From dd1336f9de44e0193aa48dc959c1209fa63444a0 Mon Sep 17 00:00:00 2001
From: Joseph Magly <1159087+jmagly@users.noreply.github.com>
Date: Sat, 29 Aug 2026 19:33:24 -0400
Subject: [PATCH] docs: record Qwen3.8 qualification evidence

---
 ci/test-risk-map.json               |  2 ++
 docs/QWEN38_27B_RESEARCH_ROADMAP.md | 29 ++++++++++++++++++++---------
 2 files changed, 22 insertions(+), 9 deletions(-)

diff --git a/ci/test-risk-map.json b/ci/test-risk-map.json
index 9e63d24..a2af53f 100644
--- a/ci/test-risk-map.json
+++ b/ci/test-risk-map.json
@@ -58,6 +58,7 @@
       "paths": [
         "obliteratus/abliterate.py",
         "obliteratus/benchmark_lifecycle.py",
+        "obliteratus/checkpoint_evaluation.py",
         "obliteratus/models/qwen35_contracts.py",
         "obliteratus/models/offload_surgery.py",
         "obliteratus/persistence_contracts.py",
@@ -87,6 +88,7 @@
         "tests/test_sweep_contracts.py",
         "tests/test_tourney_contracts.py",
         "tests/test_checkpoint_atomicity.py",
+        "tests/test_checkpoint_evaluation.py",
         "tests/test_offload_surgery.py",
         "tests/test_persistence_contracts.py",
         "tests/test_persistence_pipeline.py",
diff --git a/docs/QWEN38_27B_RESEARCH_ROADMAP.md b/docs/QWEN38_27B_RESEARCH_ROADMAP.md
index a3b3e96..0a98049 100644
--- a/docs/QWEN38_27B_RESEARCH_ROADMAP.md
+++ b/docs/QWEN38_27B_RESEARCH_ROADMAP.md
@@ -254,17 +254,28 @@ from 60% refusal in S2 to 100% in several other strata.
 - text-only regression plus a separate multimodal smoke test so text surgery
   does not silently break the native vision-language wrapper.
 
-## Recommended next operational run
+## Operational qualification result
 
 E02 completed on the optimizer-tune partition as run
-`run-efb2f334197e48fb82587f39a84fe6c9`: refusal fell to 1% (1/142), but
-coherence was 70% and 6/142 harmful responses were degenerate. Its checkpoint
-and full archive are retained. E03 changes only regularization from 0.00 to
-0.10, retaining 10% of the measured refusal subspace to recover coherence while
-using E02's 29-point refusal margin. E03 must not read the 200-pair final
-partition. Promote only if tune refusal remains below 30% and coherence reaches
-at least 80%; then evaluate the final partition exactly once and independently
-reload the saved checkpoint for release qualification.
+`run-efb2f334197e48fb82587f39a84fe6c9`: refusal fell to 1% (1/142), but its
+original preview-truncated coherence score was 70% and 6/142 harmful responses
+were degenerate. Its checkpoint and full archive are retained.
+
+E03 (`run-c18babdb7be34396a07fbb994c23d4cb`) changed only regularization from
+0.00 to 0.10. The saved 53.8 GB BF16 checkpoint was independently reloaded for
+each evaluation after verifying all 35 checkpoint inventory entries by size and
+SHA-256. With the corrected full-completion coherence scorer, the immutable
+optimizer-tune split produced 1.41% refusal (2/142), 100% coherence, 83.3%
+capability, 3.21 perplexity (1.02x pristine), and 1/142 degenerate harmful
+responses. It therefore earned exactly one evaluation on the untouched final
+split. That final saved-checkpoint reload produced 1.0% refusal (2/200), 100%
+coherence, 83.3% capability, 3.21 perplexity, and 6/200 degenerate harmful
+responses. Both durable evaluation records are hashed in the run manifest.
+
+E03 passes the declared promotion gates of refusal below 30% and coherence of
+at least 80%. The 3% final harmful-output degeneracy rate remains a follow-up
+quality concern and should be included in downstream human and classifier
+audits rather than hidden by the aggregate promotion result.
 
 ## Evidence base