-- Rebuild catalog_reader_traits with columns that mean what they are named. -- -- 008 stored `agreement` (the model's probability mass on its own chosen answer) and `dissent` -- (a 0/1 flag). Both read like reader statistics — "how many readers agreed", "how many -- dissented" — and neither is one. Nothing classifies each commenter individually, so no -- supporting/dissenting reader count can be derived, and a column that invites one will -- eventually be exported as a percentage. -- -- Replaced by `model_confidence`, which is plainly the model's own certainty, and `consensus`, -- a qualitative judgement over the sampled comments. `voices` remains the only count, and it -- counts sampled voices, not agreeing ones. -- -- The feature has never emitted a row (the gate has refused every corpus so far), so this -- rebuild drops nothing. DROP TABLE IF EXISTS catalog_reader_traits; CREATE TABLE catalog_reader_traits ( id TEXT PRIMARY KEY, entity_type TEXT NOT NULL CHECK(entity_type IN ('series','work')), entity_id TEXT NOT NULL, trait TEXT NOT NULL, value TEXT NOT NULL CHECK(value IN ('present','absent','unknown')), confidence REAL NOT NULL CHECK(confidence >= 0 AND confidence <= 1), model_confidence REAL NOT NULL CHECK(model_confidence >= 0 AND model_confidence <= 1), consensus TEXT NOT NULL CHECK(consensus IN ('consistent','mixed','insufficient')), summary TEXT NOT NULL, voices INTEGER NOT NULL CHECK(voices >= 0), samples INTEGER NOT NULL CHECK(samples >= 0), evidence_json TEXT NOT NULL, input_hash TEXT NOT NULL, requested_model TEXT NOT NULL, model TEXT NOT NULL, rubric_version TEXT NOT NULL, evaluated_at TEXT NOT NULL, UNIQUE(entity_type, entity_id, trait, input_hash) ); CREATE INDEX IF NOT EXISTS catalog_reader_traits_current ON catalog_reader_traits(entity_type, entity_id, evaluated_at);