How to audit this release
Threshold-passing with abstract means selected-label probability ≥0.75 plus a non-empty abstract. This is an operational score-and-evidence route, not a human-validated class or an accuracy claim.
Public evidence: the Version 1.1 public-minimal package contains aggregate outputs, configuration, checksums, verification code and a data-minimised agreement subset. It excludes the full bibliographic corpus and full work-level model and membership rows.
Internal agreement check: 79 of 88 agent-adjudicated reference records overlapped the frozen v0.2 snapshot, with exact three-class agreement on 73 of 79 records (92.4%). The reference labels were adjudicated by AI agents, so this is not independent human validation, external model accuracy or probability calibration.
Version note: v0.2 freezes an OpenAlex work-ID snapshot captured on 29 September 2026. For 39,888 records overlapping v0.1, it carries forward prior bibliographic inputs, Jev outputs and institution-membership evidence. It freshly fetched and classified 1,284 new work IDs. A further 2,860 records present in the v0.1 snapshot were absent from the residual live query and are excluded from v0.2; overlapping membership edges were not fully re-fetched.
Read the corrected Version 1.1 working paper or return to the aggregate Observatory dashboard.