Data and reproduction
The result store is published on GitHub as self-contained record bundles: protobuf files that carry the conversations, judge scores, outcome events, and any activation tensors. Each experiment’s report-cited dataset digest is stamped in the bundle metadata.
The Study 3 calibration: four store experiments added to the combined records bundle beside the thirteen released in August with their digests unchanged, plus a Study 3 capture bundle of pooled activation records from every steered, envelope, dose-sweep and replay cell. Study 4 collections are deliberately left out and ship with that study.
Cited by Study 3
The Study 2 store consolidated into self-contained bundles: one combined records file carrying every experiment’s report-cited digest, the three Study 2 capture bundles (fixed-input, own-trajectory and fresh-generation replays), the calibration captures, and the token-retention pass in volumes.
Cited by Study 2
The Study 2 collection in the earlier layout: one bundle per experiment plus loose activation-capture files with manifests. Superseded by the consolidated release the next day; kept because the results post cites it.
Cited by Study 2
Adds the distress-v3 battery pilots and the first activation captures at reference precision (bail and distress replays, fixture directions), the calibration that the Study 2 registration was built on.
Cited by Study 1
Study 1 complete: the registered confirmatory run and the method arm, the trial run, instrument and ladder calibration, the judge bake-off, the bail-tool arms, the Gemma distress control and the SmolLM3 probe. One bundle per experiment.
Cited by Study 1