Evidence
An answer you can
sign your name to.
When the room is loud and the clock is running, LUCID returns the answer, the reason and the grade, fast enough to matter.
An answer, not an opinion
One graded recommendation for the case in front of you, in under five seconds, not a list to weigh.
Always shows its work
The guideline Class and level of evidence on every recommendation.
Forgets nothing
Every contraindication checked against 225 decision points, every time.
LUCID against three frontier language models.
The same 225 decision points from the 2026 AHA/ASA acute ischemic stroke guideline, given to LUCID and to three general-purpose models.
| Measure | LUCID | GPT-5 | Claude 4 Sonnet | Gemini 3 Pro |
|---|---|---|---|---|
| Overall concordance | 100% | 74.2% | 78.2% | 71.1% |
| Class III-Harm sensitivity | 100% | 58.8% | 70.6% | 44.1% |
| Class III-No Benefit sensitivity | 100% | 72.5% | 80.4% | 66.7% |
| 2026 guideline updates | 100% | 59.8% | 67.8% | 52.9% |
| Evidence-level (Class and LOE) accuracy | 100% | 47.3% | 52.8% | 41.2% |
Source: HORUS validation study, 225 decision points from the 2026 AHA/ASA acute ischemic stroke guideline. LUCID 100% (95% CI 98.3 to 100%); frontier models 71 to 78% overall. All differences P<0.001 against LUCID.
Every care stage, every decision point.
Concordance was 100% at each of the eight care stages LUCID covers.
How the benchmark worked.
- 225 decision points from the 2026 AHA/ASA acute ischemic stroke guideline.
- LUCID compared head to head with three frontier language models on the same cases.
- Measured: overall concordance, Class III sensitivity, 2026 updates and evidence-level accuracy.
- Confidence intervals and paired tests reported for every measure.
Next. A validation study is underway at two UPMC centers, with a real-world retrospective review in progress.
See LUCID on your own cases.
A 30-minute walkthrough with the team that built it.