# MERGE-GATE

- **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:** [Tanjim Ahmed Zabeer](https://cerebralvalley.ai/u/Taaz), [Peter Rubin](https://cerebralvalley.ai/u/PeterRubin22), [Michael Wei](https://cerebralvalley.ai/u/michaelwei07)
- **GitHub:** https://github.com/prubinadmin/MergeGateAIEHackathon
- **Demo video:** https://youtu.be/iromugChO3s
- **Gallery:** https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/40

Merge Gate is an independent verifier for AI-written pull requests. It never trusts the PR's own tests or description. Instead, Meta's Muse model writes spec-only tests from the issue and the public interface, a sandbox runs them on base and head, fourteen deterministic anti-gaming rules (deleted or weakened tests, mocked subjects, out-of-scope test edits, perf regressions, complexity inflation) inspect the diff, and plain code decides BLOCK, REVIEW or PASS. No LLM ever decides a verdict, and every run is hash-chained and replayable.

The verdict goes to the engineer who actually wrote the changed lines (git blame plus SubQ Snap to pick who and how urgently), as a real iMessage. They can reply BLOCK, WHY, OK, FALSE ALARM, or type a rule in plain English, and that sentence becomes a pytest test stored under their name and run on every future PR touching those files. A text can stop a merge; it can never approve one. Proposed gate changes are adopted only when held-out PRs confirm them: v1 to v2 raised held-out catch rate from 0.55 to 0.73 with zero held-out false blocks. A rollback command re-verifies merged commits and opens revert PRs.

Merge Gate runs as a five-agent Fetch.ai system usable from ASI:One (orchestrator plus TestWriter, Sandbox, Rules and OwnerRouter specialists, with an interactive verdict card), can scan any public GitHub PR by name, logs every run to MongoDB Atlas, and gates its own repository: it blocked several of our PRs today and we fixed the code, not the rules.

Honest numbers on our 30-case benchmark: Merge Gate v2 catches 17 of 22 bad PRs with 2 false blocks on 8 clean ones; an LLM that simply reads the diff catches 22 of 22. Our edge is independence from the diff, evidence on every verdict, routing to the author (30/30 versus 17/30 for CODEOWNERS), learning from the author, and reproducibility.

Live demo: https://dist-phi-five-99.vercel.app (streams real runs from our gate). Repo: https://github.com/prubinadmin/MergeGateAIEHackathon. Everything was built during the hackathon; synthetic commit authors in the demo repos are disclosed in the README.

## More from AIE Code Summit 2026 Hackathon

- [rsi_today](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/37)
- [Sisyphus](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/38)
- [Museance](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/39)
- [KT-coding-agent](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/41)
- [Doppel](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/42)
- [Prysma Labs](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/43)

---

Markdown version of https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/40. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt