remove-ai-watermarks/docs/verification-plan.md

# Full verification plan

How we convince ourselves the library actually works, across its whole surface, on real data.

This is the pre-release and periodic-audit plan. It is deliberately organized by **oracle
strength** rather than by module, because the hard part is never "call the function" -- it
is "know what the right answer was". A sweep with no oracle proves only that nothing threw.

Measured throughput on the local corpus (39,430 images, M-series, 2026-07-19):

| path | per image | full corpus, 8 procs |
|---|---|---|
| `identify` | 0.58 s | ~0.8 h |
| `detect_marks` | 0.28 s | ~0.4 h |
| visible remove (cv2) | 0.58 s | ~0.8 h |
| diffusion @512px (MPS) | ~50 s | ~23 days -- sample only |

So every CPU path is affordable at FULL corpus scale; only the diffusion paths need
sampling. Plan accordingly: never sample where a full sweep costs an hour.

## Data sources

| source | size | committed | role |
|---|---|---|---|
| `data/spaces/originals/` | 39,430 imgs, 87.5 GB | no (gitignored) | the real-upload corpus |
| `data/spaces/identify/` | 39,314 JSON sidecars | no | **recorded `identify` verdict per image** |
| `data/spaces/_visible_datasets/` | 3,741 imgs, per vendor | no | mark-positive pools |
| `data/synthid_corpus/` | 39 imgs, labelled | yes | pos/neg/cleaned SynthID references |
| `data/samples/` | 11 fixtures | yes | deterministic fixtures |
| `data/*_capture/` | vendor captures | yes | detection silhouettes |
| synthesized | generated | no | constructed ground truth (tier B) |

**Data safety.** The corpus is user uploads: local analysis only. No run may copy, promote,
or commit corpus images into a tracked path, and no report may embed them. All harness
output goes to gitignored paths under `data/spaces/`.

## Tier A -- self-evident oracles (full corpus, unattended)

Properties that are true or false without anyone labelling anything. These are the
backbone: they scale to 39k images and catch regressions with zero human cost.

### A1. Sidecar regression -- the highest-value check we are not running

`data/spaces/identify/` holds 39,314 recorded `identify` verdicts, keyed by the same uid as
the image. Re-running `identify` today and diffing against them turns the corpus into a
**39k-image behavioral regression suite for free**. Any drift in verdict, platform,
confidence or signal set shows up as a diff, bucketed by cause.

Caveat that makes this honest: a diff is not automatically a bug -- the sidecars were
written by older versions, so intended improvements also show up. The output is therefore a
**classified diff** (new detections / lost detections / changed platform / changed
confidence), reviewed once, then re-baselined. Lost detections are the alarm.

Implemented as `scripts/sidecar_regression.py` (resumable, ~1.5 h at 8 workers).

#### First full run, 2026-07-19, all 39,314 sidecars

| class | n | share |
|---|---|---|
| unchanged | 37,326 | 94.9% |
| lost_signal | 1,302 | 3.3% |
| platform changed | 922 | 2.3% |
| confidence changed | 899 | 2.3% |
| new_signal | 694 | 1.8% |
| lost_ai | 747 | 1.9% |
| new_ai | 152 | 0.4% |

Two results worth keeping:

**No metadata signal regressed anywhere.** Lost families were exclusively visual
(`visible_sparkle` 1,256, `visible_doubao` 45, `visible_jimeng` 4) -- zero c2pa, synthid,
aigc_tc260, iptc, exif_generator or xai_signature losses across the whole corpus. And
`identify` raised on **none** of the 39,314 real uploads.

**The sparkle losses are mostly corrected false fires, but not entirely.** Sampling 400 of
the 1,256 and checking whether the file still carries Google provenance independently of
the sparkle: 12.5% (95% CI 9.6-16.1%) still do, i.e. ~120-200 corpus-wide are **genuine
misses**; the other ~1,050-1,135 had nothing backing them. Read the split as a trade the
FP-gate tightening made, not as a clean win.

Caveat on that split: "no Google provenance" is not proof of a false positive -- a
metadata-stripped Gemini screenshot also has none while still carrying the pixel sparkle.
So 87.5% is an **upper bound** on false fires; only the 12.5% genuine-miss figure is solid.

Doubao moved the other way (-45 / +628 net +583), which is the `scale_basis` landscape fix
showing up at corpus scale. `trustmark` +38 and `open_invisible` +9 are not behavior: those
extras were simply not installed when the sidecars were written.

### A2. Parity: whatever we detect, we must be able to remove

For every image where a signal fires: remove, re-scan with the same oracle, assert quiet.
- metadata: `scripts/metadata_removal_audit.py` (exists) -- run full corpus.
- visible: `scripts/visible_removal_audit.py` (exists) -- run **once per backend**
  (cv2 / migan / lama). It is single-process, so a full-corpus sweep is ~10 h and three
  backends ~30 h. Its expensive half is DETECTION, which does not depend on the backend,
  so run `scripts/visible_positives.py` once (parallel, ~40 min) and feed the result to
  the audit's `--paths-file` seam: a few thousand images per backend instead of 39k.

#### Metadata parity, first full run, 2026-07-19 (20,153 carriers + 1,500 clean controls)

Zero scan/strip/decode errors. Survival after strip:

| signal | carriers | survived |
|---|---|---|
| c2pa_manifest / claim_generator | 15,410 | **3** |
| synthid_watermark | 14,985 | 0 |
| aigc_label | 4,414 | 0 |
| the other 12 signal types | - | 0 |

The no-op control is clean: the strip **added** a signal to 0 of 1,500 clean images. Two
real defects fell out of the run.

**Defect 1 -- the fail-safe reports success on a file it did not strip.** All 3 parity
failures are Samsung Galaxy S22 camera PNGs (`Galaxy S22 c2pa-rs/0.37.0`) whose `caBX`
chunk survives. Cause: PIL raises `UnidentifiedImageError` on them, so
`remove_ai_metadata`'s fail-safe copies the file through byte-identical -- correct in
intent (never crash a worker on a partial upload) but it returns an output path
indistinguishable from a real strip. User-visible: `metadata --remove` prints
"AI metadata stripped ->", exits 0, and `identify` on the output still reports C2PA. The
warning is logged but the success line contradicts it. Rare here (3 of 20,153) but the
mechanism fires on ANY file PIL cannot decode. The fail-safe should stay; what needs
fixing is that the caller cannot tell a no-op from a strip.

**Defect 2 -- 16-bit PNGs are silently downconverted to 8-bit.** 5 of the 1,500 clean
controls failed the pixel-identity check; all are 16-bit PNGs, and the PIL re-save halves
their bit depth (one went 9.2 MB -> 2.5 MB). This is the known limitation recorded in
CLAUDE.md, now measured: a byte-level IHDR scan over every corpus PNG puts it at
**42 of 27,018 (0.16%)**.

Method note worth keeping: the first attempt to reproduce Defect 2 said "pixels
identical" and nearly closed it as a harness bug. That check read both files through
`image_io.imread`, which returns 8-bit -- **the reader destroyed the very property under
test**. The audit was right because it reads via `read_bgr_and_alpha`, which preserves
uint16. When verifying a fidelity property, check that the verification path can still
represent it.

### A3. Byte-level invariants

- no-op `remove_visible` returns the ORIGINAL bytes (not a re-encode)
- pixels outside the fill mask are bit-identical to the input
- JPEG metadata strip is pixel-lossless on the DEFAULT path (`--remove-all` re-encodes by
  design -- see `metadata.py`; assert the split, not losslessness everywhere)
- lossless source formats survive a misnamed extension

### A4. Idempotence and order-independence

- `remove_visible(remove_visible(x)) == remove_visible(x)`
- `strip(remove(x)) == remove(strip(x))` in signal terms
- a second `identify` on a cleaned output reports no metadata signals

### A5. Contract sweep across every parameter choice

`scripts/smoke_matrix.py` (exists, 68 rows, 0 skipped with `--diffusion`) covers every
choice-valued flag on fixtures. Extend from fixtures to a stratified corpus slice
(~500 images spanning format x provenance x aspect ratio), asserting exit-code semantics
rather than just absence of crash.

Known trap to encode: **exit 2 is triply overloaded** (no-visible-mark, no-invisible-signal,
Click usage error). A wrapper cannot distinguish them without parsing stderr. Either the
sweep asserts on stderr, or the codes get split -- the latter is the better fix.

#### Coverage before the extension (measured 2026-07-19, not estimated)

Comparing the flags the matrix actually executed against the flags the CLI declares:
**18 of 38 had never been executed even once** -- `--pipeline`, `--strength`, `--steps`,
`--guidance-scale`, `--device`, `--model`, `--upscaler`, `--tile`/`--tile-size`/
`--tile-overlap`, `--humanize`, `--unsharp`, `--adaptive-polish`, `--controlnet-scale`,
`--min-resolution`, `--hf-token`, `--auto`, `--verbose`. Plus uncovered VALUES of covered
flags: `--backend migan|lama` had never been driven through the CLI at all (only at
library level), `erase --backend` only ever ran `cv2`, `batch --mode all` never ran, and
`--pipeline` only ever ran its default.

Whole subsystems had unit tests but no real-data run: tiling (27 unit tests, never
processed a real image through the CLI), the region-targeted composite, the ESRGAN
upscaler, and the ffmpeg audio/video strip. The gap is not "logic untested" but
"never executed on real data", which is precisely what this campaign is for.

#### Bug found by the extension: `--steps` below ~7 crashes inside torch

Effective timesteps are `int(steps * strength)`. At the vendor-adaptive default strength
(0.15, or 0.10 for OpenAI) any `--steps` under 7 rounds to **zero**, and the pipeline dies
with a raw traceback:

```
$ remove-ai-watermarks invisible img.png --steps 5
RuntimeError: cannot reshape tensor of 0 elements into shape [0, -1, 1, 512]
```

Fully valid CLI arguments, no special flags, no `--force`. The value is accepted, the
crash is a torch internal, and nothing tells the user that steps and strength interact.
Fix is either a clamp to >=1 effective step or an up-front validation naming both values.

Method note: the first run of the knob rows failed 12 times with this identical error,
which read like twelve broken features. It was one bad harness parameter (`--steps 4`)
sitting on top of one real bug. An error that is IDENTICAL across unrelated rows is
evidence of a common cause, not of many faults -- check the shared input first.

## Tier B -- constructed ground truth (automatable, no labelling)

Where reality gives no answer key, build one. This is the tier that closes the two biggest
holes: fill quality has **never** been asserted, and recall was measured once at n=240.

### B1. Fill quality with a true reference

Real marks have no clean counterpart, so quality has only ever been eyeballed. Construct it
instead: take a clean corpus image, stamp a known mark at a known position (the captured
alpha maps make this exact), remove it, and compare against the **true original**.

Yields PSNR / SSIM per `--backend` (cv2 / migan / lama), sliced by background class, which
is exactly the axis where the docs say quality varies but no number exists.

Implemented as `scripts/fill_quality.py`. Two reporting rules are load-bearing: score
INSIDE the footprint (whole-frame PSNR sits near 60 dB whatever the backend does), and use
the MEDIAN (a fill that reproduces a flat background exactly scores PSNR=inf, and one inf
makes a mean inf -- the first run reported "+inf" for every flat bucket).

#### First run, 2026-07-19, 80 verified-clean sources, n=720 measurements

Median dB recovered inside the footprint (filled PSNR minus damaged PSNR):

| mark | bg | cv2 | migan | lama |
|---|---|---|---|---|
| doubao | flat | +10.79 | **+15.62** | +13.97 |
| doubao | mid | +9.25 | **+14.71** | +14.69 |
| doubao | textured | +2.15 | +0.66 | **+1.64** |
| jimeng | flat | +5.52 | +8.80 | **+10.30** |
| jimeng | mid | +10.78 | +11.05 | **+12.33** |
| jimeng | textured | +2.16 | +1.70 | **+3.03** |

**The `auto` order is CONFIRMED.** Median per-image recovery on the textured tercile:

| mark (textured) | cv2 | migan | lama |
|---|---|---|---|
| doubao | +0.06 | +1.53 | +1.38 |
| gemini | +1.79 | +2.75 | **+5.59** |
| jimeng | +1.24 | +2.69 | **+3.34** |
| jimeng_pill | +4.27 | +3.81 | **+5.16** |

LaMa > MI-GAN > cv2 holds on 3 of the 4 marks worth filling, and MI-GAN edges LaMa on
doubao. Nothing here argues for changing `--backend auto`.

**Correction, and the statistic that caused it.** An earlier version of this section
claimed the opposite -- that MI-GAN was the WORST on texture, below cv2 -- and it was
wrong. The report computed recovery as `median(filled) - median(damaged)`: a **difference
of medians**, not the median of the per-image differences. On skewed data those are
different statistics and here they disagreed in SIGN. A paired per-mark sign test settled
it: on the textured tercile cv2 vs MI-GAN is not significant for any mark (p 0.13-1.00;
pooled n=119, cv2 wins 69, p=0.099), while MI-GAN's medians are higher for 3 of 5.

The wrong statistic nearly shipped a change to `--backend auto`, which is the resolver a
memory-constrained CPU caller depends on. Two lessons: report the median of the per-image
DIFFERENCES when the question is paired, and confirm a ranking with a paired test before
acting on a table of independently-aggregated columns.

**Every backend collapses on texture**: recovery falls from ~+10-15 dB to ~+1-3 dB and SSIM
from ~0.79-0.95 to ~0.30-0.44. The documented "textured is where fills struggle" is
confirmed, with numbers, for the first time.

**The invariant held**: 0 violations of "the fill touches nothing outside its mask" across
all 720 measurements and all three backends.

#### Visible parity, first full run, 2026-07-20 (10,593 image/mark pairs, cv2)

| mark | n | detector-clean after removal |
|---|---|---|
| jimeng | 298 | 100% (CI 98.7-100) |
| gemini | 4,974 | 95.3% (CI 94.7-95.8) |
| doubao | 2,580 | 91.8% (CI 90.7-92.8) |
| samsung | 3 | 100% (CI 43.8-100) |
| jimeng_pill | 2,738 | 32.4% -- **wrong path, see below** |
| **excluding the pill** | **7,855** | **94.3% (CI 93.8-94.8)** |

Backend-independence was CHECKED, not assumed: the same 309 pairs through cv2 / MI-GAN /
LaMa agree within one image on doubao (91.9% x3), gemini (91.4 / 92.1 / 91.4) and jimeng
(100% x3). Only the pill diverges (32.5 / 25.0 / 25.0), and 17 of the 18 disagreements are
the pill. So one backend suffices for parity; quality is B1's job, not this sweep's.

**The residual is NOT one phenomenon -- it splits by mark.** Checking whether a still-
detected mark carries vendor provenance independent of the visual detector, against the
cleaned marks as a control:

| mark | still detected | cleaned | reading |
|---|---|---|---|
| doubao | 82.9% corroborated | 80.0% | indistinguishable -- the marks are REAL |
| gemini | 39.6% | 62.4% | significantly lower -- much of it is false fires that persist |

The doubao half turned out to be the front-end mismatch bug (fixed 2026-07-20, see the
CLAUDE.md rule): `remove()` was a silent no-op on 57 of 60 sampled cases because the
binarized mask came back empty. After the fix, 60/60 of those clear, so doubao parity
should now sit near 99%; **the sweep has not been re-run to confirm that**.

Method note: two intermediate checks were worthless and both looked meaningful.
`MarkDetection.region` for a text mark is the GEOMETRY box derived from frame dimensions,
not a match position, so "the detector's box did not move after the fill" was predetermined,
and "the mask covers 96.5% of the detected region" compared geometry against the same
geometry. What actually answered it: looking at the pixels, and testing whether `remove()`
changed the array at all.

#### The Jimeng pill: measure the GATED path or the number is meaningless

The parity audit calls `get_mark(key).remove`, which bypasses `_keep_pill`. On the pill
that path reports "detector still fires after removal" 68-75% of the time and reads as a
broken feature. It is not: the product gates the pill hard, and `--mark auto` behaves
completely differently.

Measured through the product path over **all 2,738 pill positives** (2026-07-20):

| | n | |
|---|---|---|
| raw detections | 2,738 | |
| corroborated real (wordmark or TC260) | 346 | 12.6% |
| the gate lets through to removal | 125 | 4.6% |
| **of those, corroborated real** | **125** | **precision 100% (95% CI 97.0-100)** |

Every single pill the product removed was corroborated. Headline recall is 36.1%, but that
denominator is wrong: **TC260 provenance maps to BOTH ByteDance products**, so a "TC260
says Jimeng" image may be a Doubao image with no pill at all. Split by evidence strength:

| corroboration | n | removed | recall |
|---|---|---|---|
| wordmark (names the product) | 49 | 49 | **100% (CI 92.7-100)** |
| TC260 only (cannot separate Doubao) | 297 | 76 | 25.6% (CI 21.0-30.8) |

So where the evidence actually names Jimeng, the pill is removed every time. The flatness
guard suppresses 186 TC260-only cases -- its measured recall cost, paid to avoid the
smeared textured fills it exists to prevent.

Two harness lessons, both of which produced a wrong number before being caught:
- **Gated marks must be measured through the gate.** A per-mark audit answers a question
  the product never asks.
- **Match mark labels exactly.** A substring test on `AI生成` also matches Doubao's label
  (`Doubao 豆包AI生成 text`), counting Doubao removals as pill removals and inflating the
  pill's precision.

#### The finding that was not being looked for: filling a faint mark is net negative

Samsung came out negative in every cell, which looked like a broken mark. It is not -- the
sign is set by how strongly the mark perturbs the image, not by which mark it is. Per-mark
recovery by damage band (a HIGH damage-PSNR means a FAINT mark):

| mark | 0-18 dB | 18-22 dB | 22-26 dB | 26+ dB |
|---|---|---|---|---|
| doubao | +9.78 (n=160) | -1.52 (36) | -3.37 (18) | -8.38 (23) |
| jimeng | +7.70 (187) | -2.49 (21) | -2.54 (15) | -2.34 (15) |
| samsung | - | **+7.23** (48) | -3.78 (102) | -10.78 (90) |

Samsung is POSITIVE where its mark is strong; doubao and jimeng go NEGATIVE where theirs
are faint. Linear fit over all 720: `recovery = -0.861 * damage_psnr + 19.47`, break-even
at **~22.6 dB**. Samsung's alpha map peaks at 0.37 against doubao 0.68 and jimeng 0.93, so
Samsung simply sits on the faint side of that line most of the time.

So: **below ~22.6 dB of mark damage, inpainting costs more fidelity than the mark did.**
The pipeline currently fills unconditionally once a mark is detected, so this cost is
invisible today.

Do not read this as "stop removing faint marks". A user who wants the watermark GONE is not
optimizing PSNR, and a faint mark is still a mark. What it says is that the fill has a real
cost, it is now measurable, and for faint marks it exceeds the thing it removes -- which
makes "how faint is too faint" a product decision that can finally be made on evidence.

### B2. Detector response curves -- RUN 2026-07-20

`scripts/detector_response.py`. Stamp marks across a controlled grid (size x contrast,
crossed with real corpus backgrounds and frame aspects) and measure two things per cell:
`detected`, and whether the same call path then yields a non-empty removal mask
(`maskable`). The second column exists because the gap between them is a silent no-op --
`identify` reports a mark that `visible` skips -- and any detection-only harness scores
that bug as a success.

**Read the nominal cell, never the aggregate.** The grid deliberately visits sizes and
opacities no engine was calibrated for, so its overall rate is an adversarial score, not
recall. Production recall still comes from the unbiased corpus sample.

What it found:

- **The size response is a COMB, not a curve.** `_tophat_score` sweeps exactly three rungs
  (0.8, 1.0, 1.25). Doubao scores ~0.99 at each rung and collapses to 0.37-0.48 between
  them, against a 0.50 gate -- so a mark ~10% off a rung is missed at FULL contrast. The
  `binary` front-end has no ladder at all: jimeng holds one lobe over 0.90-1.20, samsung
  only 0.95-1.05, i.e. samsung requires a mark at essentially its exact nominal size.
- **Contrast is nearly irrelevant** on the tophat front-end -- the response is
  max-normalized, so a mark at 15 luma levels of contrast still scores 0.984.
- **No detected-but-unmaskable cells** at any grid point, so the front-end/mask parity fix
  holds across the whole operating range, not just where the corpus happened to look.

And the follow-up that stopped a bad change: dead zones only cost recall if real marks land
in them. `scripts/ladder_headroom.py` measured that on the corpus (positives = frames whose
metadata names the vendor, negatives = frames with no metadata signal, deliberately NOT
filtered on the detector's own verdict, which would make its false-fire rate 0 by
construction). A 13-rung dense ladder recovers **28 of 368 misses (7.6%)** while false fire
goes **2.52% -> 3.05%**. So the comb is real and mostly unvisited: the fractions were
calibrated on real captures and real marks cluster at the rungs. **Do not densify the
ladder** on this evidence.

The residual looked like a clean lead and, on a full 923-positive / 3546-negative run,
turned out not to be. A targeted `plus_one` ladder (one extra rung at ~1.116) recovers 36
of the 39 marks the 13-rung dense ladder recovers -- so the win really is that one rung,
not density. But **all 36 recoveries AND all 18 of its added false fires are LANDSCAPE at
that same rung** (2.0:1 overall, 1.7:1 even gated to landscape). The rung is not vendor
signal; it is a size shift that helps and hurts equally.

And the obvious "fix it at the source" -- bump the landscape width fraction so that
subpopulation lands on the nominal rung -- is **falsified by the same data**: the 55
currently-DETECTED landscape positives already peak at scale 1.0, not 1.116. So there are
two size clusters of landscape doubao marks, one at the calibrated nominal and one ~11%
larger, and moving the nominal would drop the cluster that works to catch the one that does
not. The larger cluster is genuinely a different size and is inseparable from landscape
false fire at the NCC gate -- the same detector-discrimination wall as vendor attribution,
not a geometry constant anyone forgot to set. **Do not add the rung, and do not move the
landscape fraction.** The per-rung scores are stored in `_ladder_headroom_doubao.jsonl` so
this verdict can be re-derived without re-running the sweep.

