Skip to content

Derive canary qualification from retained evidence #53

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
vriesd merged 1 commit into from
Aug 28, 2026
Merged

Derive canary qualification from retained evidence #53

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions .agents/plans/03-assurance-hardening/decisions.tsv
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,3 +38,12 @@ ts phase decision why evidence result
2026-08-28T03:32:07Z phase-5 bound the complete release environment and grader closure first-pass review found duplicated cell construction, caller-defined host digests, and incomplete grader hashing releaseCellsFor; releaseHostConfigSha256; releaseGraderBundle import-graph scanner exact cell and block ids, seed, budgets, Linux OpenCode 1.18.6 host, and every transitive local evaluator file are authoritative
2026-08-28T03:32:07Z phase-5 removed legacy summary qualification authority the import-only summary path preserved a second executable policy surface scripts/qualify-release.ts; tests/release-qualification.test.ts V2 atomic report, exact persisted catalog, packed artifact, and canonical authority are the only qualifier inputs
2026-08-28T03:32:07Z phase-5 completed repository-owned release policy verification final review findings fixed stale authority digests, arbitrary decision-input hashes, extra model spend, and host-policy drift bun run check; bun run replay 577 pass, 1 intentional skip, 13 of 13 replays, 892 source bytes headroom
2026-08-28T04:01:46Z phase-5 merged repository-owned release policy the exact PR head passed independent shipping verification and every required check PR 52; merge cda1234 merged to main
2026-08-28T04:01:46Z phase-6 selected a minimal evidence-derived canary pipeline the independent architecture judge required exact preparation and installation binding, evidence-derived claims, deterministic actor pseudonyms, and a retained host artifact three architecture candidates; independent judge candidate A base with candidate B installation binding and candidate C evidence-first facts
2026-08-28T04:01:46Z phase-6 captured the committed canary contradiction before implementation the 8.1.2 record claimed every check passed while its retained close concluded completion-unsupported tests/eval-canary.test.ts; focused red typecheck missing derivation API failed red as expected
2026-08-28T04:01:46Z phase-6 derived canary claims from retained executable evidence caller status, checks, actors, and host identity were untrusted assertions; installation also needed exact byte proof scripts/eval-canary.ts; focused canary and release tests status and checks now rederive from measured tarball/plugin bytes plus sanitized session/transcript structure; future legacy records refused
2026-08-28T04:22:39Z phase-6 fixed all valid first-wave interrogate findings minimal record-shaped sessions, recursive transcript scans, path reads, and file-only installation proof could still admit synthetic evidence SessionSchema and assuranceProjection; OpenCode message/part parser; readWorkspaceTestReport; runtimeIdentity retained evidence is domain-valid, same-handle bounded, transcript-spoof resistant, and tied to the loaded plugin entry digest
2026-08-28T04:22:39Z phase-6 invalidated the caller-attested 8.1.2 canary grandfathering the known contradictory record would preserve the authority bypass this phase removes CanaryRecordSchema; scripts/release-metadata.ts; release metadata tests all accepted canaries now require derivation and installation evidence; the historical format is rejected
2026-08-28T04:22:39Z phase-6 bound reviewer independence through host lineage separate reviewer messages and task calls did not prove they represented the same child session and model task parentSessionId, sessionId, model metadata; negative mismatch test reviewer dispatch passes only when task lineage matches the observed manager and reviewer actors
2026-08-28T04:22:39Z phase-6 completed corrected whole-product verification the review fixes changed runtime status output, evidence parsing, and release metadata boundaries bun run check; bun run replay; pinned OpenCode live smoke 585 pass, 1 intentional skip, 13 of 13 replays, live OpenCode 1.18.6 pass, 339 source bytes headroom
2026-08-28T04:23:41Z phase-6 closed the final transcript and lineage review findings flat synthetic transcripts and unlinked reviewer observations remained broader than the real OpenCode evidence boundary root export messages plus tool parts; parent-child-model lineage match; final three-model review 586 pass, 1 intentional skip, 13 of 13 replays, pinned live smoke pass; all final reviewers green
Loading