[JacobStephens2]

Summary

Grader Scores + latest human rechecks
  → paired suite calculation under the fixed Bar
  → unreviewed borderline or decisive Scores?
      yes: waiting_on_rechecks + pending Trial IDs
      no:  pass / fail / inconclusive

thirdshift recheck <experiment> <trial> pass|fail --reason <text> appends a timestamped decision beside the original Grader Score. Optional --wrong-finding true|false corrects that flag. History, the updated Verdict and pending list are saved atomically under the Experiment's exclusive lock. experiments <id> --json exposes pending_rechecks; rechecks never rerun Trials or Graders.

Closes #1006

Evidence

• Before: cargo test --test experiments rechecks::decisive_score_waits_for_recheck_and_retains_both_scores_and_history -- --exact failed: actual pass, expected waiting_on_rechecks.
  After: the same regression passes, including finalization, immutable Grader evidence, appended history and a changed Score making another Score decisive.
• Six public CLI tests cover decisive and nondecisive borderline Scores, revisions, wrong-finding corrections, invalid/concurrent writes, legacy Experiments and dropped Void tasks.
• cargo check, cargo fmt -- --check, and cargo clippy --all-targets -- -D warnings pass.
• The first full selection caught the existing compact-help regression (66 lines, maximum 65). The defaults/manual note is shortened while keeping the new command; the existing test is retained.
• Final validation at f0e07ec: /usr/bin/time -p scripts/affected-tests 72f65acf8e6bbb7824161bc2e082255b17dd86bb passed in 1,456.56 seconds (24m 17s), including compilation. The selector ran all 43 site tests and the full Rust suite; all 52 Experiment tests passed, including the six new recheck tests. The fixed point is the captured GitHub issue-977 head, rather than the stale local ref.
• All CI checks are green on the same head, including the Console/site browser tests, Linux Rust checks and release builds.

Test seam: the whole binary against fake external Harnesses in the existing Experiment integration scenario; observations use the CLI and its saved public result format.

Review coverage

Independent Standards and Spec reviews used issue-977 at 72f65ac as the fixed point. Standards raised two duplication heuristics, both extracted and confirmed addressed; Spec raised no findings. Both axes read all 18 changed files, including generated manual HTML. No generated-output exclusions or unread changed files remain. After their final follow-up at 5c495e7, five needless borrows were removed from the already reviewed tests/experiments/rechecks.rs, and the defaults/manual help note in the already reviewed src/main.rs was shortened to retain the existing 65-line limit. These corrections are checked by clippy and the existing compact-help regression.

Reports, with each axis's final files-read list, are saved in the worktree at .thirdshift-review-F4HeAt/standards.md and .thirdshift-review-F4HeAt/spec.md.

Unaddressed findings

• Standards: None. Both findings addressed.
• Spec: None.
• Security: None. Guidance-mode review of the 72f65ac...HEAD diff (new recheck command, recheck-gated Verdict, experiments --json pending list): all new inputs come from the same local principal that owns the Experiment files; Experiment IDs are restricted to 64 hex before path use, Trial IDs are matched by string equality and never used as paths, --reason text is stored and rendered only through JSON encoding, and the exclusive lock plus atomic save reuse the unchanged helpers. No trust-boundary violation found, so no proof-of-concept failure to fix.

Security review outcome

No unaddressed introduced findings.

Merge Danger

Door: two-way

No Scores are deleted or overwritten. Previously completed Experiments acquire the recheck gate when read. The new waiting_on_rechecks Verdict value requires this version of the reader while an Experiment is waiting; older binaries do not recognize it.

Blast Radius: Experiments

The latest human decision changes the calculation and may identify further decisive Scores. Command records retain their invocation's outcome; Experiment views show the current Verdict.

Built with codex · gpt-6.1-sol · xhigh