Skip to content

Evidence index

The reviewer's entry point to the evidence behind the repository-quality assurance case: what each class of evidence supports, where it lives, and what keeps it honest. A few rows are not cited by the case directly but contextualize it; they are listed so the picture is complete. For readers who want the conclusion chain without walking the full case.

Evidence is only listed here if it is reproducible — a command, a gate, or a committed record. Prose that merely asserts quality is not evidence and does not appear.

Evidence Supports Where Kept honest by
Capability matrix What the tool can do, per capability feature-matrix.md feature_matrix_check: supported requires a cited, existing test
SACM conformance matrix What the library implements of SACM 2.3 sacm-conformance-matrix.md sacm_matrix_check: verified requires an ID-bearing test; cited paths must exist
Compliance-point tests The four claimed interchange points sacm-compliance-points.md SACM23_CP_001..004 — the unit-of-interchange tests (strict-load root assertions)
Matrix-completeness audit That the matrix is missing nothing unexamined sacm-matrix-completeness-audit.md Findings tracked as issues #333–#337; frozen into each evidence package
Verification records That verified meant a real, adversarial pass verification/ verification_index_check; FAIL rounds committed — the generated index is the count source, and FAIL verdicts outnumber passes
Preservation record That data loss was measured, fixed, and pinned sacm-integration-preservation.md Byte-pinned SACM23_LIB_002 tests; mechanical known-lost-kinds sweep
Layer gate + self-test Dependency direction cmake/check_layer_gates.cmake layer_gate_negative_check feeds it violations it must reject
Architecture record Subsystem responsibilities and ownership layers-and-ownership.md documentation_check: must describe every gated subsystem
Migration plan What is transitional and its retirement conditions legacy-bridge-migration-plan.md Observable exit criteria per phase
ADRs Why boundaries are where they are decisions/ One decision per record; append-only
Quality baseline Measured repository state repository-baseline.md Reproduction commands on the page; deliberately a dated snapshot
Hotspot register Where cleanup pays, ranked by risk hotspot-register.md Published inputs and commands; staleness criterion on the page
Code quality controls Warnings, static analysis, sanitizers code-quality-policy.md Warnings-as-errors in CI; version-pinned clang-tidy ratchet; ASan/UBSan workflow
Documentation authority One canonical home per policy documentation-map.md documentation_check: links, reachability, generated banners
Release evidence package Claims bound to an exact release sacm-conformance-statement.md evidence_package_check; generated by CI from a clean checkout, attached per release
Agent authority What AI agents may do and write .agents/ agent_definition_check: adapters generated, drift fails
Consent boundaries No data leaves without explicit consent ADRs 0005, 0007, 0010 Recorded decisions; tests on the consent seams
Security policy + parser hardening Reporting path; XML attack surface SECURITY.md SACM23_SEC_001_RejectsDoctype: DOCTYPE/ENTITY rejected before parsing
Test suite + CI Everything above, continuously .github/workflows/ (CI, coverage, sanitizers, clang-tidy) Three platforms; ctest gates labelled gate run in about a second

How to audit a claim from here

  1. Pick the claim in the assurance case.
  2. Follow its evidence link in the table above; each destination states its own reproduction command or names its gate.
  3. Check the challenge register — if the claim has a live counter-claim, the honest reading is the bounded one.
  4. For a release rather than main, download that release's evidence package; it freezes the matrix, the audit, the records and the test results at the released commit.