Review source mutation quality through deterministic owner replay

[raylee-hawkins]

Objective / new capability. Adds detection-quality review through the platform delegate. Hoxline renders measured results and verifies an entire saved report against fresh validation execution; changing its checksum cannot manufacture validity.

System role / why this matters. Review aggregation and deterministic owner replay. Platform identity is checked before and after delegation. Strict parsing and full report comparison reject metric, source, nested claim and authority edits. The existing case-growth pair is regenerated without increasing proof counts, public-safe status or closed cases. AI provides engineering labor only; this change adds no model runtime or triage approval path.

Files changed (8).

• .github/workflows/ci.yml
• README.md
• examples/case-growth/current-case-growth-index.json
• examples/case-growth/current-case-growth-index.md
• src/hoxline/cli.py
• src/hoxline/detection_quality.py
• tests/test_action_contract.py
• tests/test_detection_quality.py

Dependencies / topology. Stacked draft on existing hoxline #26, branch feature/hoxline-case-growth-convergence-v1; reviewed base branch was not edited. Depends on platform #88, validation #72 and detection #48. Organization #60 runs the new handoff in the canonical lifecycle. Existing case-growth and proof ownership remain intact.

Validation commands. Run from this repository with the scoped sibling checkouts at the reviewed revisions and documented dependencies installed.

python -B -m hoxline detection-quality --repo-root .. --detections-ref 56cad4f726c0d3988c9464693ce7c127c8f63cad --validation-ref c8f3f08995d78fa103115610e5c6a2b85ee61701 --platform-ref 0826b9916f926c204d44875f3ecd1c964a159125
python -B -m unittest discover -s tests
python -B -m pytest -q tests

Local validation. Unittest completed 93 tests with one existing skip. Full local Windows pytest: 244 passed, 1 existing skip, 2097.27 seconds. Final organization pytest: 244 passed, 1 skipped. Actual CLI integration blocked all 20 altered saved reports.

Current-head CI. Head be2cd4f08aa5fe04fff89435de47e20abef9f96e. PR-event Python 3.11/3.12 checks passed at the current head. Same-head push-event run 34038966614 also passed both versions.

Measurable and hostile results. The integrated controlled corpus covers HO-DET-001/009/010/011/012/013: 69 events, 33 positive and 36 negative. Before source repair: TP 30, TN 30, FP 6, FN 3. After repair: TP 33, TN 36, FP 0, FN 0; precision/recall/F1 1.0, false-positive rate 0.0. Real source mutants: 205 generated, 165 killed, 40 survived, 0 errors; mutation score 0.804878. Survivors remain visible. Actual Hoxline CLI replay blocked 20/20 tampered reports, including 8 authority-boundary edits and 3 AI-focused edits; recomputed checksums did not bypass fresh owner execution.

Windows and AI authority. Validation run 34038177088 passed hosted Windows Server 2025 and Ubuntu source-predicate execution and deterministic replay. This is offline controlled fixture execution, not endpoint or SIEM backend execution. AI authored candidate engineering changes and tests; it cannot change ground truth, authorize disposition, close cases, approve evidence or promote proof. Controlled auto-close was not implemented.

Proof boundary and scans. CONTROLLED_TEST_VALIDATED predicate/corpus result; NOT_PUBLIC_SAFE; human review required. No new proof record, ProofCard, runtime signal, real case closure, production performance or public-safe promotion is established. Changed-content scan found no unexplained credential, private-address, operator-path or employer-data exposure. Six path-pattern hits were escaped generic tool-name unit fixtures, not private paths. Claim wording was checked against these limits; adversarial vocabulary remains as test inputs.

Human review order and rollback. Review the existing convergence bases first, then detection #48 and validation #72 together, platform #88, Hoxline #27, and organization #60 / Website #84 as a paired integration. Proof #87 supplies unchanged proof authority. Existing presentation branches remain separate. With drafts unmerged, withholding integration is the rollback; any later revert must be a reviewed scoped change that restores owner/consumer pins together.

Merge gate. Green CI is not approval. AI review is not human approval. No merge is authorized. Visible eligible human GitHub review and human-supplied MERGE_APPROVED remain required. Current readiness: NEEDS_HUMAN_REVIEW.