# SuperHornet

- **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:** [Kartik Bhardwaj](https://cerebralvalley.ai/u/Kartik_05)
- **GitHub:** https://github.com/Kartik-99999/project-chimera
- **Demo video:** https://youtu.be/zY_v4SWAIqc
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1785

AI browser agents fail the way confident interns do: they claim success when the task isn't done. Project Chimera is the referee — a deterministic, model-free harness that judges model-driven browser
  agents from page state it captured itself, never from the model's own verdict.

  How it works: Meta's Muse Spark (muse-spark-1.3, via the Meta Model API) drives a real Chromium browser as our Baseline B agent. Each step it gets the page's element list plus visible text and replies with
  one JSON action (fill, click, wait, navigate, done). Chimera validates every action pre-execution (elements by harness-issued index only; navigation locked to the task page), observes it through a
  recording page, then scores the episode with an independent P1–P3 verifier reading harness-captured snapshots. A done{success} the verifier rejects counts as a false success — the metric this project
  exists to measure. No model is ever used in verification.

  Measured result (run model-b, 30 episodes + warm-up, 2026-11-04): 20 model claims, 18 verified, 2 false successes caught (upper 95% bound 30.1%). Both catches are genuine model failures: unsure a padded
  name "took," the model submitted 2–3 times, saw the confirmation, and claimed success; P2/P3 rejected both for duplicate entries. The 10 disabled-button episodes ended honestly with zero claims. 116 model
  calls, 181,910 tokens, 0 unauthorized actions, 0 safety violations. Every figure recomputes from per-episode traces and <run>.model.json sidecars in the repo.

  Intellectual honesty, stated plainly: at a matched 5 s budget our deterministic script equals fixed replay 20/20 — so the headline is not "our agent wins," it is "the referee works." n=30 is small, one
  task, one prompt; we claim the detector fires, not a general failure rate.

  Also in the repo: the deterministic drivers the referee was built for, plus the Workbench (run/inspect), Workflow Studio, Campaign Lab, Research Console, a six-page evidence site, 847 Python + 88
  JavaScript tests, and frozen dev/eval/audit benchmark evidence.

  • Repo: https://github.com/Kartik-99999/project-chimera
  • Demo video (180 s, captioned): https://github.com/Kartik-99999/project-chimera/releases/download/v1.0-demo/chimera-3min.mp4
  • Method + full numbers: experiments/model-baseline-v1/REPORT.md

## More from Meta Global AI Developer Hackathon

- [Heimdall](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1782)
- [Evans Mgode](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1783)
- [DecideOnce](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1784)
- [Endovascular ToolKit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1786)
- [RECALLZERO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1787)
- [WellArtDev](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1788)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1785. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt