How tiers are assigned
The tier framework is borrowed from accountability-investigation orgs (Project on Government Oversight, Government Accountability Institute) and adapted for our hybrid statistical / relational / temporal detector mix. A detector goes into a higher tier when its output requires multiple independent public-records sources to align. Tier A is the lowest because it surfaces deviations from a mathematical norm (Benford, Gini); deviation-from-norm has many innocent explanations. Tier D is the highest non-litigation tier because each leg of the chain is independently documented and the combined pattern is the signal.
Why "evidence_tier_default" matters
Each detector also carries a Direct / Indirect / Inference tag (separately from the A–F severity tier). This second axis, borrowed from Paul's Network of Networks work, asks: how many inferential steps separate the raw public record from the published claim? A finding tagged "Direct" means the public-records evidence is itself the claim (e.g., a corporate filing showing officer X). "Indirect" means the claim follows from a small chain of records. "Inference" means the claim depends on a statistical pattern that needs verification before being treated as truth-of-the-matter. evidence_tier_default is the typical classification for a detector's output; per-finding overrides are possible.
What the auditor cannot detect
This is a deliberately incomplete list of what we don't surface: private retainer income, federal-LDA flow not yet imported (~>200 results per filer pagination cap), state-court civil filings (NV Sup Ct + 8th JDC + 2nd JDC require Playwright we haven't run yet), regulatory-board votes and rulemakings (Phase 8b — partial), open meeting law minutes (#136 in progress), state employee payroll (Transparent Nevada — pending), and back-session NV legislative data prior to the 83rd (2025) Session. Each is documented in the verification page or the auditor task list.
Reproducibility
Every detector reads from the cached public-records data shipped under /data/ (the file column in each row above). Anyone can pull the underlying JSON and reproduce the findings independently. The detector source code is on the backlog to publish (see #317).