Monitoring democratic institutions through public records
Did we acquire the expected inputs, with complete content? — Source/document coverage, content completeness, pagination, period coverage, metadata classification, and fetch errors.
Of the data we have, what's the processing backlog and do we have enough reference data? — Scoring/embedding backlog, baseline presence, and L2 assessment coverage.
Does detection catch known events and reject negative controls? — Known-event recall, negative controls, and layer attribution.
Is every derived artifact consistent with and fresh against its inputs? — Edge-contract invariants (G1a–G6): eligible docs scored, aggregates present with matching counts, enrichment/narratives fresh, no orphan categories.
Does detection hold up on historical data? — Per-category precision and noise against Trump T1 (2017–2018) known events.
