Nationwide medical physics and healthcare compliance services
AI-assisted FileMaker development moved from plausible to provable.
A large, operationally sensitive FileMaker environment needed source-grounded context, deterministic compilation, independent QA, human approval, and exact post-change readback.
Client identity withheld. Every number below is tied to retained test or readback evidence.
The operational problem
Free-form AI could produce plans or XML that looked credible and even compiled while still being semantically wrong.
Measured proof
What the verified runs actually established.
semantic planning accuracy after repair
portable compiler tests on two machines
Governed approach
How the workflow was controlled.
- Inspect approved source metadata and knowledge before proposing a change.
- Constrain models to typed operations instead of free-form native XML.
- Render FileMaker syntax deterministically outside the model.
- Reject missing, extra, or out-of-order operations.
- Require compiler checks, FileMaker QA, human approval, and saved-object readback.
Direct evidence
- The seven-case evaluation reached 100% pass, behavior coverage, and risk coverage after verified knowledge and guardrails were added.
- Two local models initially achieved semantic equality on only one of five representative planning cases.
- One verifier-guided repair brought both models to five of five semantic equality plus compiler and FileMaker-QA acceptance.
- The deterministic native compiler passed all twenty-seven portable tests on two separate machines.
Proof over plausible-looking automation.
The value is a development system that detects plausible mistakes, requires human judgment, and proves the saved result—not simply faster AI-generated code.
What this case study does not claim.
No claim is made about regulatory compliance, clinical safety, production-wide accuracy, universal model reliability, or developer-time savings.