Why the misses are not a tuning problem: on the sampled misses the dense-ladder score is
bimodal -- detected marks sit at median 0.938, misses at 0.155, and **the band 0.31-0.52 is
empty**. There is no near-gate cluster, so no threshold recovers them. Eyeballing 26 miss
corners explains it: ~6 are jimeng (TC260 names no vendor, so a "doubao positive" is often
a jimeng frame), ~14 carry no visible mark in that corner at all, 2 belong to **uncovered
vendors** (`千问AI生成`, `百度 AI生成`), and only ~4 are genuine doubao failures -- on
saturated or structurally busy corners, with heterogeneous causes (the saturation gate
explains exactly one of five tested).

### B3. Invisible round-trip, positive-control gated

The open DWT-DCT detector is positive-only and carrier-fragile: "not found" on a fragile
carrier proves nothing (measured: `chatgpt-1.png` recovers 114/128, below the 118 gate).
Every invisible assertion must first embed on the SAME carrier and confirm recovery, and
**degrade to a skip rather than a pass** when the control fails. Already implemented in
`smoke_matrix.py`; apply the same discipline anywhere else this detector is used.

### B4. Resource ceilings

Peak RSS and wall time per backend x input size, up to 25 MP. The memory-constrained CPU
tier is a real constraint (MI-GAN must stay ~0.6-0.9 GB by cropping around the mask); a
regression here is invisible today and would only surface under load.

## Tier C -- human-labelled accuracy (bounded by labelling effort)

The machinery exists: `visible_recall_sample.py` -> `visible_sheets.py` ->
`visible_groundtruth.py` -> `visible_eval.py`.

- **Recall** is the known weak spot: measured once, unbiased n=240, and that single
  measurement is what exposed the landscape miss. Expand per mark, especially jimeng
  (n=14) and jimeng_pill (n=6), whose numbers currently rest on almost nothing.
- **Precision** re-runs over the existing 779-cell ground truth; benchmark every detector
  change with `--vs <snapshot>`.
- **Coverage**, the largest known gap: ~6% of sampled images carry an uncovered vendor's
  mark (百度 / 星绘 / 抖音-class; 千问 was the first of this class and is registered since
  2026-07-21) that no registered detector can fire on. This is a
  coverage problem, not a tuning problem, and no threshold work will move it.

Three harness rules are load-bearing and must not be relaxed: score a mark only within its
crop's adjudication scope; take provenance from metadata, never from labels; and never
report recall from the detector-sampled set.

## Tier D -- external oracles (manual, not automatable here)

SynthID removal cannot be verified locally by design -- no public decoder exists. Each
vendor has its own oracle and it covers only that vendor's content: `openai.com/verify` for
OpenAI (more accessible, the automation candidate), the Gemini app for Google (manual,
rate-limited). A quiet metadata proxy is **not** proof the pixel watermark is gone.

Scope honestly: this tier certifies strength floors on a handful of images per vendor, and
that is all it can do. See `docs/synthid.md`.

### D1. Sampling frame

The sidecars already classify the corpus: **15,000 images whose watermark list mentions
SynthID**, of which 9,071 carry `verify_oracle=openai` and 5,929 `verify_oracle=google`.
Stratify on the two axes that actually move removal efficacy -- vendor (the certified
floors differ: OpenAI 0.10, Gemini 0.15) and content class (photoreal vs flat graphic,
where the pipelines are documented to diverge).

### D2. The oracle is the bottleneck, not the GPU

Measured on MPS (2026-07-19), single invocation of `invisible`:

| `--max-resolution` | 256 | 384 | 512 | 768 | 1024 |
|---|---|---|---|---|---|
| wall time | 37.9 s | 58.9 s | 59.4 s | 65.2 s | 118.3 s |

384/512/768 are indistinguishable, so below ~768 the cost is dominated by **fixed model
load, not diffusion**. Confirmed by batching: 4 images in one process took 105.5 s
(26.4 s/image) against 59.4 s/image one at a time -- roughly 40 s fixed overhead per
invocation and ~15 s marginal per image at 512 (rough: run-to-run variance is large).

Two consequences for the harness:
- **Amortize the fixed cost**: one long-lived process over many images, never one
  invocation per image. That is a 2-4x win. Shrinking below 512 is not.
- The binding constraint is the **external oracle's throughput**, which is manual and rate
  limited. So do not run a uniform grid; spend each oracle check where the answer is
  uncertain -- **bisect strength per content class** to certify a floor in ~10 checks
  instead of ~100.

### D3. Mandatory control before trusting any reduced-size run

`--max-resolution` is downscale -> diffuse -> Lanczos upscale. If the **resize round trip
alone** damages SynthID, the oracle goes quiet for a reason unrelated to removal and the
result does not transfer to production at native resolution.

Before any reduced-size sweep, run the resize round trip with **no diffusion** and put the
result through the vendor oracle. If the watermark survives, the reduced size is a valid
test bed; if it does not, reduced-size results are measuring the resizer. This is the same
failure shape as the imwatermark carrier-fragility trap in B3: an oracle that falls silent
for the wrong reason reads exactly like success.

`docs/synthid.md` cites ~99.98% TPR across 30 transforms including resize, which predicts
the control passes -- but that is Google's claim about their own decoder, not our
measurement, so it is a hypothesis to test, not a reason to skip the control.

## Tier E -- robustness and adversarial inputs

Malformed and hostile inputs, since ~0.2% of real uploads are already truncated: truncation
at many offsets, corrupt headers, 16-bit and CMYK, absurd dimensions, decompression bombs,
zero-byte files, unicode and RTL filenames, symlinks, read-only output dirs, concurrent runs
on one file. The bar is never "handles it" but **never raises and never silently degrades**.

## Build order

1. **A1 sidecar regression** -- highest value per hour, unattended, needs no new labels.
2. **A2/A3/A4 parity and invariants** -- full corpus, reuses existing audit scripts.
3. **B1 fill quality** -- closes the oldest unmeasured claim in the project.
4. **B2 detector curves** -- cheap, and directly guards the geometry class of bug.
5. **A5 contract sweep at corpus scale**.
6. **B4 resource ceilings**, **E robustness**.
7. **C recall expansion** -- gated by labelling appetite.
8. **D oracles** -- manual, per release.

Every tier writes a versioned snapshot so runs are comparable over time; a run that cannot
be diffed against the last one is a one-off, not a regression suite.

## What the measurements imply for detection work

Recorded here because each item is grounded in a number from this campaign, not because
it is a prioritized plan (that lives elsewhere -- see the note at the end of this section).

### Metadata absence does not disable the detectors -- it disables the RELAXATION

The detectors are pixel-based and need no metadata. What metadata does is relax the
false-positive gate (`auto` vs `strict`). So "work better without metadata" means
strengthening the strict-path detectors themselves; it is not a gating problem.

