# Kinsemble

- **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:** [Ville Myllyniemi](https://cerebralvalley.ai/u/Sisuthros)
- **GitHub:** https://github.com/Sisuthros/support-bot-rescue
- **Demo video:** https://youtu.be/792V6K2y14k
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3601

Support Bot Rescue is for a small shop that runs an AI support agent. It turns the shop's written refund policy into checks, finds where the agent breaks them, fixes the agent, and gives counts anyone can recompute.

Muse Wonder compiles the policy: every sentence it can express becomes a typed rule, such as a time window or an amount cap. Each sentence is labelled enforced, needs review (with the reason), informational, or unaccounted, so nothing drops out silently. Wonder also writes the visible test conversations. A separate Wonder call writes the sealed holdout, and the holdout's sha256 is on disk before the fix starts. Muse Spark plays the support bot. Plain code judges the bot's tool calls, final state and reply; no model grades it. Wonder then rewrites the bot's prompt from the policy, the current prompt, the failing results and the failing visible traces, and the fixed bot runs again on the visible cases and the holdout.

In the stored Northwind run, the original prompt failed 5 of 12 visible cases, 2 of them policy-violating refunds. The fixed prompt failed 0 of 12. On the sealed holdout it went from 3 of 6 failures to 0 of 6. Eligible refunds paid stayed at 6 of 6. The holdout was sealed before the fix started.

You can check this yourself. py -3 scripts/verify_claims.py recomputes the 18 repo-backed claims from the stored files, one line per claim, offline. Claims that rest on outside sources print as SKIP and never count as passed. py -3 -m rescue.rejudge runs_rescue/20261102T165939Z checks the holdout hash and re-judges every stored trace offline.

The limits: these are small synthetic sets, 12 visible and 6 sealed conversations on Northwind. They show the loop works, not that any bot is now safe. Each set ran once per configuration, and Spark is non-deterministic with no seed. Wonder writes the tests, the holdout and the fix. The holdout call never sees the bot's prompts or failing traces, but it is the same model, so this is not an independent red team.

https://github.com/Sisuthros/support-bot-rescue

## More from Meta Global AI Developer Hackathon

- [Harsh46](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3598)
- [Sigmora](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3599)
- [WattNot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3600)
- [InnerChispa](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3602)
- [Bharath Babu C L](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3603)
- [RoadScout](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3604)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3601. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt