[stacknil]

Problem

Equivalent dependency representations could change explicit evidence-policy decisions, while malformed policy or provider evidence could cross validation boundaries. This change adds a reproducible, offline decision-stability qualification and the narrowly justified repairs it exposed.

Invariants

Canonical identity, observed metadata, recorded evidence, policy authorization and consumer decisions remain distinct. Representation-only equivalence is STABLE; changes to active rules or relevant evidence can produce JUSTIFIED_REVERSAL. Contradictory identities remain invalid input. Unknown evidence never becomes proof of authenticity or package safety.

Design decisions

• A standard-library-only expected-value module specifies 72 literal policy microcases independently of the production evaluator. The campaign covers 55 transitions and fourteen reached negative controls, N01–N14.
• C02 uses PEP 503 package-name equality for PyPI missing-attestation exceptions, preserving name-level scope and the unavailable/unverified boundaries.
• C03 uses canonical ecosystem/name/version for explicitly enabled PyPI lookup and cache isolation, including valid PURL-only versions. Both CLI input indexes are validated before either provider starts; direct/batch PyPI APIs validate identities before lookup. Observed fields remain unchanged and versionless components never request latest.
• C04 binds explicit Scorecard repository identity using ASCII-only case equivalence. Policy parsing rejects Boolean integer fields/non-string keys; Scorecard parsing rejects Boolean/nonfinite scores.
• Tests check JSON, summary, policy sidecar, Markdown, the supported SARIF subset and exit-code consistency, including post-serialization evidence against literal expectations.

Failure model and rejected alternatives

The regressions cover contradictory/duplicate inputs, cache contamination, omitted violations, severity/suppression corruption, unavailable/missing/unverified confusion, malformed provider evidence and misleading report agreement. Universal alias matching, Unicode repository alias folding, latest-release inference and a new unknown-score blocking policy were rejected. A failed cache experiment based on an invalid distinct-release assumption is retained as a rejected oracle hypothesis, not a production defect.

Compatibility and non-claims

Existing policy/report schema identifiers, risk taxonomy, strict SemVer classification and offline defaults remain unchanged. No dependency or workflow changes are included. Invalid Scorecard evidence follows existing error/unknown handling; an available-score-only policy can therefore pass without usable score evidence. This is not a vulnerability scanner or package safety certification. No live PyPI/Scorecard enrichment or production PyPI publishing was exercised. The qualification's network-denial audit recorded zero actual attempts; explicitly enabled production lookup remains the existing opt-in behavior.

Validation

• Baseline: 473 passed.
• Final candidate: Windows CPython 3.12, 817 passed × seeds 103/117/120; native Linux/ext4 CPython 3.11.17, 817 passed, plus targeted reruns.
• N01–N14 detected; golden artifacts passed; network-denial audit: 0 actual attempts.
• Changed-file Ruff correctness checks passed. Full Ruff output exactly matches the baseline's existing F821 in test_policy_scorecard.py.
• Original signed/DCO commit: dce5082e3f943c0e796e74607aa65c02cf86f10c; 30-file frozen diff; semantic tree 0760a5543df366f3a8237c60a187cde7f3aae0b0.

The owner supplied the external disposition ACCEPT_FOR_LOCAL_COMMIT, conditional on exact source identity, after targeted independent execution and review for the defined scope. That external review was not an independently executed rerun of the entire 817-test suite. Frozen qualification documents retain their historical audit-pending checkpoint; the later external disposition is preserved separately. Private transcripts, machine receipts and the private audit package are not part of this PR. Hosted CI is a separate integration gate and will be assessed from actual jobs, not inferred from local results.

Rollback and review

Revert this single contribution to remove the qualification and bounded repairs; no schema migration or data conversion is required. This is a draft because the coherent qualification spans 30 files. It does not close or claim completion of #103 or #117. Merge requires a separate decision; no merge or release is authorized by this integration task.