Building audit-grade AI agents in the open — github.com/ebt55
• ProofPack — pre-approval review agent for Medicaid-audited disability programs. Five-stage pipeline: PDF extraction → YAML checklists → deterministic eligibility checks → browser research agent → evidence report. Integrity gates make fabricated evidence structurally impossible: every "Found" cites a hashed screenshot; quotes must match page text verbatim. Gemini (main) + Claude Agent SDK branch; $0.02–0.19/review; 45 offline tests.
• IncidentGate (in development) — lab measuring how policy gates, an action monitor and single-use human-approval tokens change an incident agent under crashes & hostile input. LangGraph · FastMCP · Postgres. 434/434 kill-point recoveries, zero duplicate mutations.
• Digital Grimace Scale — preregistered study of adverse-treatment markers in LMs (Apart Research sprint). Primary test failed & published; a second frozen plan found an effect DPO couldn't remove.
• ExactDoc — PDF→editable DOCX
