How to read a finding

Disclaimer

The research on LatentEval is published for education and independent verification. Each finding comes with its method, its sources and, where it applies, an honest account of the uncertainty around it. This page explains, in plain language, how to read the research and where its limits are before you rely on it.

Last updated: July 5, 2026

General terms

Research and education: read each finding with its method, its sources, and its limits.

The research on LatentEval is published for education and independent verification. Each finding reports behavior that was measured or published under specific conditions, with its method, its sources and, where it applies, an honest account of the uncertainty around it. A meta-analytic summary inherits the limitations of the studies it pools.

Agent reliability is a fast-moving field, and every measurement carries uncertainty. A finding here reflects the assumptions, harness, and model versions it was measured with, which may not match the system you have in mind. If you plan to rely on it, test and validate it against your own system, your own data, and your own eval runs first. This work is offered without any warranty of accuracy, completeness, or fitness for a particular purpose.

Findings carry dates and can be superseded as models, harnesses, and baselines change. The work here is meant to inform the judgment of anyone building, evaluating, or trying to understand agent systems. Treat it as one input to your own decisions.

Reuse and citation

The research, guides, and tools on LatentEval are our own work, and we hold the rights to them; where a finding rests on someone else's work, the source is named and linked. You are welcome to quote or cite any of it, in print, on the web, or inside AI-generated answers, as long as the quote is attributed to LatentEval with a link to the page it came from. Republishing a whole page, or enough of one to stand in for it, needs our permission first. For permission, corrections, or takedown requests, write to contact@latenteval.ai.