Do two coding agents catch each other's mistakes? Codex and Claude build the same task in separate worktrees, review each other's diffs, and plain code decides what passes.
Result: In the one recorded live Codex-versus-Claude contest, Codex's review caught a Unicode lowercasing bug in Claude's first implementation, and Claude fixed it in revision. That is one observation, not a rate; it motivated the controlled experiment in agent_reliability_lab.
Why it matters: A model's review can reject a candidate but can never rescue one that fails a required check. Scoring reads only deterministic check results, so the harness, not the models, owns the pass/fail decision.
Status: Prototype. Its design fed into institutional-ai. Later additions (a single-agent run mode, task queues, a local control panel and an Anthropic API adapter) have only smoke runs (recorded in STATE.md and results/; the control panel was exercised only with fake CLIs) and no controlled results.
- Each agent works in its own Git worktree cut from one frozen commit, and every candidate is re-checked in a fresh detached worktree of its snapshot.
- 197 offline tests with scripted providers and fake runners; no credentials or model calls needed.
- The contest record is prose in VALIDATION.md; its audit archive is kept outside the repository.
Run mode (October 2026): the harness runs Claude Code or Codex headless on a task file, checks each step itself, prices each call from a user-supplied pricing table (unpriced calls are recorded as null) and writes a build journal per queue. See docs/RUN_MODE.md for run and the cost ledger, and docs/GUIDE.md for the control panel; queues and build journals are so far described only in STATE.md (2026-10-04 entry).
pip install -e '.[dev]' && pytest -q # 197 tests collected; all passed on 2026-10-06