This directory holds bounded reviews and evidence snapshots. Each record states what was inspected, at which revision, and what was concluded at that moment.
Read every record here as a snapshot, never as current status. Repository state,
open pull requests, branch protection, and issue status all move; an audit is
retained because it preserves what was inspected and why a decision changed, not
because its situational claims are still true. Current authority lives in
../decisions/, ../specifications/README.md,
../architecture/README.md, and
../../PROJECT_RULES.md.
Each entry below carries the baseline the record itself declares. Four records declare no explicit baseline revision; that is noted rather than invented.
Whole-repository snapshots, ordered by date.
main at 8614a669. Executable-contract phase reached, first
real worker runtime not yet on main.e86fec87. Argues the scarce resource is independent
contact with reality. Self-declared as a proposal, not a status claim.649df7b7. Structural correctness pass across code,
contracts, README, and navigation documentation.Each answers one bounded question rather than surveying the repository.
Which module does the v0.4 calibration actually verify with?
— 2026-09-05; baseline c35b31e. Traces a red canonical-calibration guard to a
superseded patch verifier that was restored after #171 and imported by nothing.
A workflow step named it as “the calibrated evaluator path” while dispatch went
elsewhere, and the path filter guarding that run watched a file the run did not
read.
Do the gates check what they claim?
— 2026-09-02; baseline 31b8f18. Mutation-tests each Phase 0 gate by
introducing the defect it claims to catch. Five of six caught it; the schema
gate reported OK: schemas valid while covering 5 of 32 schemas.
origin/main at 566bee13. Read-only
review for issue #151. Explicitly AI-assisted independence, not human or
organizational independence.e16a8d0. Answers which open issues name a precondition that is
a measured property of the repository rather than a task, and what that
precondition currently measures. Backed by tools/issue_evidence_gate.py,
which fails when a registered precondition becomes met rather than when one
stays unmet.Bounded classification passes for issue #152. Each freezes a source revision, classifies a named cohort, and reports a measured observatory delta. None treats a lower warning count as a standalone objective. Later passes do not edit earlier records; a superseded baseline is corrected by a new record, not in place.
d0bafb7f; seeded 15-document sample drawn before
classification so the cohort cannot be cherry-picked.accepted_decision_without_document_link warnings.7f28bc3c; continuation pass for the single warning that appeared
after the record above, and a measurement of how far the #152 baseline had
drifted.State the date and the exact baseline revision in the document, give the file a dated name, and add one line to the group above that it belongs to. Do not edit an existing audit to reflect later repository state — write a new record that names what changed.