Every filter state is a URL; the URL is the query; the query is citable via /q/⟨hash⟩. The page, the API and the export parse the same parameters.
The current cohort, streamed from the database: every work column, the machine labels, the provisional scores, and the per-row validation status. Exports are capped at 100,000 rows. Mints a permanent /q/ link for this exact query. The same filters always produce the same link, whoever asks.
Labels cover 2 of 80 works in this cohort. The rest are unlabeled, which is not a negative label: the label table is sparse today and grows as labeling rounds land.
Distilled predictions cover 80 of 80 works in this cohort. Predictions are machine_predicted_unvalidated. The Gemma side is a direct model label for every work (title-only); the Codex side is a distilled, calibrated classifier. Candidate is the union; consensus is the intersection.
