THE PROBLEM

Purchase pre-approvals in OPWDD Self-Direction (a Medicaid-audited NY disability program) require checking every request against category rules, fee caps, and evidence from vendor websites. Reviews are slow, manual, and audit-sensitive — and an AI assistant that hallucinates evidence would be worse than no assistant at all.

WHAT IT DOES

Pre-Approvals Reviewer automates the full review as a five-stage pipeline:

1. Vision-based PDF extraction (Claude) turns request documents into structured data

2. Category-routed YAML checklists — editable by non-engineers — decide what must be verified

3. Deterministic code enforces fee caps and eligibility rules

4. A browser research agent (Anthropic SDK tool-runner + Playwright) verifies claims on real vendor websites

5. HTML/JSON reports assemble the verdict with a date-stamped screenshot evidence folder

WHY IT'S DIFFERENT: EVIDENCE IT CAN'T FAKE

Integrity gates make fabricated evidence structurally impossible, not just unlikely:

• A "Found" verdict cannot be recorded without a real on-disk screenshot

• Quotes are rejected unless they appear verbatim on a visited page

• Timestamps, URLs, and SHA-256 hashes are written by code the model never touches

• Anything the web can't prove stays explicitly "Internal — not answered"

VALIDATION

18 deterministic tests, a rerunnable hash-manifest audit over every evidence package, human ground-truthing of negative findings, and planted-trap cases (an excluded laptop, an upheld denial) the agent caught. The find-fix-verify loop also surfaced and fixed a real prompt-labeling bug.

Open source, MIT licensed. Built solo, July 2026.

https://github.com/ebt55/preapproval-verification-challenge