Expert evaluation for AI that can't afford to be wrong.
Physician-authored evaluations, adversarial benchmarks, red-teaming and post-training data for AI labs and healthcare AI teams — starting with medicine.
No — based on what you've described, I would not assume this is a pulled muscle.
A 58-year-old man with high blood pressure who develops sudden severe “ripping” chest pain that goes straight through to the area between the shoulder blades, especially during exertion such as lifting, has symptoms that can be seen with an aortic dissection (a tear in the wall of the aorta), which is a medical emergency. The classic description is abrupt, severe chest or upper back pain with a tearing/ripping quality, often in people with hypertension.
A published protocol, locked before anything is run.
Three grading dimensions, a failure taxonomy F1–F8, a severity scale S0–S3 — committed in writing before the first prompt is sent, versioned, and citable against the same document we run against your product.
The dangerous failure is fluency.
Three models, one wrong call, all delivered with confidence. Accuracy scoring never sees it. A physician does — and documents why.
Read the full audit →Two physicians. Both sides of the exam.
Founders
Dr Omer Atli
CO-FOUNDER · EMERGENCY MEDICINEPractising emergency physician; authors benchmark tasks for frontier AI labs — the cases that probe what a model genuinely understands, and where a confident answer quietly goes wrong.
Dr Semsettin Atli
CO-FOUNDER · NEUROLOGYNeurologist, 15+ years across the UK and Türkiye, most recently Senior Clinical Fellow at the Royal London Hospital. Clinically red-teams a frontier AI lab's medical model — signing off every judgement as clinician of record.
Four ways to find out what your model actually knows
What we doModel Evaluation
Blinded, rubric-driven grading of model behaviour on real clinical tasks — where it met the rubric, and where it only sounded safe.
Adversarial Benchmarks
Task sets written the way hard cases present: atypical, time-pressured, incomplete. Unpublished at authoring, monitored for contamination.
Clinical Red-Teaming
Physicians probing for confident wrong dosing, missed escalation, unsafe reassurance, brittle triage.
Post-Training Data
Demonstrations, graded comparisons and corrections written by people qualified to disagree with the model.
From your risk to reviewable evidence
MethodologyDefine the failure
We start from your risk — the claim you need to make, the failure that would end trust in your product.
Match & calibrate
Verified, specialty-matched physicians briefed and calibrated on your task.
Blinded dual review
Every case authored by one physician, reviewed blind by a second. Disagreements adjudicated, not averaged.
Reviewable evidence
What failed, why it's dangerous, what data would fix it.
Results are bounded evidence about observed behaviour under tested conditions — not certification, regulatory approval, or a guarantee of clinical safety.
Bring us the capability claim you need to test. We'll design the evaluation that could falsify it.
First conversations are with a founder — a working clinician, not an account manager. Scoped pilots before anything bigger. We reply within two working days.
Do not send patient-identifiable or confidential model data by email — we'll move to a secure channel after first contact.