// Evidence classes

A	Primary sources: court records, audit reports, company filings, operator documents we fetched ourselves.
B	Serious press and academic research.
C	Community reports (Reddit, forums) - paraphrased, linked, labeled as anecdote.
D	Our own measurements. Empty until the test phase completes; provider pages say so explicitly.

// Scoring

Seven dimensions: privacy & logging, ownership & track record, security, censorship resistance, performance, honest value, ease. Weights differ per kind (a free app is judged harder on trust than a paid service). Sub-scores are derived deterministically from documented facts; the formula ships in scripts/seed.js and is reproducible from the data.

Performance and ease are null until we test. The total re-normalizes over the known dimensions - no provider is penalized for our missing tests, and none is flattered either.

// Affiliate blindness

Scoring paths never read affiliate fields; a source-level test enforces it on every run. See how we make money.

// Corrections

Sources rot and facts change. Every page carries its check date; every correction with evidence is applied and logged on the change feed.