Can AI answers about medicines be classified reliably under a written standard? A pre-registered study (pre-registered protocol)

The frozen protocol of a pre-registered study of how reliably Annex B of the open Evidence-Authority Standard (version 3.0) can be applied to sentences of real AI answers about medicines, deposited before any of the runs it governs. A pilot, in which three readers classified 123 items of the core set before the protocol was written, is reported in its section 8. The record holds the protocol, the text of the standard it applies, the labelling instructions, the held-out split's manifest, the human-labelling subset's manifest, and a manifest of the sha256 of every material: the test sets, the reference labels and their adjudication logs, and the scripts, which are released with the results. The held-out sentences themselves are not published. Version 3.0 of the standard has a reserved DOI (10.5281/zenodo.23594909) and is published after the runs, with their results. Results are reported against this protocol, with every deviation and its reason.