ProofPack

What it is

Before a purchase from a self-directed, Medicaid-audited budget is approved at a New York disability-services nonprofit, a reviewer has to verify the provider's public website and file date-stamped evidence. ProofPack does the verification and files the evidence; the human still decides.

Five stages: a PDF form read into schema-validated fields, routed to category checklists written in YAML a non-engineer can edit, run through deterministic fee-cap and eligibility checks, handed to a browsing agent that navigates and captures, then assembled into a report with an evidence folder and a SHA-256 manifest. Exactly one stage is agentic.

Version 0.2.0 added a localhost reviewer workbench. It holds a queue of applications and finished packages, a live agent trail that streams each page opened, search made and capture taken during a review, and a package view that re-hashes every capture against the manifest on demand.

What I measured

Seven committed sample reviews, on synthetic forms covering six of the seven form types, cost $0.02–$0.19 each in Gemini spend. Negative cases were ground-truthed by hand: where a class page genuinely publishes no price, the correct output is "Not Found", and the tool refuses to guess one.

The integrity gates live on a session object, so offline tests cover every rejection path without an SDK, a browser or an API key. A "Found" cannot be recorded without a real capture in the manifest; a quote is rejected unless it appears verbatim on a page visited that session; timestamps, URLs and hashes are written only by code the model never touches.

Limitations

Unverifiable items stay marked "Internal — not answered" rather than being resolved by the model. The measured costs are the tool's own estimate from token counts at list prices. The engineering audit listing what would have to change before anyone relies on this is committed.