Per mark, what actually goes away when metadata is stripped:

- **The pill loses an entire arm.** Its TC260 arm is dead without metadata, leaving only
  the wordmark arm. Measured: where the wordmark corroborates, pill recall is **100%**
  (49/49); on TC260-only evidence it is 25.6%. So on stripped uploads the pill's fate
  rests entirely on Jimeng wordmark detection -- whose own recall is **71% on n=14**.
  This is the weakest link with the most leverage: every point of wordmark recall pulls
  the pill along with it.
- **The sparkle already runs on pixels**, and the FP-gate tightening cost ~120-200 genuine
  detections corpus-wide (12.5% of the 1,256 lost). That headroom exists but the
  precision trade behind it was deliberate.
- **The largest gap is metadata-independent by nature**: ~6% of sampled images carry an
  uncovered vendor's mark (百度 / 星绘 / 抖音-class; 千问 is registered since 2026-07-21)
  that no registered detector can fire on at all.

### Where the evidence points

1. ~~A generic CJK AI-mark detector~~ **superseded for 千问 (2026-07-21).** The AUC ~0.5
   non-separability from Doubao was measured at the WRONG size; at the fitted geometry an
   exact-size 6-glyph template separates 千问 from 400 doubao-marked frames with ZERO
   cross-fire at the gate, so per-vendor registration won and shipped. The generic
   class-detector shape may still be right for the long tail of compliant vendors, but
   its motivating measurement is gone.
2. **Port the `tophat` front-end to the remaining marks.** It took Doubao from 89% to 92%
   recall at unchanged 99% precision. But the gate is front-end specific and **must be
   recalibrated, never ported**: a naive 0.40 produced 8 false fires instead of 1 and
   silently halved the pill's recall (because `_keep_pill` suppresses the pill whenever
   Doubao fires).
3. **The Jimeng wordmark.** Weak on its own (71%/71%) and it gates the pill. Its
   silhouette is also non-discriminative against Doubao's, which was patched with a 0.85
   threshold -- a patch on a detector problem, not a fix.

### Measure before improving

Jimeng recall rests on **n=14** and the pill's on **n=6**. Improving what is measured by
six samples means not knowing whether it improved. Tier B2 (detector response curves on
stamped marks) is the instrument to build first: recall as a function of size, contrast
and background texture, with no new hand labelling, and it catches the geometry class of
bug (`scale_basis`) directly.

This section records what the measurements imply technically. Prioritization is tracked
separately, outside this repo.

## Real-example end-to-end run (2026-07-20)

The corpus sweeps all call the library IN-PROCESS. This run drives the actual
`remove-ai-watermarks` entry point as a user would, over real corpus images and the
committed real fixtures, and checks the OUTCOME rather than the exit code
(`scripts/real_examples_e2e.py`). **26 of 28 behaviours passed.**

| Command | Real examples | Result |
|---|---|---|
| `identify --json` | 6 provenance classes (OpenAI, Adobe, Midjourney, Doubao/TC260, Grok, FLUX) | 6/6 correct verdict + platform |
| `metadata --check` / `--remove` | the same 5 real AI-metadata files | 10/10 detected, and the output re-scans clean |
| `visible --mark auto` | a live positive per mark | doubao / jimeng / gemini removed and re-detect clean; pill correctly DECLINED by the gate; samsung partial (below) |
| `erase --region` | one real image x 3 backends | cv2 / MI-GAN / big-LaMa all wrote output |
| `batch --mode visible` | a real 5-image directory | 5/5 outputs |
| `invisible` (MPS, `--max-resolution 512`) | a real Gemini and a real OpenAI carrier | both wrote a genuinely CHANGED image |
| `all` (MPS) | a real Gemini carrier | one transient exit 1, not reproducible (see below) |

**Samsung is a real, reproducible partial.** It is the faintest registered mark (peak alpha
~0.38) sitting on a 0.40 gate, so the margin between "detected" and "removed" is razor thin.
Over the entire corpus population (n=3, all of it) the CLI clears 2 of 3 outright
(0.446 and 0.440 -> below gate) and on the weakest one reduces 0.431 -> **0.404**, which is
still fractionally over the gate. The glyph IS filled and the confidence IS reduced; the
residual re-detects. This is the faint-mark residual class on the `binary` front-end, which
has no equivalent of the `tophat` faint-mask fallback. **Not fixed:** with n=3 corpus-wide
there is no way to tell an improvement from noise, which is the same
improving-what-3-samples-measure trap recorded under Open items.

**The pill "failure" was the harness, not the product.** `detect_marks` fires `jimeng_pill`
but `remove_auto_marks` returns no label -- `_keep_pill` correctly declines an uncorroborated
low-confidence pill, so `visible` writes nothing and exits 2. The harness now asks the
product what it DECIDED and treats a correct decline as a pass.

**The `all` exit 1 did not reproduce** -- the same invocation ran clean standalone twice and
again in the exact three-run sequence that produced it (all exit 0, step 2 executed, no skip
banner). It stays recorded as a transient because it could not be diagnosed: the harness
discarded the command output, and `cmd_all` has three distinct `SystemExit(1)` paths
(unreadable input, unreadable intermediate, and the deliberate synthid-skipped banner) that
an exit code alone cannot distinguish. The harness now retains the output tail on any
failure, so a recurrence is identifiable.

## Tier E: robustness (2026-07-20) -- RUN, and it found two real crashes

`scripts/robustness_suite.py` drives the real CLI over adversarial and degenerate inputs and
scores GRACEFULNESS, not success: a non-zero exit with a readable message is a pass, an
unhandled traceback or a hang is a fail. **33/33 graceful after two fixes; 31/33 before.**

Covered: truncated and corrupt files, zero-byte, a text file named `.jpg`, 1x1 and
1x4000 slivers, a 729 KB decompression bomb declaring 16000x16000, Unicode and RTL
filenames (read AND write), a mismatched extension, a nonexistent nested output dir, a
read-only output dir, a directory passed as a file, a missing path, 4x concurrent runs
against one input, and batch over both an empty and an undecodable directory.

**Both defects were invisible to the 849-test suite**, because unit tests feed well-formed
fixtures into writable directories. Both are now regression-guarded by
`tests/test_cli_robustness.py`.

1. **A failed write crashed on the size report.** `image_io.imwrite` is contractually
   non-raising -- it returns `False` when the path cannot be written. But
   `write_bgr_with_alpha` discarded that bool and returned `None`, so **no caller could
   distinguish a failed write from a successful one**, and all five write sites then ran
   `output.stat()` to print the size. A read-only output directory produced a bare
   `FileNotFoundError` traceback pointing at the stat rather than at the write. The signal
   existed the whole way down and was thrown away by one wrapper. Fixed at that wrapper
   (propagate the flag) plus one shared `cli._write_output_or_exit`, so all sites are
   covered rather than the one that happened to be caught.
2. **A directory passed as the image crashed the scanner.** `click.Path(exists=True)`
   accepts directories unless told otherwise, so `identify <dir>` reached `open()` and
   raised `IsADirectoryError`. Fixed by `dir_okay=False` on all six `source` arguments --
   argument parsing now refuses it, which is where it belongs. (`batch` was already correct:
   it declares `file_okay=False`.)

3. **The worst instance was found by the /simplify review, not by the sweep: `batch` into a
   read-only directory wrote ZERO files for 2 inputs and exited 0.** No traceback, no
   error, a success code and an empty output directory -- a wrapping service would treat
   that as a completed run. `graceful()` structurally cannot see this class, because it
   scores exit code and traceback markers and this failure has neither. The suite now
   carries a `batch_readonly_outdir` case that asserts on the ARTIFACTS (how many files
   exist) rather than on the status. **Any check for a silent no-op has to assert on the
   output, not the exit code.**

The fix is NOT uniform, and that matters: the single-image commands exit via
`cli._write_output_or_exit`; `api._write_visible_result` RAISES so a library caller gets an
accurate error rather than a confusing `FileNotFoundError` from the downstream metadata
strip; and the batch sites raise too but must never `SystemExit`, because the batch loop
counts per-image exceptions and aborting would kill the whole run instead of failing one
image.

