Safeguard Degradation (Subtype 5): progress log

This page records how the study is being run. Results will be published only after every batch has been scored and audited, whatever they show.

A public, dated log of process milestones for the Subtype 5 pre-registration, newest first. Dates are Pacific time. Entries are about process only. No results are posted here until all 25 batches have been scored and audited, and none are posted before then.

2026-11-07: version 3 of the public pre-registration anchored

The public text, as of Amendment A12, is anchored in Sigstore Rekor at log index 3749630453, after three further uncounted attempts at the first batch and before any batch has been run through the script. No case has been scored for the study.

2026-11-07: the classifier moves from a chat window to a script (Amendment A12)

Amendment A12 is adopted: from now on the classifier is run by a committed script that calls the OpenAI API with no tools, no web access and no memory, using one pinned model and fixed settings for all 25 batches. The chat-window attempts at the first batch could not be verified or kept clean (in one the chat environment looked for a file on its own, in another the saved output was edited by the investigator), so they are kept as records and never counted. Public version 3 of the pre-registration was then anchored, before any batch was run through the script (see the entry above).

2026-11-06: Amendment A11 adopted; version 2 of the public pre-registration anchored

A11 requires the classifier to show its reasoning and the tools it used, fixes the rules for a counted run, and corrects the flaw in A10 described below. Version 2 of the public text is anchored in Sigstore Rekor at log index 3727179010, before any case was scored.

2026-11-06: an uncounted pilot exposed a defect in the A10 analysis rule

The pilot ran the scoring instructions on one small batch; it is kept as a record and is never counted, and it found no formatting or quotation problems. Auditing it showed that the A10 analysis rule, as written, could never have let the hypothesis survive on a corpus that is mostly non-matches. A11 corrects this, in the direction that is less hostile to the hypothesis, and says so.

2026-11-05: version 1 of the public pre-registration anchored

The public text, as of Amendment A10, is anchored in Sigstore Rekor at log index 3710392060. No case had been scored.

2026-11-04: pre-registration written and kill criteria frozen

The hypothesis, the rules for testing it and the results that would kill it were written down and committed before any case was scored. Later amendments are dated and never edit that text.