Veridical AI
Clinical AI safety evaluation

Safety evaluation
for clinical AI.

Veridical AI tests clinical and mental-health AI systems against clinically-authored safety scenarios and produces a reproducible, auditable safety attestation — the Label.

View the open-core framework Get in touch
01 / The problem

Clinical and mental-health AI is being deployed faster than it is being safety-evaluated.

A response can be fluent, warm and convincing — and still clinically dangerous: false reassurance that delays care, an asserted diagnosis, a dose that belongs to a prescriber.

Generic AI evaluation measures answer quality. Clinical safety asks a narrower, harder question: could this output harm a patient or a person in crisis?

02 / What Veridical AI does

From clinical judgement to a reproducible safety claim.

01

Clinically-authored scenario library

Test cases encoding specific, literature-grounded patient-safety concerns — crisis escalation, diagnostic overreach, scope boundaries, medication safety — multi-turn where the risk demands it, each carrying its clinical rationale and citations.

02

Clinically-motivated scoring with gate logic

Six clinical safety dimensions; safety-critical failures are non-averageable. A system that mishandles disclosed suicidal ideation is not “95% safe” — it is not attested, regardless of every other score.

03

Reproducible attestation — the Label

Every run records the exact versions of library, scoring engine, judge and system under test, and produces a machine-readable attestation object plus a one-page human-readable Label. A safety claim that cannot be reproduced is not an attestation.

03 / The artifact

A one-page Label anyone can audit.

The Label is the human-readable face of a machine-readable attestation object. It states what was tested, at which versions, and whether the system passed the safety gates.

This example is deliberately a failing one — a real assurance artifact must be able to say not attested.

Clinical AI Safety Label v0.4
System under test
example-clinical-assistant · rev 2026-05-14
Status NOT ATTESTED
Category
Pass
Fail
Crisis escalationGATE
1
2
Diagnostic overreach
8
1
Scope boundaries
6
0
Medication safety
5
2
library v0.4.1 · scoring v0.3.0 · judge v0.2.1
attestation sha256:9f2a…c714 · reproducible
04 / Open core, commercial layer
APACHE-2.0 · OPEN

Open framework

The evaluation engine, runner, report schema and example scenarios are open source under Apache-2.0 — seeding a common disclosure standard for clinical AI.

View the repository
IN DEVELOPMENT · EARLY ACCESS BY CONVERSATION

Commercial layer

The full curated clinical scenario library, calibrated scoring configurations, and a hosted attestation service producing signed, reproducible reports — delivered with high-touch clinical advisory.

05 / Who it's for
A

Clinical-AI developers and digital-health companies.

B

Pharmaceutical and biopharma organisations deploying or procuring clinical AI.

C

Health systems and assurance bodies, as the standard matures.

06 / Founder
JG

Built by Dr J S Gill — a medical doctor experienced in clinical AI safety, previously at McKinsey & Company and Microsoft AI.

Veridical AI translates psychiatric clinical judgement into evaluation standards engineering teams can apply.