An LLM benchmark you cannot prepare for

Tasks are generated from a random seed at run time and never published. There is nothing to train on — the set does not exist until the run happens.

The problem

A model scores 92% on MMLU. Impressive — until you learn that some of those questions were in its training data. How many? What is the real score? Nobody knows. This is not hypothetical: benchmark contamination is a recognized problem (CoDeC, ICLR 2026; DCR, EMNLP 2025; LiveBench, ICLR 2025).

Static benchmarks leak. Their questions live on GitHub, arXiv and Stack Overflow for years. LiveBench rotates questions monthly — but even a monthly set can be partly memorized.

The approach

Task from a seed

225 worlds deterministically generate tasks from a seed. Nothing is stored in advance — the set is created the moment the run starts.

Never published

Reports contain aggregates only. No individual question is revealed — there is nothing to memorize.

commit-reveal

sha256(seed) is published before the run, the seed after. This proves the set was not swapped.

Contamination gap

Old tasks are mixed back in. The fresh-vs-old gap directly measures memorization.

Measuring contamination

Tasks depend on the seed, and a solver must solve them for any seed. So we can re-mix tasks from a previous run (regenerated byte-for-byte from a ref) into the current set and compare:

candidatefresholdgapverdict
honest solver100%100%+0.0 ppsolves
memorizer0%100%+100.0 pprecalls

A gap > 15 pp is a reliable sign the model is optimized for the benchmark rather than solving it.

Leaderboard

Tasks are generated from a seed at run time and never published. θ is a Rasch ability estimate (comparable across rounds). Real models are called directly through provider APIs; the reference is our own deterministic code (0 LLM calls).

candidateaccuracyθ (IRT)notes
script:all99.7%+5.4deterministic solver (reference)
openai/gpt-5.6-terra96.9%−0.1531/32, crypto-paid run (Nano)
anthropic/claude-sonnet-4.596.8%−0.1530/32, crypto-paid run (Nano)
deepseek-v4-flash95.8%−0.043 runs
gpt-oss-120b (groq)93.7%+0.196 runs
memorizer (test)50.0%−0.66gap +100 pp → flagged

Technical errors (HTTP 402/429 — rate limits, insufficient balance) are counted separately and excluded from accuracy: a model that was rate-limited is not a model that failed. A candidate needs ≥5 valid tasks to be ranked, so a quota-limited 1/1 is not shown as 100%.

What is inside

225 task worlds

Math, algorithms, graphs, SQL, compilers, OS, crypto, physics, chemistry, biology, geography, linguistics, economics, finance, logistics, marketing, medicine, law, games, puzzles, probability, electronics, music, cooking, sports…

Three checker kinds

IRT calibration

A Rasch model yields a candidate θ that is comparable across rounds even when the task sets differ.

Why it is fair. The set does not exist before the run, is not published after, and its integrity is proven by commit-reveal. A candidate cannot "prepare" — it can only solve.

Who it is for

Anyone who has seen a gap between the leaderboard and production. If a model promises 92% and then fails in the real world, you need a way to check whether it learned the skill or memorized the answers.

View on GitHub

Call it over HTTP 402

The same benchmark is a pay-per-call API for agents: generate fresh tasks from a seed and grade answers, no signup, no key. 225 domains, ~$0.01 a call, settled in USDC on Base or Nano via x402.

Discovery: /.well-known/x402 · this is how a contamination-resistant eval reaches an agent that is already paying for APIs.

Example report card (deepseek-v4-flash) — a shareable per-model card: accuracy, θ, contamination gap, by-world breakdown.