AI Safety · Technical Analysis

Evaluating Frontier Models: A Practical Guide

How institutions and technical teams should test large, general-purpose AI models — rigorously, adversarially, and continuously — before and after they are deployed.

Published 2026For Institutions & PractitionersPublisher: FAIR Labs, Washington, DC
Download the PDFAll reports

Listen

Evaluating Frontier Models: Rigor, Risk, and the Instrument Panel

A ~14-minute audio interview about the report with Dr. Krzysztof Pietroszek, President of FAIR Labs. Produced with www.allais.com.

Summary

The bottom line for decision-makers

You cannot manage what you have not measured, and you cannot trust a measurement you did not design to resist gaming. Evaluating a frontier model well means asking the right questions—what can it do, where does it fail, can it be pushed into harm, does it know what it does not know—and answering them with methods robust to contamination, to spec-gaming, and to the model's own fluency. This report supplies a taxonomy for organizing those questions, protocols for answering them adversarially, and a pipeline for turning the answers into a defensible deploy-or-hold decision that is then re-opened, on schedule, for as long as the model is in service.

How institutions and technical teams should test large, general-purpose AI models — rigorously, adversarially, and continuously — before and after they are deployed.

Key findings

  1. Measure constructs, not scores. Decide what real-world capability or hazard a test is meant to stand for, then defend the claim that the test measures it. A number without a validated construct behind it is theater.
  2. Assume the benchmark is contaminated. Treat public benchmark scores as an upper bound corrupted by training-data leakage. Weight held-out, private, and freshly authored evaluations far more heavily in any decision that matters.

How to cite this report

Evaluating Frontier Models: A Practical Guide. FAIR Labs Technical Analysis. Fair Artificial Intelligence Research Labs, 2026. Available at https://fairlabs.ai/research/evaluating-frontier-models

FAIR Labs (Fair Artificial Intelligence Research Labs) is a nonpartisan 501(c)(3) research institute in Washington, DC. Its reports may be reproduced with attribution for non-commercial purposes. For interviews, briefings or data requests, use the contact form.

Related FAIR Labs reports

AI Safety Framework for Public Institutions2026 · Policy Report · PDF