When your AI fails, can you explain what happened?
A workflow and reference kit for AI agent builders who refuse to let "the AI said so" replace human accountability.
I'm a clinician.
For years, I made calls when proof wasn't fully there — because someone had to decide, and the body in front of me couldn't wait for certainty.
That work taught me what reliability really means: not perfect knowledge, but a chain of evidence that lets someone stand up afterwards and say: "this is on me."
I see AI agents being shipped without that chain. "It works on my machine" replacing actual proof. "The AI said so" replacing human accountability.
This kit is for builders who refuse to let AI become a place where responsibility quietly disappears.
When something fails, someone has to be able to say: "this is on me."
Proof-first is how we keep that sentence possible.
WHY_PROOF_FIRST.md— the principle behind this kitAUDIT.md— paid Mini Audit service ($300, fixed scope)examples/agent_audit_checklist.md— what to check before shippingexamples/deterministic_eval_flow.md— how to test agents without trusting themexamples/rollback_decision_log.md— when and how to roll back
- Solo builders shipping AI to real users
- CTOs explaining AI reliability to investors and customers
- Tool makers building common accountability standards
A clinician from a regulated field, building proof-first systems for AI. Half-anonymous on purpose. Text-only communication.
MIT — use it, fork it, ship better systems.