An AI care-plan verifier called CITE caught 5 of 17 checkable errors in test cases, but clinicians acted on only 1 of 9 valid live flags
Synopsis
Waymark tested a verifier called CITE in Medicaid primary care: a second language model reads an AI-drafted care plan before the physician signs it and flags wrong recommendations, statements the record does not support, and missing care. In retrospective testing CITE caught 5 of 17 checkable errors at its primary threshold with 97.3% specificity; in a pre-registered randomized trial it raised no false alarms at that primary threshold, physicians judged 9 of 11 flags valid, but the signed notes acted on only 1, and safety events were similar across arms.
Interpretation
AI-drafted care plans now get an independent check before signing: CITE, a second language model, reads the draft and lists problems beside it, leaving the decision to the physician. No large randomized trial in live care had tested such a verifier; the closest prospective study was a single-arm study of 98 patients, while health AI governance groups ask for testing where the tools are used. On 447 authored Medicaid test cases, three models each wrote one plan for 1,341 plans total, and CITE caught 5 of 17 checkable errors at its primary threshold (29.4%, 95% CI 10.3 to 56.0), with specificity of 1,288/1,324 (97.3%).
The verifier added little alert burden in live care: across 77 encounters CITE checked, it raised 11 flags in 7 encounters, a mean of 0.14 per encounter, under the pre-registered ceiling of a median of 3. Clinicians override 46.2% to 96.2% of decision-support alerts and repeated alerts lower acceptance, so flag volume is itself a design constraint. Physician adjudicators judged 9 of the 11 flags clinically valid (81.8%, 95% CI 48.2 to 97.7), all 5 unsupported-statement flags and 4 of 6 missing-care flags; the signed notes acted on only 1 of the 9 valid flags.
Guardrailed AI drafts were similar to unassisted physicians on safety events, the first question health systems ask about whether a model can draft for physician review without adding harm. The trial randomized 240 patients to unassisted physicians, the guardrailed assistant, and the guardrailed assistant plus CITE, a comparison previously missing in live care. The safety composite occurred in 4/79 (5.1%), 4/82 (4.9%) and 5/79 (6.3%) encounters across the three arms; over-escalation for guardrails plus CITE minus guardrails alone was −1.2 percentage points (95% CI −6.6 to 3.5), with the upper bound below the pre-specified 5-point margin.
Perspective
The demonstrated setting is Waymark's multidisciplinary primary care service for Medicaid managed-care members in Ohio, Virginia, Washington and Illinois, with 240 patients each having one encounter and every live draft written by Gemini 3.1 Pro under six-part safety guardrails; the retrospective part used 447 authored test cases.
Checkable errors were far rarer in live drafts than in test cases (1 of 161, 0.6%, versus 16 of 447, 3.6%, for the same model), and that one draft was in the guardrails arm CITE did not check, so the trial could not estimate CITE's live sensitivity; the trial was underpowered for its main question, with post hoc power only for a composite increase of about 13.7 points; it did not record whether clinicians viewed, accepted or overrode each flag, so action was measured from the signed note; and in retrospective testing Claude Opus 4.8 drafted the ratings and was also the verifier's model and one of the rated models.
