2025-05-30 · Sakana AI
Replications
We found no full replication of any of the 11 claims with a measured result. One has been partly rerun by another group. For each claim, whether someone other than the people who made it has run it again.
Partly rerun 1
Someone else reran part of it, or reran it under different conditions.
No rerun found 6
Others could rerun it, but we found no one who has.
2026-10-12 · Sakana AI
MASS
The code is public. We found no rerun by another group.
2026-10-09 · Epoch AI
InnovationEval
We found no rerun of the evaluation by another group.
2026-09-22 · METR
Claude Opus 5.5 evaluation
We found no outside rerun of METR's capability testing of this model (the five capability tasks only; the assessment also draws on information from Anthropic).
2026-09-22 · Weco AI
AIDE²
We found no public code for AIDE² and no rerun by another group.
2026-01-29 · METR
Time Horizon 1.1
We found no rerun of the TH1.1 estimates by another group.
2025-05-14 · Google DeepMind
AlphaEvolve
No one has reported rerunning the search behind the 4x4 result. DeepMind's public repository holds a results notebook, not the system. AlphaEvolve has been offered on Google Cloud since 2026-07-10; we found no rerun of the 4x4 search there or with an open reimplementation.
Rests on internal data 4
The result rests on a lab’s internal data or systems, so it can’t be rerun as published.
2026-10-13 · Google DeepMind
Gemini 4 Argon safety report
Read off an index built partly from internal models.
2026-09-17† · Anthropic
R&D Automation Index
Built from Anthropic's internal work records.
2026-09-06 · OpenAI
Research acceleration: The view inside OpenAI
Rests on OpenAI's internal usage data.
2026-08 · Anthropic
August 2026 risk report
Rests on Anthropic's internal research work.
Nothing to rerun 11
A statement, forecast or description of a method, with no measured result to rerun.
2026-10-18 · Ben Horowitz
Horowitz: zero of four preconditions met
A spoken statement in a podcast clip, with no measured result behind it.
2026-10-14 · Google DeepMind
Brin: the RSI treadmill
A spoken statement in a launch video, with no measured result behind it.
undated · Sakana AI
RSI Lab
An announcement of a new lab, with no measured result behind it.
2026-09-21 · OpenAI
Building standards for the next phase of AI
A statement of position, with no measured result behind it.
undated · Anthropic
When AI builds itself
A statement of position; the figures on the page measure how much Claude speeds up work.
2026-09-17† · Z.ai (Zhipu)
GLM inference infrastructure post
A statement of position; the figures on the page measure one model launch.
2026-09-14 · Dream-RSI authors
Dream-RSI paper
The quoted line describes how the method is built, not a measured result.
2026-09-10 · Yi Duan et al. (35 authors)
The Last AI Built by Humans
A roadmap and survey paper that reports no new measurement.
2026-09-06 · OpenAI
An Alien Mind
A forecast. It rests on internal results that the page does not show.
2026-09 · Anthropic
We Must Pace the Frontier
A policy proposal, with no measured result behind it.
2026-08-26 · METR
OpenAI / Hugging Face incident investigation
An incident investigation. It reports what happened; there is no experiment to rerun.
“No rerun found” means we searched arXiv, GitHub, and the project’s repository and issues where one exists, up to the date shown on each entry. A derivation from someone else’s published result, or a check that it is correct, is not counted as a rerun. Know of one we missed? Write to [email protected].