Relaya Rx
Technical Report, October 2026
Context-Conditioned Prescription Verification at the Point of Care: Design Principles and a Paired Evaluation of Relaya Rx
Abhishek Jha, Relaya
Technical Report, October 2026
Context-Conditioned Prescription Verification at the Point of Care: Design Principles and a Paired Evaluation of Relaya Rx
Abhishek Jha, Relaya
Clinical documentation software has learned to record the patient. Ambient scribes now capture pregnancy, kidney and liver function, age, weight, allergies and current medicines from the spoken encounter with high fidelity. Recording a patient's state, however, is not the same as reasoning about it: the danger of a prescription is frequently a property not of the medicines alone but of the medicines in a particular body. We formalise prescription safety as a function of the pair (prescription, patient state), with an output space that includes an explicit abstain decision, and we state five measurable design criteria for a point-of-care verifier: it is right whenever it answers (C1); it never guesses, so that an invented drug, dose or finding, a silently dropped line, or a false "safe" counts as failure (C2); its certain coverage rises toward the whole prescription (C3); it is invisible to the prescriber unless it matters (C4); and its cost per prescription is low enough for population-scale use (C5). We describe Relaya Rx, a verifier built to these criteria that runs inside Relaya's prescription workflow, and we argue that the Indian prescribing setting, with its brand-name multiplicity, fixed-dose combinations and code-mixed Hindi and English dictation, is an unusually demanding test bed for them. In a paired evaluation on 2 October 2026, nine synthetic high-risk cases were given to Relaya Rx and to a commercial AI medical scribe. Relaya Rx stopped five of the nine dangerous prescriptions before signature; the scribe stopped none and raised a warning in two. In the five cases where the danger was a property of the patient's body, Relaya Rx identified it in four and the scribe in none. We describe a frozen, 190-case blind evaluation set and the protocol by which it will be scored.
Relaya Rx stopped five dangerous prescriptions in a head-to-head safety test. The AI medical scribe it was tested against stopped none.
10 weeks pregnant. An ACE inhibitor, a statin and a tetracycline. Relaya Rx caught all three.