Star
Search corralai.dev

Completed + replay + agent windows

The Completed tab lists every finished audit run the brain itself coordinated (a corral certify --adversarial or start_adversarial_run pool run); a certify --local or --repo run never contacts a brain, so it does not appear here. Opening one’s replay plays back the whole audit — shard claims, survivor findings, and jail executions — through the same control bar this site’s recordings page uses. See mission history + replay for how the stream is reconstructed from durable rows, not a live recording.

The Completed tab, expanded into the “what you just watched” hero breakdown for “certify more-itertools: full_permutation”: a verdict strip reading “CERTIFIED · killed 18/20 · 2 survivors”; below it an expandable panel showing what was tested (the goal and the audited file), the decorrelated herd that ran it (gemini-3.5-flash/mutant-generator, claude-sonnet-5/test-writer, claude-haiku-4-5/test-critic, each with its host and jail), what happened in the jail (20 mutants planted, 18 killed, 2 survivors — 1 proven by the test-writer, 1 still unproven), and an honest-limits section quoting the test-critic’s full argument: “test_full_permutation asserts range(3) == (0, 1, 2), which is always False in Python 3 — the assertion never fires, so this test proves nothing”; “details” and “replay” buttons below

Every replay opens on that same hero verdict strip — CERTIFIED or NEEDS-REVIEW, the kill-rate, and the survivor count at a glance — before you even press play, and it stays expandable throughout: what was actually tested, the decorrelated herd that ran it, what happened inside the jail, and the honest-limits section with the test-critic’s actual argument, not a summary of it, so you can judge the advisory second opinion yourself instead of taking the badge’s word for it.

Task-story modal

Clicking into an individual shard opens its task-story modal — what that seat did, in order. For a test-critic seat this now surfaces the critic’s full argument (its finding evidence: the line it flagged, the reasoning, and what it would have said in review), not just the one-line target it flagged, so the advisory opinion is auditable rather than a black box.

The replay control bar after playing an audit run through to the end: a play button, a fully-advanced scrub bar, the position label “45 / 45”, a 1x speed selector, and an exit-replay button

Files

Alongside the task-story modal, a replay’s Files tab reconstructs who touched which path and when, from the herd’s recorded file claims — scrub- driven, the same way the story lens is. It’s a tree, colored per agent, with a claimed/touched count in the header; paths only, since the tape doesn’t capture file contents. For an audit run, the same tab shows the authored test instead (whichever the run actually produced) — the file tree is the build-run view of the identical scrub-driven space.