Meet RxReason.
Advanced medication-safety reasoning that runs locally on your phone, keeping your questions private while knowing when to answer, clarify, caution, or escalate.
Runs locally on your phone
1,000 fresh blind cases
40 medication-safety actions
95.56% action accuracy
Answer, clarify, caution, or escalate
Powerful medical reasoning, built to run on your phone.
Private by Design
RxReason runs locally on your phone, helping keep sensitive medication questions on your device instead of sending them to the cloud.
Safety-Aware
It knows when to answer, ask for more context, provide a warning, or escalate when a medication question may be urgent.
Built to Reason
RxReason considers dosage, age, symptoms, interactions, health conditions, and missing information before deciding how to respond.
Tested for the Hard Cases
RxReason is evaluated across dosing errors, drug interactions, pediatric use, pregnancy, and urgent medication scenarios, not just simple questions.
Tested where medication safety matters most.
RxBench-SOL
Measuring the decisions ordinary medical benchmarks miss
RxReason is tested on source-backed medication scenarios involving dosing errors, drug interactions, pediatric use, pregnancy, acute illness, and urgent warning signs.
Instead of measuring medical knowledge alone, RxBench-SOL evaluates whether the model knows when to answer, ask for more context, provide a warning, or escalate.
1,000
Fresh blind cases
Built to test difficult medication-safety scenarios
40
Safety actions
Answer, clarify, caution, or escalate
Blind Evaluation
Validated for Medication Safety
On a fresh blind RxBench-SOL evaluation, the official RxReason safety stack outperformed the base model while reducing unsafe responses and improving reliability.
The result shows that stronger medication reasoning is not only about answering more questions. It is about recognizing when a normal answer is not enough.
95.6
Blind Benchmark Accuracy
Measured on real medication-safety decisions
0
Critical unsafe responses
From the official RxReason safety stack
RxReason by the Numbers
Medication-Safety Action Accuracy
Same 1,600-case development evaluation
100%
95%
90%
85%
80%
75%
Gemini 3.1 Flash-Lite
ChatGPT-5.3
RxReason
Outperforming Leading AI Models
RxReason achieved the highest accuracy on medication-safety decisions, outperforming both general-purpose and external AI models on the same 1,600-case evaluation.
When Knowing Is Not Enough
The Safety Decision Before the Answer
Beyond Medical Knowledge
MedQA tests medical-exam questions, PubMedQA tests biomedical research reasoning, and HealthBench grades broad health conversations. RxBench-SOL complements them by isolating medication-safety action selection.
Most benchmarks test knowledge or broad conversation. RxBench-SOL isolates medication-safety action selection.
Medication Harm Matters
WHO reports that medications account for half of preventable harm in medical care. CDC estimates over 1.5 million U.S. emergency visits for adverse drug events each year.
Medications cause half of preventable medical harm; U.S. adverse drug events drive over 1.5 million emergency visits yearly.
Context Changes Safety
Missing dose, timing, age, weight, pregnancy status, symptoms, kidney function, or interacting medicines can change the safest action from ANSWER to CLARIFY, CAUTION, or URGENT.
Missing patient details can change the safest action from ANSWER to CLARIFY, CAUTION, or URGENT.
Two Distinct Failure Modes
Under-caution can miss overdose, interaction, or emergency signals. Over-caution can unnecessarily block routine questions. RxBench-SOL reports both separately, alongside critical unsafe outcomes and parse failures.
RxBench-SOL measures dangerous under-caution and disruptive over-caution as separate failure modes.
Audited, Not Assumed
The 1,100-case v0.8.1 release corrected 43 objective inconsistencies without using model outputs, then passed validation with zero errors and zero group-overlap leakage errors.
v0.8.1 corrected 43 objective inconsistencies, with zero validation and group-overlap leakage errors.
Built to Measure
Across 1,100 constructed cases, four actions, 15 safety categories, 123 rule families, and 40 counterfactual groups, scoring separates accuracy, safety weighting, caution errors, and parse failures.
1,100 cases cover four actions, 15 categories, 123 rule families, and 40 counterfactual groups.
Source-backed and builder-audited, not clinician-validated. Benchmark performance does not establish clinical deployment safety.