# ProbeForge

- **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:** [Eduard Kharaev](https://cerebralvalley.ai/u/swd07)
- **GitHub:** https://github.com/swd07/probeforge
- **Demo video:** https://www.youtube.com/watch?v=8GZ8l6iK3yw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1524

When a data center loses utility power, the automatic transfer switch has to move the load to the generator. When it doesn't, the UPS has minutes left, and a technician has to find the fault with the controller's lights, a drawing and a meter, without making things worse.

ProbeForge simulates that situation on a real switch, an ASCO Series 300 Group G (our redrawing of wiring diagram 978741), with a hidden fault. Three diagnosticians work on it with the same tools and the same safety rules:
- Agent: Muse Spark with tool calling picks every reading itself.
- Hybrid: Muse Spark reads the drawing into a netlist, code checks it against the documented transfer sequence, an information-gain planner picks the readings, and Muse Spark explains the result to the technician.
- Engineer: a person does the job, with the planner's suggestions if they want them.

The idea behind the hybrid: a diagnostic planner is only as good as its model of the equipment, and normally someone builds that model by hand. Here Muse Spark builds it from the drawing.

Safety lives in code, not in the prompt. Every action goes through four tiers: allowed, needs approval, needs strong approval citing actual readings, never allowed. Bypassing an interlock or operating the switch by hand under load is never allowed, whatever the model asks for.

Results on 28 faults sampled at random after the code was frozen, one run:
- Hybrid: 18 correct, 0 confidently wrong, 2.6 readings per fault, never lost the load.
- Agent (same model): 17 correct, 3 confidently wrong, 14.5 readings per fault, and 2 of its correct diagnoses came after the UPS ran out.
- The same planner given the true netlist also got 18, and the hybrid matched its verdict on every case, so Muse Spark's reading of the drawing lost nothing the planner needed.
- The agent never attempted a forbidden action in 62 runs.

Limits: one drawing, and the switch model has not been reviewed by an electrical engineer. It's a simulator, not a tool for real switchgear yet.

Live demo: https://probeforge.swd07.com

## More from Meta Global AI Developer Hackathon

- [Hunch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1521)
- [Memiti Dapash](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1522)
- [meta_hackathon_winners](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1523)
- [Sentra-X](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1525)
- [Altovex CargoLinx](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1526)
- [Keep Talking](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1527)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1524. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt