+    Image.new("RGB", (64, 48)).save(source)
+
+    remover = WatermarkRemover(device="cuda", pipeline="auto")
+    runtime = MagicMock()
+    runtime.run.return_value = Image.new("RGB", (64, 48), (50, 60, 70))
+    monkeypatch.setattr(remover, "_load_qwen_zimage_pipeline", lambda: runtime)
+
+    remover.remove_watermark(source, vendor="google")
+
+    assert remover.model_profile == "sdxl-zimage"
+    assert remover.torch_dtype == watermark_remover.torch.float16
+    assert runtime.run.call_args.kwargs["strength"] == pytest.approx(0.50)
+
+
 def test_auto_profile_text_manifest_uses_the_measured_qwen_engine_once(tmp_path, monkeypatch):
    """A manifest does not override the researched OpenAI-to-Qwen decision."""
    from remove_ai_watermarks._internal.text_restoration import VerifiedTextLine, VerifiedTextManifest
@@ -1373,3 +1402,30 @@ def test_auto_profile_text_manifest_uses_the_measured_qwen_engine_once(tmp_path,
    assert remover.model_profile == "qwen-zimage"
    assert runtime.run.call_count == 1
    assert runtime.run.call_args.kwargs["text_manifest"] is manifest
+
+
+def test_auto_profile_is_resolved_fresh_for_each_image(tmp_path, monkeypatch):
+    """A Google image must not pin the next OpenAI image to SDXL."""
+    from remove_ai_watermarks._internal.watermark_remover import WatermarkRemover
+
+    _mock_watermark_runtime_deps(monkeypatch)
+    source = tmp_path / "source.png"
+    Image.new("RGB", (64, 48)).save(source)
+    remover = WatermarkRemover(device="cuda", pipeline="auto")
+    runtimes = []
+
+    def load_runtime():
+        runtime = MagicMock()
+        runtime.run.return_value = Image.new("RGB", (64, 48))
+        runtimes.append(runtime)
+        return runtime
+
+    monkeypatch.setattr(remover, "_load_qwen_zimage_pipeline", load_runtime)
+
+    remover.remove_watermark(source, tmp_path / "google.png", vendor="google")
+    assert remover.model_profile == "sdxl-zimage"
+    assert runtimes[-1].run.call_args.kwargs["strength"] == pytest.approx(0.50)
+
+    remover.remove_watermark(source, tmp_path / "openai.png", vendor="openai")
+    assert remover.model_profile == "qwen-zimage"
+    assert runtimes[-1].run.call_args.kwargs["strength"] == pytest.approx(0.07675)
@@ -12,6 +12,8 @@ sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "scripts"))

 import corpus_gap_scan as gap
 import sidecar_regression as regression
+import visible_eval
+import visible_positives as visible


 def _report(**overrides):
@@ -70,22 +72,113 @@ def test_gap_checkpoint_keeps_complete_rows_and_ignores_a_torn_tail(tmp_path):
    assert gap._read_checkpoint(checkpoint) == {row["path"]: row, second["path"]: second}


+def test_visible_checkpoint_is_repaired_before_resume(tmp_path):
+    checkpoint = tmp_path / "visible.jsonl"
+    first = {"path": "/corpus/2026-09-10/first.png", "keys": [], "status": "ok"}
+    checkpoint.write_text(json.dumps(first) + '\n{"path":"torn', encoding="utf-8")
+
+    visible._repair_checkpoint(checkpoint)
+    with checkpoint.open("a", encoding="utf-8") as stream:
+        stream.write(json.dumps({"path": "/corpus/2026-09-10/second.png", "keys": [], "status": "ok"}) + "\n")
+
+    assert [row["path"] for row in visible._read_records(checkpoint)] == [
+        "/corpus/2026-09-10/first.png",
+        "/corpus/2026-09-10/second.png",
+    ]
+
+
+def test_visible_since_filter_uses_corpus_date_directories(tmp_path):
+    old = tmp_path / "2026-09-01" / "old.png"
+    current = tmp_path / "2026-09-10" / "current.png"
+    nested = tmp_path / "2026-09-10" / "nested" / "current.webp"
+    undated = tmp_path / "loose.png"
+    for path in (old, current, nested, undated):
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.touch()
+
+    assert visible._files(tmp_path, date(2026, 9, 10)) == [current, nested]
+
+
+def test_visible_discovery_candidates_include_unmapped_provenance_and_unbiased_quiet_rows():
+    rows = [
+        {"path": "known.png", "keys": ["qwen"], "status": "ok", "uscc": "unmapped-a"},
+        {"path": "a.png", "keys": [], "status": "ok", "uscc": "unmapped-a"},
+        {"path": "b.png", "keys": [], "status": "ok", "platform": "Google Gemini"},
+        {"path": "c.png", "keys": [], "status": "ok"},
+        {"path": "bad.png", "keys": [], "status": "unreadable", "uscc": "unmapped-b"},
+    ]
+
+    selected = visible._discovery_candidates(rows, per_cohort=2, random_quiet=10, seed=7)
+
+    assert {row["path"] for row in selected} == {"a.png", "b.png", "c.png"}
+    assert next(row for row in selected if row["path"] == "a.png")["discovery_strata"] == [
+        "tc260:unmapped-a",
+        "unbiased-quiet",
+    ]
+
+
+def test_visible_evaluation_inventory_follows_the_registry():
+    from remove_ai_watermarks.watermark_registry import mark_keys
+
+    assert tuple(mark_keys()) == visible_eval.MARKS
+
+
 def test_gap_worker_error_is_a_complete_resumable_row(monkeypatch, tmp_path):
    path = tmp_path / "bad.bin"
    path.write_bytes(b"not an image")
    calls = []

-    def fail(_path, **kwargs):
-        calls.append(kwargs)
+    def fail(_path):
+        calls.append(_path)
        raise ValueError("broken\ninput")

-    monkeypatch.setattr(gap, "identify", fail)
+    monkeypatch.setattr(gap, "extract_provenance_evidence", fail)
    row = gap._scan_one((str(path), "bad.bin", "1.2.3"))

    assert set(row) == set(gap.REPORT_FIELDS)
    assert row["candidate_classes"] == "identify_error"
    assert row["error"] == "ValueError: broken input"
-    assert calls == [{"check_visible": False, "check_invisible": False}]
+    assert calls == [path]
+
+
+def test_gap_worker_records_c2pa_review_fields(monkeypatch, tmp_path):
+    path = tmp_path / "signed.png"
+    path.touch()
+    evidence = SimpleNamespace(
+        c2pa_info={
+            "issuer": "Example signer",
+            "claim_generator": "Example generator",
+            "actions": "created",
+            "source_type": "trainedAlgorithmicMedia (AI-generated)",
+            "c2pa_validation_state": "Valid",
+        },
+        scan=b"c2pa",
+    )
+    monkeypatch.setattr(gap, "extract_provenance_evidence", lambda _path: evidence)
+    monkeypatch.setattr(
+        gap,
+        "identify_from_evidence",
+        lambda actual: (
+            SimpleNamespace(
+                is_ai_generated=True,
+                platform="Example",
+                confidence="high",
+                watermarks=["C2PA"],
+                signals=[SimpleNamespace(name="c2pa")],
+                integrity_clashes=[],
+            )
+            if actual is evidence
+            else None
+        ),
+    )
+
+    row = gap._scan_one((str(path), "signed.png", "1.2.3"))
+
+    assert row["c2pa_issuer"] == "Example signer"
+    assert row["c2pa_claim_generator"] == "Example generator"
+    assert row["c2pa_actions"] == "created"
+    assert row["c2pa_source_type"] == "trainedAlgorithmicMedia (AI-generated)"
+    assert row["c2pa_validation"] == "Valid"


 def test_gap_summary_does_not_report_a_negative_hidden_count(capsys):
@@ -0,0 +1,123 @@
+"""Unit contracts for the lightweight source-pipeline runtime."""
+
+from __future__ import annotations
+
+import re
+from hashlib import sha256
+from typing import TYPE_CHECKING
+
+import numpy as np
+import pytest
+
+from remove_ai_watermarks import source_classify
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+
+def test_model_identity_is_immutable() -> None:
+    assert re.fullmatch(r"[0-9a-f]{40}", source_classify.WEIGHTS_REVISION)
+    assert re.fullmatch(r"[0-9a-f]{64}", source_classify.MODEL_SHA256)
+
+
+def _artifact(path: Path, *, activation: str = "gelu_exact", output_bias: tuple[float, ...] = (0.0, 0.0, 0.0)) -> None:
+    np.savez_compressed(
+        path,
+        classes=np.asarray(source_classify.CLASSES),
+        kind=np.asarray("fused"),
+        mean=np.zeros(source_classify.FEATURE_WIDTH),
+        scale=np.ones(source_classify.FEATURE_WIDTH),
+        margins=np.asarray((0.5, 0.5)),
+        layer0_weight=np.zeros((128, source_classify.FEATURE_WIDTH), dtype=np.float32),
+        layer0_bias=np.zeros(128, dtype=np.float32),
+        layer1_weight=np.zeros((64, 128), dtype=np.float32),
+        layer1_bias=np.zeros(64, dtype=np.float32),
+        output_weight=np.zeros((3, 64), dtype=np.float32),
+        output_bias=np.asarray(output_bias, dtype=np.float32),
+        activation=np.asarray(activation),
+    )
+
+
+def test_loader_accepts_the_exact_pickle_free_schema(tmp_path: Path) -> None:
+    path = tmp_path / source_classify.MODEL_FILE
+    _artifact(path)
+
+    model = source_classify._load_model(path, expected_sha256=None)
+
+    assert model.mean.shape == (source_classify.FEATURE_WIDTH,)
+    assert model.margins.tolist() == pytest.approx([0.5, 0.5])
+
+
+def test_loader_rejects_the_mislabeled_activation(tmp_path: Path) -> None:
+    path = tmp_path / source_classify.MODEL_FILE
+    _artifact(path, activation="gelu_tanh")
+
+    with pytest.raises(ValueError, match="activation must be gelu_exact"):
+        source_classify._load_model(path, expected_sha256=None)
+
+
+def test_provider_threshold_mutation_changes_the_real_decision(tmp_path: Path) -> None:
+    path = tmp_path / source_classify.MODEL_FILE
+    _artifact(path, output_bias=(2.0, 0.0, 0.0))
+    model = source_classify._load_model(path, expected_sha256=None)
+    features = np.zeros(source_classify.FEATURE_WIDTH, dtype=np.float32)
+    scores = source_classify._scores(features, model)
+
+    accepted = source_classify._result(scores, np.asarray((1.5, 0.5), dtype=np.float32))
+    rejected = source_classify._result(scores, np.asarray((2.5, 0.5), dtype=np.float32))
+
+    assert accepted.label == "openai"
+    assert accepted.reason == "classified"
+    assert rejected.label == "unknown"
+    assert rejected.reason == "abstained"
+
+
+def test_a_float32_threshold_tie_preserves_the_inclusive_training_decision() -> None:
+    scores = np.asarray((2.0, 0.0, 0.0), dtype=np.float32)
+    result = source_classify._result(scores, np.asarray((2.0, 0.5), dtype=np.float32))
+
+    assert result.label == "openai"
+    assert result.reason == "classified"
+
+
+def test_small_image_returns_honest_unknown(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    from PIL import Image
+
+    weights = tmp_path / "weights"
+    weights.mkdir()
+    _artifact(weights / source_classify.MODEL_FILE)
+    monkeypatch.setenv(source_classify.WEIGHTS_ENV, str(weights))
+    monkeypatch.setattr(
+        source_classify,
+        "MODEL_SHA256",
+        sha256((weights / source_classify.MODEL_FILE).read_bytes()).hexdigest(),
+    )
+    image = tmp_path / "small.png"
+    Image.new("RGB", (128, 128), "white").save(image)
+
+    result = source_classify.classify_source(image)
+
+    assert result.label == "unknown"
+    assert result.reason == "feature_unavailable"
+    assert result.scores == {}
+
+
+def test_model_hash_is_a_runtime_precondition(tmp_path: Path) -> None:
+    path = tmp_path / source_classify.MODEL_FILE
+    _artifact(path)
+
+    with pytest.raises(ValueError, match="model SHA-256 mismatch"):
+        source_classify._load_model(path, expected_sha256="0" * 64)
+
+
+def test_spectral_feature_is_finite_and_has_the_frozen_width() -> None:
+    from PIL import Image
+
+    from remove_ai_watermarks._internal.source_spectral import feature_from_image, make_geometry
+
+    rng = np.random.default_rng(20260916)
+    pixels = rng.integers(0, 256, size=(300, 400, 3), dtype=np.uint8)
+    feature = feature_from_image(Image.fromarray(pixels), make_geometry())
+
+    assert feature.shape == (768,)
+    assert np.all(np.isfinite(feature))
@@ -0,0 +1,119 @@
+"""Tests for the abstaining source-pipeline classifier research tool."""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+import pytest
+from click.testing import CliRunner
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "scripts"))
+
+import source_pipeline_classifier as classifier
+
+
+def test_command_help_exits_cleanly() -> None:
+    result = CliRunner().invoke(classifier.main, ["--help"])
+
+    assert result.exit_code == 0
+    assert "Usage:" in result.output
+    assert "train" in result.output
+    assert "predict" in result.output
+
+
+def _training_rows() -> tuple[np.ndarray, np.ndarray]:
+    rng = np.random.default_rng(20260915)
+    rows = np.vstack(
+        (
+            rng.normal((-3.0, 0.0, 0.0), 0.1, size=(12, 3)),
+            rng.normal((3.0, 0.0, 0.0), 0.1, size=(12, 3)),
+            rng.normal((0.0, 3.0, 0.0), 0.1, size=(12, 3)),
+        )
+    )
+    labels = np.repeat(np.arange(3), 12)
+    return rows, labels
+
+
+def test_fit_and_abstention_separate_provider_rows() -> None:
+    features, labels = _training_rows()
+    model = classifier.fit_ridge(features, labels, penalty=1.0)
+    raw_scores = classifier.scores(features, model)
+    margin = classifier.calibrate_margin(raw_scores, labels, unknown_fpr=0.01)
+    predictions = classifier.predict_scores(raw_scores, margin)
+
+    assert np.array_equal(predictions[:24], labels[:24])
+    assert np.all(predictions[24:] == classifier.UNKNOWN_INDEX)
+
+
+def test_model_round_trip_is_pickle_free_and_strict(tmp_path: Path) -> None:
+    grid = 4
+    feature_count = classifier.FEATURE_WIDTH + 3 * grid * grid
+    model = classifier.RidgeModel(
+        mean=np.arange(feature_count, dtype=np.float64),
+        scale=np.full(feature_count, 2.0),
+        weights=np.arange(feature_count * 3, dtype=np.float64).reshape(feature_count, 3),
+        margin=0.25,
+        grid=grid,
+        penalty=100.0,
+    )
+    path = tmp_path / "model.npz"
+
+    classifier.save_model(path, model)
+    loaded = classifier.load_model(path)
+
+    assert loaded.grid == grid
+    assert loaded.margin == model.margin
+    assert np.array_equal(loaded.mean, model.mean)
+    assert np.array_equal(loaded.scale, model.scale)
+    assert np.array_equal(loaded.weights, model.weights)
+
+    with np.load(path, allow_pickle=False) as payload:
+        assert payload.files
+
+
+def test_model_loader_rejects_extra_fields(tmp_path: Path) -> None:
+    path = tmp_path / "model.npz"
+    np.savez(
+        path,
+        classes=np.asarray(classifier.CLASSES),
+        mean=np.zeros(136),
+        scale=np.ones(136),
+        weights=np.zeros((136, 3)),
+        margin=np.asarray(0.1),
+        spectral_grid=np.asarray(2),
+        penalty=np.asarray(100.0),
+        unexpected=np.asarray(1),
+    )
+
+    with pytest.raises(ValueError, match="model fields must be exactly"):
+        classifier.load_model(path)
+
+
+def test_model_loader_rejects_nonfinite_margin(tmp_path: Path) -> None:
+    grid = 2
+    feature_count = classifier.FEATURE_WIDTH + 3 * grid * grid
+    path = tmp_path / "model.npz"
+    np.savez(
+        path,
+        classes=np.asarray(classifier.CLASSES),
+        mean=np.zeros(feature_count),
+        scale=np.ones(feature_count),
+        weights=np.zeros((feature_count, 3)),
+        margin=np.asarray(np.nan),
+        spectral_grid=np.asarray(grid),
+        penalty=np.asarray(100.0),
+    )
+
+    with pytest.raises(ValueError, match="margin must be finite"):
+        classifier.load_model(path)
+
+
+def test_calibration_requires_unknown_rows() -> None:
+    with pytest.raises(ValueError, match="must contain unknown"):
+        classifier.calibrate_margin(
+            np.zeros((2, 3)),
+            np.asarray([0, 1]),
+            unknown_fpr=0.01,
+        )
@@ -0,0 +1,81 @@
+"""Tests for the phase-free spectral pipeline research harness."""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "scripts"))
+
+import spectral_pipeline_probe as probe
+
+
+def test_log_spectrum_discards_translation_phase() -> None:
+    rng = np.random.default_rng(17)
+    channel = rng.normal(size=(32, 32))
+
+    original = probe.phase_free_log_spectrum(channel)
+    translated = probe.phase_free_log_spectrum(np.roll(channel, shift=(5, -7), axis=(0, 1)))
+
+    assert np.allclose(original, translated, rtol=1e-12, atol=1e-12)
+
+
+def test_spectral_feature_is_finite_and_has_declared_shape() -> None:
+    geometry = probe.make_geometry(size=32, grid=4)
+    yy, xx = np.mgrid[:32, :32]
+    carrier = np.cos(2 * np.pi * (5 * xx + 3 * yy) / 32)
+    pixels = np.stack((carrier, np.roll(carrier, 1, axis=0), np.roll(carrier, 2, axis=1)), axis=2)
+
+    feature = probe.spectral_feature(pixels, geometry)
+
+    assert feature.shape == (3 * 4 * 4,)
+    assert np.all(np.isfinite(feature))
+    assert np.linalg.norm(feature) > 0
+
+
+def test_discovery_keeps_only_stems_shared_by_every_class(tmp_path: Path) -> None:
+    for class_name in ("first", "second"):
+        directory = tmp_path / class_name
+        directory.mkdir()
+        for stem in ("001", "002"):
+            Image.new("RGB", (8, 8), "white").save(directory / f"{stem}.png")
+    Image.new("RGB", (8, 8), "black").save(tmp_path / "first" / "unpaired.png")
+
+    paths, labels, groups, names = probe.discover_grouped_paths(tmp_path, ("first", "second"))
+
+    assert names == ["001", "002"]
+    assert [path.stem for path in paths] == ["001", "002", "001", "002"]
+    assert labels.tolist() == [0, 0, 1, 1]
+    assert groups.tolist() == [0, 1, 0, 1]
+
+
+def test_grouped_ridge_never_trains_on_the_test_group() -> None:
+    groups = np.tile(np.arange(14), 3)
+    labels = np.repeat(np.arange(3), 14)
+    class_centers = np.eye(3)[labels] * 8.0
+    group_nuisance = np.column_stack((groups % 2, groups % 3, groups % 5)) * 0.01
+    features = class_centers + group_nuisance
+
+    models = probe.fit_grouped_models(features, labels, groups, class_count=3, folds=5, ridge=1.0)
+    predictions = probe.predict_grouped(models, features)
+
+    assert np.array_equal(predictions, labels)
+    fold_assignment = np.full(len(groups), -1)
+    for fold, model in enumerate(models):
+        assert not np.any(fold_assignment[model.test_rows] >= 0)
+        fold_assignment[model.test_rows] = fold
+    assert np.all(fold_assignment >= 0)
+    for group in np.unique(groups):
+        assert len(np.unique(fold_assignment[groups == group])) == 1
+
+
+def test_confusion_matrix_uses_actual_rows_and_predicted_columns() -> None:
+    labels = np.asarray([0, 0, 1, 1, 2])
+    predictions = np.asarray([0, 1, 1, 2, 0])
+
+    matrix = probe.confusion_matrix(labels, predictions, class_count=3)
+
+    assert matrix.tolist() == [[1, 1, 0], [0, 1, 1], [1, 0, 0]]
@@ -3449,7 +3449,7 @@ wheels = [

 [[package]]
 name = "remove-ai-watermarks"
-version = "0.40.2"
+version = "0.41.0"
 source = { editable = "." }
 dependencies = [
    { name = "c2pa-python" },
@@ -3584,6 +3584,13 @@ qwen-zimage = [
    { name = "torchvision" },
    { name = "transformers" },
 ]
+source-classify = [
+    { name = "huggingface-hub" },
+    { name = "numpy", version = "1.26.4", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.13'" },
+    { name = "numpy", version = "2.3.5", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.13'" },
+    { name = "opencv-python-headless", version = "4.11.0.86", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version < '3.13'" },
+    { name = "opencv-python-headless", version = "4.14.0.94", source = { registry = "https://pypi.org/simple" }, marker = "python_full_version >= '3.13'" },
+]
 text-draft = [
    { name = "huggingface-hub" },
    { name = "paddleocr" },
@@ -3637,6 +3644,7 @@ requires-dist = [
    { name = "huggingface-hub", marker = "extra == 'dev'", specifier = ">=0.20.0" },
    { name = "huggingface-hub", marker = "extra == 'lama'", specifier = ">=0.20.0" },
    { name = "huggingface-hub", marker = "extra == 'migan'", specifier = ">=0.20.0" },
+    { name = "huggingface-hub", marker = "extra == 'source-classify'", specifier = ">=0.20.0" },
    { name = "huggingface-hub", marker = "extra == 'text-draft'", specifier = ">=0.20.0" },
    { name = "invisible-watermark", marker = "extra == 'dev'", specifier = ">=0.2.0" },
    { name = "numpy", marker = "python_full_version >= '3.13' and extra == 'pixels'", specifier = ">=2" },
@@ -3665,10 +3673,11 @@ requires-dist = [
    { name = "remove-ai-watermarks", extras = ["pixels"], marker = "extra == 'classify'" },
    { name = "remove-ai-watermarks", extras = ["pixels"], marker = "extra == 'detect'" },
    { name = "remove-ai-watermarks", extras = ["pixels"], marker = "extra == 'diffusion'" },
+    { name = "remove-ai-watermarks", extras = ["pixels"], marker = "extra == 'source-classify'" },
    { name = "remove-ai-watermarks", extras = ["pixels"], marker = "extra == 'visible'" },
    { name = "remove-ai-watermarks", extras = ["qwen-zimage", "lama"], marker = "extra == 'text-restoration'" },
    { name = "remove-ai-watermarks", extras = ["video"], marker = "extra == 'dev'" },
-    { name = "remove-ai-watermarks", extras = ["video", "heif", "detect", "trustmark", "text-restoration", "migan", "classify"], marker = "extra == 'all'" },
+    { name = "remove-ai-watermarks", extras = ["video", "heif", "detect", "trustmark", "text-restoration", "migan", "classify", "source-classify"], marker = "extra == 'all'" },
    { name = "remove-ai-watermarks", extras = ["visible"], marker = "extra == 'lama'" },
    { name = "remove-ai-watermarks", extras = ["visible"], marker = "extra == 'migan'" },
    { name = "remove-ai-watermarks", extras = ["visible"], marker = "extra == 'video'" },
@@ -3684,7 +3693,7 @@ requires-dist = [
    { name = "transformers", marker = "extra == 'diffusion'", specifier = ">=5,<6" },
    { name = "trustmark", marker = "python_full_version < '3.13' and extra == 'trustmark'", specifier = ">=0.8.0" },
 ]
-provides-extras = ["pixels", "heif", "visible", "video", "detect", "diffusion", "qwen-zimage", "text-restoration", "text-draft", "trustmark", "classify", "classify-onnx", "lama", "migan", "dev", "all"]
+provides-extras = ["pixels", "heif", "visible", "video", "detect", "diffusion", "qwen-zimage", "text-restoration", "text-draft", "trustmark", "classify", "classify-onnx", "source-classify", "lama", "migan", "dev", "all"]

 [[package]]
 name = "requests"
Author	SHA1	Message	Date
Victor Kuznetsov	c7021b900b	Add source pipeline classifier runtime	2026-09-15 23:17:48 -07:00
Victor Kuznetsov	3f55df6f99	Correct source classifier activation schema	2026-09-15 22:55:46 -07:00
Victor Kuznetsov	4795149d34	Publish source classifier model card	2026-09-15 22:36:34 -07:00
Victor Kuznetsov	cf30b4f95a	Release 0.40.6	2026-09-15 22:00:05 -07:00
Victor Kuznetsov	8ce8bf14da	Document GPT Image 2.5 compatibility	2026-09-15 21:32:46 -07:00
Victor Kuznetsov	3fb94b220d	Wait for PyPI simple index before distribution	2026-09-15 21:05:14 -07:00
Victor Kuznetsov	d2c5aa66fc	Merge remote-tracking branch 'origin/main'	2026-09-15 19:21:01 -07:00
Victor Kuznetsov	380820fa7e	Add SDXL verified-text restoration	2026-09-15 19:20:37 -07:00
Victor Kuznetsov	12e9a87865	Publish abstaining source pipeline classifier	2026-09-15 19:18:35 -07:00
Victor Kuznetsov	19fccf27cb	Release 0.40.4 with calibrated invisible profiles	2026-09-15 16:49:41 -07:00
Victor Kuznetsov	a385aab433	Calibrate provider-specific invisible removal profiles	2026-09-15 16:38:17 -07:00
Victor Kuznetsov	1d7a6270d4	Add phase-free spectral pipeline probe	2026-09-15 16:20:31 -07:00
Victor Kuznetsov	b5cd16aef0	Add local Microsoft InvisMark payload disruption	2026-09-14 12:07:00 -07:00
Victor Kuznetsov	bf63e8e187	Attribute discovered C2PA signers	2026-09-14 11:51:19 -07:00
Victor Kuznetsov	d4181f3277	Complete visible signal discovery scanner	2026-09-13 17:46:29 -07:00