The lesson worth keeping: **a non-raising IO contract needs its return value checked at
every call site, and a wrapper that swallows it silently disables the contract for
everyone downstream.** Grep for other `-> None` wrappers over non-raising primitives --
`invisible_engine.py:346` still discards `imwrite`'s flag and is the same shape.

## Tier B4: resource ceilings (2026-07-20) -- RUN

`scripts/resource_ceilings.py`, one FRESH process per cell (peak RSS inside a long-lived
process is contaminated by whatever ran before it, and the question is what a per-request
worker needs). The mask is a fixed, small corner region at every input size, so the thing
under test is whether the learned backends really crop around the MARK rather than
processing the frame.

| backend | peak RSS 1MP -> 25MP | wall | scaling |
|---|---|---|---|
| cv2 | 74 -> 440 MB | 0.02-0.12 s | **grows 5.9x with input size** |
| migan | 603 -> 775 MB | ~0.6 s | flat |
| lama | 4679 -> 4779 MB | ~3.8 s | flat |

**The documented claims hold.** CLAUDE.md's "migan ~0.6-0.9 GB regardless of upload size"
and "lama ~4.7 GB peak" both reproduce to the digit, and the crop-around-the-mask design is
confirmed: both learned backends are flat in input size. This is also the measurement
behind the deployment split (free tier `migan` fits a small droplet under 1 GB; paid tier
`lama` needs ~5 GB).

**New information: cv2 is the only backend that scales with the input.** It inpaints the
full frame rather than a crop, so at 25 MP it costs 440 MB -- still the cheapest tier, but
6x its small-image footprint, which is worth knowing when sizing a worker that accepts
phone-camera uploads. Its wall time stays trivial throughout.

Caveat on the wall times: each is a COLD process including model load, so migan's ~0.6 s
here is not comparable to the ~0.19 s warm figure quoted elsewhere. Cold is the honest
number for a per-request worker; warm is the honest number for a long-lived one.

**Two method notes, both cases of the harness corrupting its own measurement.**

1. The first version called `erase()` with a positional `boxes` argument, which the
   keyword-only signature rejects. All 12 cells failed identically and still reported
   plausible-looking RSS (60-129 MB) -- the process's numpy footprint with no backend work
   at all. Twelve identical failures were one bad call, and only the printed error made
   that visible. The child now asserts the output actually differs, so a no-op cannot be
   reported as a measurement.
2. That assertion was then itself the contaminant: `(out != img).any()` allocates a boolean
   temp the size of the IMAGE before `getrusage` is read -- ~75 MB at 25 MP, and it grows
   with the input, so the harness partly manufactured the very "cv2 scales with input size"
   conclusion it was measuring. Now it compares only the mask box. **Re-measured: the
   conclusion survived, the digits moved.** cv2 6.1x -> 5.9x, migan max 847 -> 775 MB, lama
   max 4849 -> 4779 MB; the table above is the corrected run. Worth keeping because the
   contaminated numbers were already committed to CLAUDE.md before the check was
   questioned -- a verification step is part of the instrument and needs the same scrutiny
   as the thing it verifies.

## Open items (as of 2026-07-20)

Everything below is known, measured, and deliberately not done yet. Each line says what it
would take, so none of it has to be rediscovered.

### START HERE next session

Item 1 from the previous session is **DONE (2026-07-21)**: 千问 is registered --
see "The 千问 harvest (2026-07-21)" below for the decision and the numbers. In
priority order:

1. **Decide the exit-code split** (open defect 2). It is a deliberate product call, not
   research: it is breaking for existing wrappers, so it needs a yes/no rather than more
   measurement.
2. **The two small correctness items** (open defects 3 and 4) -- both are contained, both
   have the fix written out below.
3. **The bonus vendors from the harvest** (元宝 n=50, 可灵 n=30, cat-logo n=19) repeat
   the 千问 playbook each: font-rendered silhouette, `--fit-geometry`, gate calibration
   against the contamination-guarded clean arm, crossfire against doubao/jimeng. 可灵
   additionally stamps a second mark bottom-LEFT, which no current text-mark config
   expresses (the pill is top-left; a bottom-left CJK mark needs a `corner="bl"` CJK
   config -- samsung is `bl` but Latin-script and width-based). 星绘/百度 are NOT in the
   corpus in labelable quantity -- verified, do not hunt them again.

Do NOT restart the sweeps to "check". Their artifacts are on disk and listed under
"Completed full runs" below; re-running costs hours and answers nothing new. The fast way
to confirm the whole surface still works after a change is
`uv run python scripts/real_examples_e2e.py` (~2 min, real corpus examples through the real
CLI) plus `uv run python scripts/robustness_suite.py` (~3 min, adversarial inputs).

### The 千问 harvest (2026-07-21) -- RESOLVED, registered the same day

**The unlock: the TC260 label is not anonymous.** Its `ContentProducer` field carries the
producer's Chinese Unified Social Credit Code (`001191110102MACQD9K64010000` -> USCC
`91110102MACQD9K640`), which names a legal entity. So carriers partition into per-VENDOR
cohorts from METADATA ALONE, owing nothing to any pixel detector -- which is exactly what
broke the previous attempt, whose only way to find 千问 frames was to eyeball the misses of
a detector that cannot see them. A cohort is a LABEL: eyeball one frame, and every frame in
it is a labelled example. (CLAUDE.md's "the generic TC260 label names no specific vendor" is
about the label MARKER; the producer FIELD inside the block is a different thing.)

New tools, both lint-clean, both **uncommitted**:
- `scripts/vendor_cohort_harvest.py` -- full-corpus metadata scan -> cohorts. Joins which
  detectors fired from the completed `_visible_positives.jsonl` rather than re-running the
  pixel pass. Artifact `data/spaces/_vendor_cohorts.jsonl` (**4441 carriers, 46 entities**).
  `--sheets N` writes full-width top/bottom band crops per cohort for eyeballing.
- `scripts/vendor_mark_calibrate.py` -- scores a cohort against the 432 hand-labelled
  `present: []` negatives from the 2026-07-18 round, and writes score-SORTED corner crops so
  mark presence and the gate are read in one visual pass.
- `src/.../assets/qwen_alpha.png` + `xinghui_alpha.png` regenerated (they were listed in
  `render_vendor_silhouettes.py` but had never been committed).

**What the corpus actually contains** (this corrects the previous list of targets):

| Cohort USCC | n | quiet | Visible mark | Verdict |
|---|---|---|---|---|
| 91440101MA9Y9T4H7A | 117 | 112 | `千问AI生成` bottom-right | **the target, confirmed by eye** |
| 91340100MAEB4N8H76 | 73 | 70 | mostly none; one `RunningHub AI生成` | metadata-mostly |
| 913502007378955153 | 113 | 109 | none seen | metadata-only |
| 91440300708461136T | 50 | 46 | `元宝AI生成` (Tencent Yuanbao) | bonus vendor, bold |
| 91441900557262083U | 49 | 45 | none seen | metadata-only |
| 91110108335469089C | 30 | 28 | `可灵AI 3.0` (Kling) + an `AI生成` pill bottom-LEFT | bonus vendor |
| 91110108562144110X | 19 | 19 | cat-logo + `AI生成`, in **19/19** | bonus vendor, very clean |

**`百度` and the `星绘`/`抖音` class are NOT in this corpus in labelable quantity.** No brand
token for them appears in any AIGC label field, and every remaining cohort is <= 16 frames.
Do not spend another session hunting them here; the previous "one confirmed positive each"
is all there is. The corpus offers 千问 richly plus three DIFFERENT vendors instead.

