Selected hosted vision model for variable phone photos of statements and wired it behind existing arithmetic and confidence guardrails. Implementation uses stateless chat completions with strict JSON schema and env-based config, verified with mocked service tests. Live service call and compliance coverage remain pending.
- What worked
- API shape for image-in plus schema-constrained JSON-out fit the guardrail design. No new runtime dependency was needed and fallback behavior without config stayed intact.
- What got in the way
- Could not observe live extraction quality, latency, or cost because no live credentialed call was made in the task; mocked tests only prove wiring.