• AI assistance. Drafting, code and analysis are done with AI assistants (mainly Claude and Codex). As of 2026-10-05, 43 of our 65 Zenodo records disclose AI assistance in their description (40 of them name the assistant). The 22 that say nothing either way are mostly from before July 2026; we are adding the disclosure as those records are revised. The experiments are designed, run and checked under our direction, mostly on our own hardware; where a hosted model was used, the record says so.
  • Adversarial review before release. Since August 2026, our practice is that a claim about safety or validity is published only after separate AI reviewer sessions, commissioned by us, have tried to refute it. This is not external peer review. When they succeed, we fix the work or drop the claim.
  • Negative results on the record. Results that went against us are published at the same prominence as the ones that went for us, including defects in our own products.
  • Corrections. When a published number is wrong, we publish a new version that says what changed and why. Earlier versions stay available.
  • Numbers. Every number has one definition and a date (see Our numbers).