up

bench/reachability-benchmark/docs/governance.md

@@ -0,0 +1,41 @@
+# Reachability Benchmark · Governance & Maintenance
+
+## Roles
+- **TAC (Technical Advisory Committee):** approves material changes to schemas, truth sets, and scoring rules; rotates quarterly.
+- **Maintainers:** curate cases, review submissions, run determinism checks, and publish baselines.
+- **Observers:** may propose cases and review reports; no merge rights.
+
+## Release cadence
+- **Quarterly update window:** publish new/updated cases and hidden test set refreshes once per quarter.
+- **Patch releases:** critical fixes to schemas or scorer may be shipped off-cycle; must remain backward compatible within `version: 1.x`.
+
+## Hidden test set
+- A reserved set of cases is held back to prevent overfitting.
+- Rotation policy: replace at least 25% of hidden cases each quarter; keep prior versions for audit.
+- Hidden cases follow the same determinism rules; hashes and schema versions are documented internally.
+
+## Change control
+- All changes require:
+  - Schema validation (`tools/validate.py`).
+  - Deterministic rebuild (`tools/build/build_all.py` with `SOURCE_DATE_EPOCH`).
+  - Updated truth files and baselines.
+  - Execution log entry in `docs/implplan/SPRINT_0513_...` with date/owner.
+- Breaking changes to schemas or scoring rules require TAC approval and a new major schema version.
+
+## Determinism rules (global)
+- No network access during build, analysis, or scoring.
+- Fixed seeds and sorted outputs.
+- Stable timestamps via `SOURCE_DATE_EPOCH`.
+- Telemetry disabled for all tools.
+
+## Licensing & provenance
+- All public artifacts are Apache-2.0.
+- Third-party snippets must retain attribution and be license-compatible.
+- Each release captures toolchain hashes (compilers, runners) in the release notes.
+
+## Incident handling
+- If a nondeterminism or licensing issue is found:
+  1) Freeze new submissions.
+  2) Reproduce with `ci/run-ci.sh`.
+  3) Issue a hotfix release of truth/baselines; bump patch version.
+  4) Announce in release notes and mark superseded artifacts.