Also worth knowing: a cohort is a strong grouping key but names the SIGNING ENTITY, not
always the consumer brand -- the 千问 cohort contains one `造点AI生成` frame. And TC260
provenance does NOT imply a visible mark, which is why four large cohorts above are
metadata-only. The cohort is the candidate pool; the eye settles mark presence.

**RESOLVED 2026-07-21: 千问 is registered** (`qwen_engine.py`, strict-only, no rival
margin). The ladder trade-off that was the open decision is settled in favor of a
**per-mark ladder**, not the shared one and not a wider shared one -- and the path there
found two more geometry defects the "single fraction on the shipped ladder" framing had
missed.

What the final calibration measured, in the order it happened:

1. **The locate box was clipping the mark.** Scoring with the fitted fractions still
   collapsed the cohort (p50 0.209). Frame-level diff against the wide-ladder fit showed
   why: the real mark sits ~0.025 of the short side off the right edge, while doubao's
   inherited box anchors at 0.004 -- the box's left edge cut into the 千 glyph, and an
   exact-size template scored 0.26 where the fit's wider box scored 0.73. So the locate
   fractions are as mark-specific as the template size; `--fit-geometry` now records the
   absolute match rects and fits the box too (margins ~0.021, width 0.231, height 0.074).
2. **The size distribution is cleanly BIMODAL.** With the box fixed, frac_short clusters
   at ~0.124 (13 frames) and ~0.203 (38 frames), nothing between -- two stamp sizes,
   ratio 1.64, just over the shared ladder's 1.5625 span. That is why no single fraction
   covers the mark.
3. **A per-mark 2-rung ladder beats both alternatives.** Candidates, both arms scored on
   the shipped code path: (A) shipped 3-rung @ frac 0.167 -- cohort p50 0.538; (B1)
   2-rung (0.78, 1.27) @ frac 0.160, one rung centred on each mode -- cohort p50
   **0.662**; (B2) 4-rung (0.8, 1.0, 1.25, 1.5625) @ frac 0.155 -- p50 0.449, strictly
   worse (its big-mode rung sits 4.6% off the mode, and the extra rungs cover nothing).
   B1 also costs one matchTemplate LESS than the shipped 3. `TextMarkConfig.ladder` was
   added with the default `(0.8, 1.0, 1.25)`, so every other mark's computation is
   byte-identical (the full 876-test suite plus e2e + robustness confirm); a shared
   densification was already ruled out by B2's false-fire measurement. The doubao
   false-fire check the plan asked for reduces to that equivalence-by-construction --
   doubao's ladder never changed.
4. **`alpha_height_frac` measured, not inherited:** aspect fit at the winning width, p50
   0.260 (tight, p10-p90 0.250-0.270) -> 0.0416. The silhouette's own aspect (0.2219)
   and doubao's ratio were both measurably off.
5. **The clean arm was contaminated, and fixing it flipped the verdict.** The 2026-07-18
   `present: []` labels are in the vocabulary of the REGISTERED marks only -- 146 of the
   432 "clean" frames sit in a TC260 cohort, including 15 qwen-cohort frames VISIBLY
   carrying 千问AI生成, and they were the clean arm's entire top tail (clean p99 0.37 ->
   0.69 with the fitted geometry). `load_sets` now drops every frame in ANY cohort.
   Final arm: 286 frames, clean p99 0.301 / max 0.316.
6. **Gate 0.45, strict-only, no rival margin.** Every cohort frame >= 0.45 carries a
   visible mark (83 of ~96 eyeballed visible marks fire = **86% recall of visible
   marks**; the misses are white-on-near-white contrast losses); 0/286 clean fires;
   crossfire at the gate: 0/400 on doubao-marked frames, 0/298 on jimeng-marked frames
   (the shared `AI生成` tail correlates at p50 0.224, far below gate -- the AUC-0.5
   attribution fear from 2026-07-18 was a mis-sizing artifact). A 0.10 rival margin
   would have cost ~10% of genuine qwen detections, so `rivals=()`. The band just below
   the gate is dominated by non-qwen banners (夸克 anti-forgery strip 0.274, 造点 mark
   0.253), so a provenance-relaxed arm would be mostly false fills:
   `provenance_ncc_factor` is pinned at 1.0 and qwen has NO provenance mapping.
7. **Parity confirmed end to end:** detect -> cv2 fill -> re-detect clean on **83/83**
   real cohort marks, no empty masks; `real_examples_e2e.py` now drives a live qwen
   positive through the real CLI (qwen bucket = symlinks under the gitignored
   `_visible_datasets/`); a confident qwen detection suppresses the jimeng pill exactly
   like doubao's does.

Method note worth keeping: **the wide ladder flattered the clean arm exactly as
predicted, but the trap that actually bit was the LABEL vocabulary.** "present: []"
meant "no registered mark", not "no mark" -- and a calibration clean arm has to be
re-filtered per candidate, or the gate is read off frames that carry the very mark being
calibrated.

The bonus vendors (元宝, 可灵, cat-logo) need their own font-rendered silhouettes before
any of this repeats for them; 可灵 additionally puts a second mark bottom-LEFT, which no
current text-mark config expresses.

### Open defects

| # | Defect | Measured impact | What the fix takes |
|---|---|---|---|
| 1 | 16-bit PNGs are downconverted to 8-bit by a metadata strip | 42 of 27,018 corpus PNGs (0.16%); one went 9.2 MB -> 2.5 MB | a byte-level PNG chunk stripper, so the PIL re-save is skipped entirely |
| 2 | Exit code 2 means three different things (no visible mark / no invisible signal / Click usage error) | any wrapper must parse stderr to tell them apart | split the codes; **breaking for existing wrappers**, so it needs a deliberate call |
| 3 | `visible_removal_audit.py` measures the UNGATED per-mark path | reports the pill at 32% where the product runs at 100% precision | teach it the product path (`remove_auto_marks`) for gated marks, or at minimum say so loudly in its docstring |
| 4 | `invisible_engine.py:346` still discards `imwrite`'s success flag | same shape as the crash fixed 2026-07-20, on the diffusion output path; not yet observed failing | check the bool and raise/report, mirroring `cli._write_output_or_exit` |
| 5 | samsung leaves a residual just over its gate on the weakest of its 3 corpus positives | 0.431 -> 0.404 against a 0.40 gate; the other two clear outright | a faint-mask fallback for the `binary` front-end (samsung has none), but **do not tune on n=3** -- blocked behind item 1 above |

### Closed 2026-07-20 (kept for the reasoning, not for action)

- **The visible-parity re-run** confirmed the doubao front-end fix: **91.8% -> 99.3%**
  (2562/2580), gemini/jimeng/samsung unchanged. Artifact `_visible_parity_cv2_v2.csv`.
- **The faint-mask fallback filled the whole corner box** instead of the glyph. Fixed with
  the detector's own best-match box; full reasoning immediately below, because how it
  escaped both parity and its own regression test is the instructive part.
- **Three crash-class defects** found by Tier E and the /simplify review (failed writes
  crashing on the size report, batch losing data silently, directories crashing the
  scanner) -- see the Tier E section.

