# Closeout

- **Event:** [AIE Code Summit 2026 Hackathon](https://cerebralvalley.ai/e/aie-code-2026-hackathon)
- **When:** Sat, Nov 7 at 9:00 AM – 10:00 PM (PST)
- **Where:** San Francisco, CA
- **Team:** [Benjamin Shyong](https://cerebralvalley.ai/u/BenjaminBear)
- **GitHub:** https://github.com/benikigai/closeout
- **Demo video:** https://youtu.be/z_zCCMzQvvQ
- **Gallery:** https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/1

Factory failures get closed on a yield bump. Then the defect comes back. Closeout makes a fix earn its closeout.

We built an agent-assisted failure analysis loop around a fictional beacon assembly line: triage failed units, rank competing causes, choose units for physical investigation, patch the firmware, challenge the claim, and require a fresh build to pass.

The demo starts with a cold-start failure blamed on materials. The parts are fine: firmware writes calibration at 50 ms, but the cold LED driver is not ready until 180 ms. The correction waits for readiness. Judges can run both C firmware revisions on the live server, inspect the evidence, and try the shortcut of widening the optical limit. That shortcut admits a known-bad physical defect, so the negative control rejects it.

SubQ Snap 1.1 scores hypotheses and recommends the next probe. Meta's Muse Wonder generates candidate C patches. Gemini 4 Argon challenges the root-cause claim from held-out evidence without the fixer's reasoning. The MongoDB Atlas ledger implementation stores the Major Issues List, category history and FA evidence, with checks against confirmation without findings or closure before monitoring passes.

Live demo: https://closeout-web.vercel.app
Choose Reproduce the failure, Inspect the evidence, Test the shortcut, Run the corrected firmware, then Run 20 held-out builds. The site reports 320 synthetic units: cold failures fall from 113 to 0, all 15 physical defects remain detected, and no warm controls regress. These results come from compiled C execution.

Built for Trust but Verify. All products, units and measurements are synthetic. Browser interactions execute real C against a simulated hardware layer; model assessments and the saved engine trace are explicitly replayed. The browser does not generate a new patch or mutate the Atlas ledger. Scores guide investigation; they are not calibrated probabilities.

Failure in. Evidence out.

## More from AIE Code Summit 2026 Hackathon

- [EvoArena](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/2)
- [ClauseLens](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/3)
- [Self recovery mech interpiblity](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/4)
- [self improving on-call agent](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/79)
- [Gregg](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/80)
- [Pathlock](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/81)

---

Markdown version of https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/1. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt