[JacobStephens2]

Summary

The write-up for #429's replays: 183 read-only review sessions over past work, graded against the bars settled in Q6, Q29 and Q30 of the grilling log.

docs/research/
├── second-opinion-replays.md        # method, results, what passes, caveats
└── second-opinion-replays/
    ├── replay.py                    # builds each throwaway repo + prompt, runs one Harness's units
    ├── stages.sh                    # runs stages in order, pausing Weeding/Security cron lines
    └── timing-checklist.md          # the checklist arm's list, written blind to the bugs

Piece	Result
Second opinion in a Run	Fails both ways: Codex caught 1 of 6 timing bugs (fresh Claude caught it too), and Claude caught 0 of Codex's 3. The new Harnesses caught 1 between them
Timing checklist	No gain: 4 caught, the same as fresh Claude
Plan check	Passes only for Claude checking Claude's plans (8 of 10 stops right). On Codex's plans, which Weeding writes now, neither arm passes
Second opinion after the Spec review	Passes for either family: Codex in 16 Specs (2 of 25 wrong), fresh Claude in 19 (9 of 40 wrong), mostly different bugs
Spec review on its own Harness	Fails: Codex 22 Specs vs the recorded review's 23

Findings in private repositories, and ones still unfixed, are left out of the note on purpose.

Evidence

• Before: Multi-family review: have another model family review a Run's work #429's bars were set, but no session had run.
  After: all 183 sessions finished:

  • every one exited 0 within its timeout;
  • none changed its repository (HEAD and git status compared before and after);
  • the two gh flags were quoted search patterns.

  Grades are in ~/.thirdshift/replays/2026-10-09/grading/.

Merge Danger

Door: two-way

Docs and a standalone script only; nothing the binary builds or runs.

Blast Radius: none

🤖 Generated with Claude Code