SkillForge
An AI fix should earn the right to run again. SkillForge turns a verified, human-approved repair into a versioned skill that handles the next compatible incident without calling the model.
01 / THE PROBLEM
Repeated problems should not require repeated improvisation.
The model proposes a bounded repair. Independent code verifies the result, tests the generated skill in isolation, and runs hidden cases before a person can promote it. Reuse checks the incident fingerprint and the pinned skill files before executing known code.
02 / THE SYSTEM
From input to a checked result.
Python · Qwen · vLLM · GB10 · React · TypeScript
03 / EVIDENCE & LIMITS
What has been demonstrated.
On GB10, a model-authored skill passed 9/9 hidden checks; after host approval, a compatible incident reused it with zero model calls. Incompatible lookalikes were refused.
All demonstrated incidents use synthetic data. Generation remains unreliable: the documented thinking-on evaluation passed 3 of 7 attempts. The hosted walkthrough is a recorded simulation, not a live model or a connection to the GB10. Full Telegram-to-OnCall execution has not passed its integration gate.