Continue improving the repository and do something concretely useful.
The repository had one intentionally blocked open PR (#159) requiring genuinely independent human review. Rather than bypass that gate or create another broad controller, the next unblocked high-value task was selected from issue #180: calibrate Phase B2 successor-v2 Task 003.
Task 003 is directly relevant to the repository’s current branch-convergence problem. Its frozen objective is to ensure that historical merged-PR evidence cannot classify a branch as integrated/cleanup-ready when the branch’s current head SHA was not observed.
The successor-v2 scaffold is deliberately pre-outcome. Publishing the canonical
fix first would contaminate the task before its provisional evaluator had been
calibrated. Therefore this iteration adds only evaluator-owned calibration
machinery against immutable source a69aa0ae....
The calibration constructs:
head_sha is not None before an
exact merged-PR head can match; andThe canonical EvaluatorPlan v0.4 must accept the first and reject the second. A separate behavioral matrix must also prove that:
The first exact calibration head passed in GitHub Actions:
33201309072;98951164080;9697855139;sha256:9212f58c076324cb9aa6659c1c5cd54f9c61746d1de51647bc3f5576cf972136.Straightforward candidate:
passed;accept_candidate;1/1;1/1;safe_unobserved_head_matrix_passed=true;sha256:b5c2ce3beba36f4a6d8ed45497f1384063a542dbedf0153718f38f165c7ea5d3;sha256:1ec0fa91f62a3bc86c2562919b2790eae057dbb749994037d72b857f89c3a078.Inert lexical decoy:
failed;reject_candidate;1/1;0/1;sha256:1fbd1a3bb430bb7e2ba90f9869ef175a67a6dfbf1779a0fcb0026d14c7215a77;sha256:7bb19325c1ed6c10f3a6b54e1f972145aff600e077aacc619a5e3c7a240df796.The workflow also confirmed that calibration candidates are not benchmark outcomes, the canonical verifier remains metadata-only, and no candidate has canonical-write, push, merge, or automatic-selection authority.
A temporary same-PR scaffold-state update was deliberately removed. PR #198 is
kept as calibration machinery/evidence only, and the scaffold file is restored
byte-for-byte from main. If #198 lands unchanged, calibration registration
should be a separate small current-main PR that adds the run/artifact receipt,
removes only Task 003 from calibration_pending_task_ids, keeps
freeze_ready=false, and leaves actual Task 003 benchmark evidence pending.
This avoids coupling evidence generation to evidence registration and prevents a
fast-moving main from being overwritten by a stale scaffold snapshot.
This is calibration evidence, not a benchmark outcome and not a production patch. The successor-v2 scaffold remains unfrozen, task evidence stays pending, and a fresh novelty audit is still required before any future freeze.
The workflow uses read-only GitHub permissions, no secrets, no persisted checkout credentials, and no push/approval/merge/automatic-selection authority. Candidate code execution occurs only in a disposable evaluator-owned calibration checkout; the canonical v0.4 verifier remains metadata-only.
Related: #127, #180, PR #186, PR #189, PR #198.