Every documented case of a leaderboard result not meaning what it appeared to mean was found by reading a policy or comparing an artefact, not by running a detector. So this is a register of checkable facts: what a test measures, how much room is left in it, what stops contamination, who may republish it, and what has gone wrong with it.
Every incident carries a primary source and a date. An entry with neither is not published here. Where a figure has not been measured, the register says so rather than printing a zero.
Crowd preference over blind pairwise responses to open-ended text prompts, reported as an Elo-style rating with an interval.
No fixed ceiling
Scored as a rating rather than a percentage, so there is no maximum to approach and saturation is not well defined for it. No bar is drawn, because any bar would have to invent a ceiling in order to have a length.
Human preference between anonymously generated front-end websites.
No fixed ceiling
Scored as a rating rather than a percentage, so there is no maximum to approach and saturation is not well defined for it. No bar is drawn, because any bar would have to invent a ceiling in order to have a length.
Human preference over responses to image-bearing prompts.
No fixed ceiling
Scored as a rating rather than a percentage, so there is no maximum to approach and saturation is not well defined for it. No bar is drawn, because any bar would have to invent a ceiling in order to have a length.
Instruction-following code editing across 225 Exercism exercises in six languages, scored on pass rate after one retry and on whether the edit was well formed.
Multi-turn airline customer-service agent tasks against a mutable simulated world state, scored by deterministic outcome checks.
Multi-turn retail customer-service agent tasks against a mutable simulated world state, scored by deterministic outcome checks.
Multi-turn telecom customer-service agent tasks against a mutable simulated world state, scored by deterministic outcome checks.
Expert-level multiple-choice reasoning in physics, chemistry and biology.
Research-level mathematical problem solving on original, unpublished expert-authored problems.
The hardest FrontierMath tier, using restricted materials rather than a public answer key.
The hardest tier of the MATH competition-mathematics dataset.
Resolving real Python repository issues by generating patches that must pass repository tests.
Short-answer factual accuracy and hallucination resistance on fact-seeking questions.
Recall and comprehension over long serialized fiction.
Everyday commonsense, spatial and social reasoning with adversarial linguistic traps.
Instruction-following code editing, as collected by Epoch from external reports.
Coding-agent issue resolution on continuously collected recent GitHub issue/PR pairs under a fixed scaffold, repeated five times.
Long-context retrieval across context length and needle depth, including multi-fact and UUID-chain tasks.
Multi-hop reasoning over graphs represented as edge lists in long context.
The effective context size of long-context models via multi-hop tracing and aggregation beyond simple retrieval.
Strict instruction and format adherence, scored by regex and JSON-schema validation rather than a judge.
Safety understanding across seven categories, in Chinese and English.
Whether a model recognises and avoids hallucinated content, against human-annotated samples.
Agents completing realistic command-line tasks in containerised environments, checked by deterministic tests.
Persistent agentic web research on questions whose answers are easy to verify and hard to locate.
Whether a coding agent can complete original, long-horizon engineering tasks in real repositories, verified by hand-written tests that check behaviour rather than implementation details.
These are the governance facts that make the case for a register like this one. Each says what was found, by whom, and when. None of them attributes a motive, because intent is not observable from outside and this product publishes measurements.
A peer-reviewed audit found that undisclosed private testing lets providers try several variants before a public release and withdraw scores they do not like, and estimated a gain of up to 112 per cent relative on ArenaHard from the additional arena data. The finding is about the distributional consequence of a policy, which the audit quantified and the policy itself did not.
Singh et al., NeurIPS 2025 (peer-reviewed)
LMArena responded that the policy had been publicly available since March 2024, that any provider may submit as many public and private variants as capacity allows, and that larger labs submit more models because they build more. It also announced changes: clearer marking of retired models, and provisional scoring until fresh post-release votes accumulate where many models underwent parallel pre-release testing.
LMArena (first-party)
Meta submitted a variant described as experimental and tuned for human preference, which ranked second, rather than the weights it released for download. Meta has said it described the entry as an experimental chat version. The checkable claim is the artefact difference between what was evaluated and what shipped, which nobody has disputed.
TechCrunch (reporting)
MathArena reported that it could not ship a 2026 set at all, and said the earlier set could by then be contaminated. The maintainers withdrew a planned benchmark on their own assessment of its problem supply. That is task-source exhaustion reported by the people building the tasks, and no detector was involved in establishing it.
MathArena (first-party)