Direct PT/DX × Bowers Enterprises · Phase 0 Validation

The LLM Vision Shootout

The assumption we're testing

Vision AI can reliably read clinical documents across every format Direct PT/DX receives.

Names, dates of birth, dates of service, CPT codes, unit counts — pulled from real scans and faxes. This one assumption carries the triage and reconciliation workflows.
Four models, head to head
Claude Sonnet 4.6 Claude Opus 4.8 GPT-5.5 Gemini 3.1 Pro
The plan, in order
01 Label the answer key Debbie's team · ~2–3 hrs
02 Build the test harness Chase
03 Run every document through every model Same prompt, no favorites
04 Score accuracy, cost, and failure modes Against the answer key
05 Deliver the shootout chart Go / no-go + model pick
What passing looks like
≥80%
Document-level accuracy on top-tier providers
≤$0.03
Cost per document at the winning model
Predictable
Failures we can route to human review