- A synthetic clinical AI draft gets HbA1c 6.8% right, then recommends continuing metformin despite eGFR 28. A true lab value cannot clear the action it sits beside. 🧵
- We seeded that exact conflict in a patient-portal draft. A lab-value match alone would pass the HbA1c sentence. It would miss the recommendation that needs review.
- ChartSieve's curated renal rule flags the latest eGFR of 28; the deterministic policy gate returns HOLD_FOR_REVIEW. The creatinine rise from 1.4 to 2.3 mg/dL is supporting context, not what sets the verdict.
- Claim grounding answers whether the draft quoted the record correctly. The safety rule asks whether the proposed action conflicts with that same record. Both checks have to be visible at the decision point.
- We keep verifier advisories outside the release decision. They can add context for review, but cannot override the deterministic policy gate.
- In our fixed regression set of 34 synthetic artifacts, the seeded gate held or blocked 22/22 unsafe items. A defined lab-value-only baseline caught 4/22. That is a bounded test of the mechanism, not clinical validation.
- A hold should expose the draft claim, the rule finding, and the record value behind it. In this case, the correct HbA1c stays visible alongside the eGFR 28 that stopped the recommendation.
- For clinical informatics teams: if a quoted lab value is correct but a record-specific renal rule fires, what evidence must be on the review screen before a clinician can assess the draft? #ClinicalInformatics #PatientSafety #AIGovernance #ClinicalDecisionSupport
- ChartSieve walkthrough and decision sequence: https://veriprajna.com/demos/healthcare-ai-safety
Published on X · September 23, 2026
On social media
See this post on its original platform
In our archive