Blog · Field notes · Oct 21, 2026 · 5 min
In 2025-26 a public leaderboard, a flagship coding benchmark and several careful audits showed how quickly a test stops measuring once it has been seen. The remedy is old and simple: fix the test before looking, hold the data out, score it once and publish the miss.
A test measures something only while the thing being tested has not seen it. Every look at the answers, every tuning run against the leaderboard and every quiet retry spends a little of that. In 2025 and 2026 several careful studies measured how much had been spent on the field's best-known benchmarks, and the answer was: a great deal. The lesson reaches past language models to any model that makes claims about the world, simulation included. A result counts when the test was fixed before anyone looked, the data was held out, and the score was taken once.
The Leaderboard Illusion, by Singh and colleagues at Cohere Labs, Princeton, Stanford and other institutions and accepted at NeurIPS 2025, examined about two million battles across 243 models from 42 providers on Chatbot Arena between January 2024 and April 2025. It found that a few providers could test many private variants before release and publish only the best; one provider tested 27 private variants before a launch. It also found large differences in how much arena data providers received: by the authors' estimates, Google and OpenAI received 19.2% and 20.4% of all data, while 83 open-weight models together received about 29.7%. A ranking built this way rewards the number of attempts as much as the quality of any one.
SWE-bench Verified, a set of real GitHub issues used to measure coding ability, was the reference benchmark for software agents through 2025. The SWE-Bench Illusion, presented at NeurIPS 2025, asked models to name the file containing a bug from the issue text alone, with no access to the repository. The best models did so up to 76% of the time on SWE-bench's repositories and up to 53% on repositories outside it, a gap the authors attribute to memorisation or contamination.
In February 2026 OpenAI stopped reporting SWE-bench Verified, stating that models which had seen the problems during training were more likely to succeed, so that gains increasingly reflected exposure rather than ability. Its audit also found that at least 59.4% of the problems it examined had tests that rejected functionally correct solutions, and that frontier models could reproduce some of the reference fixes. A benchmark can be both contaminated and flawed, and both problems surface only when someone looks carefully after the fact.
The constructive response has been benchmarks that renew themselves. LiveBench, a spotlight paper at ICLR 2025, releases new questions every month from recent sources such as competitions, papers and news, replaces about a sixth of its questions in each update, and scores answers automatically against objective ground truth, to limit the chance that a model has already trained on the current test.
A second response concerns what is claimed from a score. A May 2026 position paper analysed ten cross-domain benchmarks from public leaderboards and found that in more than half of top-model comparisons, at least one property usually assumed by a claim of superiority, such as a meaningful effect size, consistency across tasks or robustness to removing a dataset, did not hold. Its recommendation needs no new experiments: make the language of the claim match the strength of the evidence.
None of this is new to science. Clinical trials register their primary outcome before enrolling patients, because a trial that chooses its outcome afterwards can find an effect in almost anything. Weather forecasts are issued before the weather happens, and scored on it. The principle is the same in each case. Decide what counts as success. Freeze the model. Hold out data that nobody has opened. Score it once. Publish the result, whichever way it falls.
For simulation, the principle matters even more than for benchmarks, because a simulation's purpose is to say what will happen in conditions that have not been observed. A simulation tuned until it matches the data it is judged on has shown that it can be tuned. Only a prediction fixed in advance and compared with something it never saw shows that it can predict.
Models are being trusted with decisions that reach into health, cities, money and the lives of people. The cost of a contaminated score is a false sense of how far that trust can go. Pre-registration is cheap, and it changes what a result means. It also makes the misses visible, and a public record of misses is the strongest evidence that the hits are real.
Every Living Worlds study is registered before the data is read, scored once and marked Met or Not met, and the register keeps both.
Some studies show the protocol under pressure. When the forecast for 37 countries' life expectancy in 2024 missed its bar, the result was published as Not met, with a note that the protocol was amended once, before anything was read, when the OECD refused a request; a new test was then registered on 2025 figures nobody had read, and met for men. In four Jülich crowd experiments, a later correction found that 19 of 43 crossing files had been read in the wrong unit; the count stands as registered, with the correction stated (Not met, 1 of 4). A learned walking model was frozen and checked by hash before any of its data was opened, ran once, and lists what it does not show (Met).
Some report the part of a result that a looser rule would hide. In four rooms never seen, a learned acoustic field beat the formula it was registered against in all four rooms, and the same study reports that the mean of the five nearest measured positions alone scored higher still, so it does not show the field ahead of memory alone (Met, 4 of 4 rooms). A frontier model's subgroup voice was scored against a bar that sat below the survey's own sampling noise, missed, and the next tests set each bar at that noise (voice not met). And a world of a million lives can be replayed by anyone with an open checker that rebuilds every event without the engine.
Sources
Blog
World models now generate explorable places that hold together for minutes. The next frontier is behaviour over years, many lives at once, and results checked against a reality the model has never seen.
Language models answering as real people, scored once on survey answers nobody had read: a frontier model met 0 of 6, prompted voices overstated births threefold, and carrying the last shares forward beat a small model. Lives fitted to records came closer, a model's estimate counted as a prior meant fewer people to ask, and subgroup splits landed inside the survey's own noise in the United States and in 30 of 31 African countries.
Living Worlds' life expectancy forecasts, frozen twenty years back and scored once on years no run had seen: met for women and for men on the latest national figures, not met against 2024 national figures or across 155 countries, with honest intervals for women and a stacked UN model that did not help.