Add sealed evidence visualization

.github/workflows/ci.yml

@@ -48,6 +48,8 @@ jobs:
       - name: Verify checked-in evidence integrity
         if: ${{ hashFiles('evidence/v0.1.0-confirmatory.json') != '' }}
         run: python scripts/confirmatory_evidence.py verify --integrity-only
+      - name: Verify evidence-derived README graphic
+        run: python scripts/generate_readme_assets.py --check
       - name: Replay evidence from the implementation commit
         if: ${{ matrix.python-version == '3.11' && hashFiles('evidence/v0.1.0-confirmatory.json') != '' }}
         run: python scripts/confirmatory_evidence.py verify

README.md

@@ -6,7 +6,12 @@
 
 **Fail-closed preflight and regression CI for supervised fine-tuning.**
 
-![SFTGuard confirmatory dashboard](docs/assets/sftguard-confirmatory-dashboard.png)
+![SFTGuard confirmatory evidence matrix showing every preregistered injected fault detected and no false positives across clean controls](docs/assets/sftguard-confirmatory-evidence.svg)
+
+*Generated from the sealed v0.1 artifact: 270/270 required signals across
+nine synthetic fault templates and 30 frozen seeds, with 0/30 clean-control
+false positives. This is internal detector regression evidence, not an external
+fine-tuning benchmark or safety certification.*
 
 I built SFTGuard to catch a narrow set of expensive-to-discover pipeline
 mistakes before an adapter is released: malformed chat records, conflicting or
@@ -155,6 +160,8 @@ The viewer is a static React application with no analytics, report endpoint,
 external font, or server-side storage. An imported report stays in the current
 browser tab.
 
+![SFTGuard browser-local report viewer displaying the confirmatory matrix](docs/assets/sftguard-confirmatory-dashboard.png)
+
 ```bash
 cd web
 npm ci
@@ -244,6 +251,7 @@ python -m pip install -e ".[dev]"
 python -m pytest
 python -m ruff check src tests/python scripts
 python -m ruff format --check src tests/python scripts
+python scripts/generate_readme_assets.py --check
 
 cd web
 npm ci