'You cherry-picked the convergences.' The register rules and the quarantine list.

Michael Darius Eastwood · Independent AI alignment researcher

Published 3 July 2026

Any register of confirming events invites the suspicion that disconfirming ones were quietly dropped. The answer is not to protest good faith. It is to publish the rules and the rejects, which this programme does.

The rules

One external event per row. One primary source per row: an arXiv identifier, DOI, patent number or institutional URL a stranger can open. Gaps measured from a named artefact with a published hash. Every row carries a classification, and the classifications are not all favourable: CONCURRENT rows explicitly credit the other party's priority, as the Sharma and Chopra entry does in terms.

The quarantine

Candidate events without a confirmed primary source sit on a published pending list and are excluded from the headline count. When this register was audited in July 2026, the count fell from a claimed 31 to 19, because duplicates and unsourced rows were cut, in public, with the audit trail preserved. A programme willing to shrink its own headline number by a third is not optimising for impressiveness.

The standing invitation

The falsification dashboard lists what would kill each claim, including this one: the register fails if its rows do not survive source-checking, or if someone demonstrates a systematic pattern of relevant disconfirming events omitted. A weekly monitor script checks that every cited source still resolves. If you find a row that fails its source, or a disconfirmation that belongs on the record, the correction gets published the same way the retraction was. That is the deal the whole programme is built on.