**The faint-mask defect in full, because it is instructive.** The fallback added on
2026-07-19 reads
`np.where(resp >= _FAINT_GLYPH_LEVEL)` with the constant at `0.5`, and its comment says
"thresholded relative to its own peak". But `tophat_response` returns **uint8 0..255**, so
`>= 0.5` selects every pixel with value >= 1 -- the entire non-zero response, not half the
peak. Measured against alternatives on 14 real frames (cv2 fill, detector re-run after):

| mask | detector clean after | median filled area (% of corner box) |
|---|---|---|
| `thr0.5` (as shipped) | 100% | 120.9% |
| `thr190` | 100% | 68.5% |
| largest connected component at 190 | **21%** | 10.5% |
| **the detector's own best-match box** | 100% | **58.7%** |

The fix is the last row: the correlation already located the mark at a position and scale,
so thresholding its response was always a weaker proxy for information we had. The
connected-component variant is rejected outright -- it is the tightest but removes the mark
on only 21% of frames, i.e. it does not cover it.

Two things about how this was found are worth keeping:

- **Parity could not see it.** Parity asks whether the detector is clean after removal, and
  a mask that fills everything passes trivially. The defect was in the COST, and nothing
  measured cost on that path. A green parity run is not evidence about mask size.
- **The regression test could not see it either, by construction.** Its fixture is a FLAT
  frame, where the top-hat response is non-zero only on the glyph, so every threshold gives
  the same bounding box. Mutating the constant to an absurd 99.0 left it green. The fixture
  now carries texture, which is the condition under which sizing matters and what real
  corner backgrounds look like -- and it reproduces the corpus number exactly (127% of the
  corner box, against 120.9% measured).

### Latent, pre-existing, not fixed this pass

The detect-fires / mask-empty silent no-op that fix 4 closed on the `tophat` front-end has a
**narrower cousin on the `binary` front-end** (jimeng/samsung), surfaced by the /simplify
altitude review. Binary detection gates on `coverage >= detect_min_coverage` (a FRACTION)
while the mask gates on `xs.size >= _MIN_GLYPH_PIXELS = 20` (an absolute COUNT), both on the
same blob. For samsung (`detect_min_coverage = 0.01`) they disagree in a small-image band
(width ~200-294 px): detection can fire at 10-19 glyph px while the 20-px mask floor returns
None -> the identical observable. It is NOT introduced by this work (the `else: return None`
fall-through predates it), it sits well below real mark sizes (captured positives are
1086-2048 px wide), and fixing it means changing binary detection's gate to match the mask's
-- which needs its own per-mark measurement. So the "cannot drift by construction" claim is
scoped to `tophat` (where score and box are one computation); the binary path is coupled but
by a threshold pair that can still disagree at the edges. Fix only alongside a binary-mark
detector change, never on its own.

### Dependency alert

`GHSA-rrmf-rvhw-rf47` (torch, `torch.jit.script` memory corruption) is open again and
**the reason it was dismissed no longer holds**. It was dismissed `not_used` on 2026-06-10
partly because no patched version existed; a patched **torch 2.13.0** now does, and the
current alert range is `<= 2.12.1`. `docs/release-and-distribution.md` still says "no
patched torch version exists -- do not re-triage it", which is now stale. Either bump torch
(it is transitive from the optional `gpu` extra) or re-dismiss on the remaining grounds
(the codebase never calls `torch.jit`, grep-verified) and correct that note.

### Where detection work should go next

This is the EVIDENCE behind "START HERE" item 1, not a competing list -- it records what was
measured and ruled out, so the next session does not re-run any of it. In the order the
evidence supports:

1. **Not the ladder, not the threshold, not the landscape rung.** All three were measured
   to completion and all three are dead ends: the dense ladder buys 7.6% of misses for a
   21% relative rise in false fire; the score band below the gate is empty so no threshold
   recovers the misses; and the one targeted rung that helps (1.116, landscape) adds false
   fire at 1.7:1 because the recoveries and the false fires are the same landscape size
   shift. Moving the landscape width fraction is also out -- detected landscape marks
   already sit at the nominal, so it would break more than it fixes. Do not spend here.
2. **Coverage of uncovered vendors is the largest lever.** 千问 was the head of this item
   and is now CLOSED (registered 2026-07-21, see the harvest section above): the blocker
   turned out to be evidence, and the TC260 producer-USCC cohort trick removed it. The
   remaining named vendors are 元宝 (n=50), 可灵 (n=30) and cat-logo (n=19) -- each needs
   a font-rendered silhouette, then the same calibrate-and-crossfire chain. `百度` and the
   星绘/抖音 class are NOT in the corpus in labelable quantity (verified twice; do not
   hunt them again). Nothing may be registered off a single frame.
3. **A generic shared-tail template is not a shortcut.** `AI生成` is guaranteed across
   compliant vendors by GB 45438-2025, so one template covering all of them is the obvious
   idea -- and measured on the tophat front-end it separates a bold 千问 positive from clean
   corners by only 0.407 vs a clean p99 of 0.298. A 4-glyph run is simply less specific
   than a 6-glyph one. Treat it as a harvesting aid, not a detector.

### Completed full runs -- do not re-run to "check"

Each of these took from tens of minutes to hours and its artifact is on disk (gitignored,
under `data/spaces/`). A later session asking "did we actually cover X" should read the
artifact, not relaunch the sweep. Row counts are what the file held when written.

| Run | What it covered | Rows | Artifact |
|---|---|---|---|
| A1 sidecar regression | the whole corpus re-run through `identify`, diffed against recorded verdicts | 39,314 | `_sidecar_regression.jsonl` |
| A2 visible parity (v1 pre-fix, v2 post-fix) | detect -> remove -> re-detect per mark | 10,594 x2 | `_visible_parity_cv2{,_v2}.csv` |
| metadata removal audit | strip-and-verify, detection/removal parity | 21,654 | `_metadata_removal_audit.csv` |
| B1 fill quality | stamped ground truth, per backend and background | 1,200 | `_fill_quality.jsonl` |
| pill gate audit | the pill through the PRODUCT path, not the ungated one | 2,738 | `_pill_gate_audit.jsonl` |
| B2 / ladder headroom | the scale-ladder and threshold questions, to exhaustion | 4,469 | `_ladder_headroom_doubao.jsonl` |
| E robustness | 34 adversarial CLI cases | -- | rerun is ~3 min, no artifact needed |
| B4 resource ceilings | peak RSS per backend, 1 MP -> 25 MP | 12 cells | rerun is ~2 min |
| real-example E2E | every command over real corpus examples | 28 checks | rerun is ~2 min (+ diffusion) |

### Verification tiers not run

- **C recall expansion** -- gated by labelling appetite. This is now the ONLY unrun tier,
  and it is blocked on data rather than effort: every remaining detector question needs
  labelled positives (30+ per uncovered vendor; jimeng still rests on n=14, the pill on
  n=6, samsung on n=3).

(Tiers E and B4 are now RUN -- see the sections above.)

## Standing gap

None of this is in `maintain.sh`, and it should not all be -- the sweeps take hours. But
that means **no detector-accuracy or CLI-contract regression is caught automatically**
today. The endpoint of this plan is a cheap subset (fixtures-only smoke + a sidecar diff on
a fixed 500-image slice) that CI can run, with the full sweeps staying pre-release.