← ClaudeAtlas

mutation-matrix-evidence-disciplinelisted

Load when editing hooks/mutation_matrix.py or hooks/mutations.json, running or citing an "authoritative"/"55/55 killed" mutation result, marking a mutation equivalent/unreachable, adding a test whose name asserts a property class, building any verification probe or self-check, or correcting a threat-model/README overclaim. A runtime-hook change that also adds/edits a test or makes a test/evidence claim ALSO loads skill-vetting-security-invariants. Harness-only work uses this skill rather than the security skill — that split does not suppress security-hardening-review-ops during a hardening campaign or a security-sensitive push/PR/merge.
F-e-u-e-r/opus-pack · ★ 2 · AI & Automation · score 68
Install: claude install-skill F-e-u-e-r/opus-pack
# Mutation-matrix and evidence-integrity discipline `hooks/mutation_matrix.py` (data in `hooks/mutations.json`) reverts each landed fix one at a time inside a throwaway git worktree at a named commit and requires the test suites to go red. It is this repo's instance of the pack's proof-gate doctrine (`ground-truth-gates`: a gate must be able to FAIL under the broken behavior, and "proof gate" is defined there) — but the tool that measures the tests is itself security-relevant, and during PR #83 it produced a false "55/55 killed" more than once. These rules make an authoritative result trustworthy; each traces to a defect the campaign hit and fixed. ## Two commands, do not confuse them - **`python3 hooks/mutation_matrix.py --authoritative`** — the closure run a report may cite: on a clean committed checkout it reverts every landed fix in a throwaway worktree and requires the suites to go red (`--help`: "the mode a closure report may cite"). Slow; produces the run-scoped record. - **`bash hooks/test-mutation_matrix.sh`** — the harness's own UNIT suite (it `exec`s `hooks/test-mutation_matrix.py`). It tests the tool, NOT your code, and is NOT an authoritative mutation run. Do not cite it as one. ## The core principle **A verification tool must be proven able to FAIL before its "pass" is trusted** (`ground-truth-gates`), and the evidence it emits must be bound to the exact code and run it claims. A mutation harness, a closure verifier, a status check — each is a claim