Nineteen percent of the benchmark claims in the tensor.news record are independently reproduced, and every single one of them comes from Epoch AI. All 559 reproduced claims are Epoch's own evaluation runs. No second laboratory has independently reproduced any claim in the ledger.
That is not a defect in the record. It is what the record can currently prove, and it changes what the label means. A claim tagged independently reproduced has been run by Epoch — independent of the vendor — not confirmed by a second party. Every third-party leaderboard in the ledger, from the ARC Prize board to SimpleBench, DT Bench, WeirdML, and the LM Council, sits in the third-party tier — a named outside party ran it, but it is not an independent reproduction, because the record flags a score as independent only when Epoch produced it.
| Status | Claims | Share |
|---|---|---|
| Independently reproduced | 559 | 18.7% |
| Third-party measurement | 1,317 | 44.0% |
| Unverified | 734 | 24.5% |
| Self-reported | 381 | 12.7% |
| Contradicted | 2 | 0.1% |
| Total | 2,993 | 100% |
Unverified means the record asserts no provenance at all. A score published on the ARC Prize leaderboard is recorded as a third-party measurement because it names who ran it, and as unverified when nobody can be named — not because anyone doubts it. The split is the point: a named outside evaluation and an unattributed number are different epistemic states, and until now they shared one bucket.
In July the record held 1,943 claims, 573 of them independently reproduced — 29.5%. It now holds 2,993 claims and 559 reproduced ones — 18.7%. The ledger grew by 1,050 claims, 54%. The reproduced tier lost 14.
Two thirds of the claims sit on benchmarks where nothing has been independently reproduced. Of the 63 benchmarks with tracked claims, 51 have no reproduced claim at all, and they hold 2,074 of the 2,993 claims, 69.3%. The twelve benchmarks that do carry reproduced claims hold the other 919, led by GPQA diamond with 135 and OTIS Mock AIME with 126.
At the model level, 120 of the 265 models with tracked claims have no reproduced claim at all, 45.3%. Exactly one model's entire ledger is reproduced: GPT-5.3 Codex, with a single SWE-Bench verified claim.
The ledger holds exactly two contradicted claims, both on GPQA diamond, both from the same vendor report. Epoch measured DeepSeek-R1 at 64.88 and DeepSeek-V3 at 43.8; the DeepSeek-R1 paper, Table 4, reports 71.5 and 59.1. The gaps of 6.62 and 15.3 points both clear the 3.0-point threshold the record uses for a contradiction. Two findings out of 2,993 are a calibration point, not a pattern, and they say the vendor and the evaluator disagree — not which number is right.
It proves that the independent tier is thin, concentrated in one evaluator, and shrinking as the record grows. Anyone weighting benchmark scores for a decision should know that a reproduced tag means Epoch ran the evaluation, and that most of the record rests on scores no one in the record has independently rerun.