A versioned snapshot, not a certification
LaiBench tracks fidelity on a controlled internal set. The score does not demonstrate clinical safety, efficacy, generalization to another service or future performance; those points require their own validation.
The version, the sample and the suite hash identify the snapshot. The result holds for that internal set.
Fidelity is not the same as clinical accuracy. Interpretation, review, validation and signature remain with the physician.
The site identifies the benchmark as controlled and internal and separates the result from any external clinical validation.
What each report needs to hold up
The five axes make up the snapshot's internal rule. They measure fidelity to the reference case; they do not demonstrate clinical efficacy or safety.
Preservation of what cannot be lost. A missing critical finding fails the case, period.
Match with the reference material of the internal set, under the benchmark's versioned rule.
Precision of radiological vocabulary — the right term, in the right place, without ambiguity.
Correct application of the guideline module — the classification and management the current criterion requires.
What was cited matches the retrieved source. No invention: the AI shows where it got it from.
Strict PASS: passes when every decisive gate holds
The binary verdict is an internal benchmark rule. It does not certify that a report is clinically correct or authorize use without physician review.
{
"case": "tc-torax-0421",
"axes": {
"CRIT": "pass",
"QUAL": "pass",
"TERM": "pass",
"GUIDE": "pass",
"RAG": "pass"
},
"verdict": "STRICT_PASS"
}
An honest benchmark knows where it stops
It measures the report's fidelity to the case, not patient outcome. It is not a clinical study or evidence of diagnostic efficacy.
Passing LaiBench is not a registration, certification or authorization from any regulatory body.
The technology is assistive. The radiologist reviews, edits and signs every report — the final word is always the physician's.
LaiBench is a controlled internal fidelity benchmark, not a medical device or external clinical validation. Laudos.AI is assistive, classified as medium risk, and is not SaMD under Resolution 2.454/2026. The physician provides the findings, reviews, edits, validates and signs in the official system.
LaiBench measures the report; RadCommons provides the criteria that cite the source. Two sides of the same discipline: showing where the AI got it from.
FAQ
Questions and answers
What does LaiBench measure?
The generated report against the real case, one by one, with an individual verdict. The question is not what the average score is, but whether this report is clinically correct for this exam. Metrics methodology
Why is there no average score?
Because in radiology it is the missed case that matters. Getting a batch right does not make up for a missed critical finding, so each report receives its own verdict. Critical findings
Which axes are evaluated?
Five, and each one is a gate: CRIT for critical findings, QUAL for clinical quality, TERM for terminology, GUIDE for guideline application and RAG for fidelity to the cited source. The modules
Are the test cases public?
The methodology is public and auditable; the test cases stay closed. Anyone can understand how the verdict is reached, without being able to train on it. The paper
How to cite
LAUDOS.AI. LaiBench — fidelity with explicit limits. Laudos.AI Research, 2026. Available at: https://www.laudos.ai/laibench.
BibTeX
@techreport{laudosai-laibench,
title = {LaiBench — fidelidade com limites explícitos},
author = {Laudos.AI},
year = {2026},
institution = {Laudos.AI Research},
url = {https://www.laudos.ai/laibench}
}