# KINETIC-ORACLE

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Jack W](https://cerebralvalley.ai/u/Jacksthename)
- **GitHub:** https://github.com/p1nk-n0de/kinetic-oracle-hackathon
- **Demo video:** https://p1nk-n0de.github.io/kinetic-oracle-hackathon/
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/492

KINETIC-ORACLE gives tool-using AI agents an independent way to find out whether what they did actually worked.

What is new, and what is not. Built during the hackathon, on 30 Oct 2026: the SAM 3.1 tracker, the Muse-compiled frozen contracts, the residual verifier, the Muse Spark → Muse Wonder repair loop, all 21 runs, the replay verifier, the tests and the dashboard. The demo video was made on 2 Nov. Pre-existing and disclosed: the Godot test game and the Godot MCP server that drives it, both built on 23 Oct 2026 and not part of the entry. The README lists every component with its date. The public repository is a sanitised copy: a 2 Nov commit replaces the name of the author's private tooling with visible redaction markers and adds checks for that. No result changed (docs/SANITIZATION.md).

Agents treat a tool's "OK" as proof, and it isn't. In our demo, a Godot game is told to walk the player into a low hurdle. The tool returns is_error: false, and the player walks straight through.

KINETIC-ORACLE watches instead:
1. Muse Spark turns the objective into a falsifiable contract, which is frozen (hashed) before anything is measured.
2. The action is rendered, and SAM 3.1 tracks the objects named in the contract through every frame.
3. A deterministic residual decides PASS, FAIL or INCONCLUSIVE from those measurements alone.
4. On FAIL, Muse diagnoses blind, using only the numbers and an allowlisted set of inspection tools, and patches a working copy of the game.
5. Everything is measured again under the same contract.

In the hero run, Muse Spark patched the game and stated that the hurdle now blocked the player. The verifier re-measured: same failure, same frame (FAIL at frame 60, +236 px). The loop escalated to Muse Wonder, which removed the fault Spark had missed, and the same check then passed (max E −13 px). The models never get to declare success.

Results: 21 real closed-loop runs and 40 measured passes. 39 agree with the game engine's hidden position record, 1 abstained, 0 were wrong. Every verdict replays offline from the stored SAM output with zero API calls, and the test suite passes: 306 tests from the build, plus 27 that check the public copy's redactions. This shows the mechanism working on one scene family with deliberate variations; it is not an accuracy rate.

## More from Meta Global AI Developer Hackathon

- [SeismoRescue](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/489)
- [ZionMatrix - ASO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/490)
- [Swastik Mishra](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/491)
- [zhalpakov](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/493)
- [Kuron Universe](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/494)
- [Savy Electrical](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/495)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/492. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt