Make the conclusion easier to audit than to accuse.

Publication rules

1. Score atomic factual propositions, not people, parties, speeches or candidates.
2. Use the same evidentiary standard for materially equivalent claims regardless of speaker.
3. Seek the strongest credible evidence supporting the claim and the strongest credible evidence contradicting it.
4. Prioritize primary and directly inspectable sources; disclose source interests and limitations.
5. Track ancestry: derivative copies share one evidentiary root.
6. Separate provenance certainty from claim support.
7. Never treat missing provenance as proof of falsity.
8. Publish Evidence Support and Uncertainty as separate numbers.
9. Refuse a percentage when the proposition is insufficiently specified.
10. A challenge is a new research issue. It never edits the parent claim directly.
11. A resolved challenge affects the parent only through a specific, inspectable evidence operation weighted by challenge confidence, uncertainty and materiality.
12. All corrections, challenges, evidence additions, method changes and human overrides append new ledger events; history is never erased.
13. AI may search, extract, compare and explain; deterministic systems validate hashes, signatures, timestamps and certificate status.
14. Show what evidence would materially change the assessment.

Two-number confidence

Evidence Support X% estimates how strongly the current independent evidence supports the proposition as written. Uncertainty ±Y expresses how fragile that estimate is given missing evidence, source dependence, contradiction, weak provenance, context gaps, recency and method limitations.

The first release is explicitly an evidence-support index. It becomes a calibrated probabilistic estimator only after a historical benchmark set demonstrates calibration.

Challenge protocol

A challenge creates a separate Challenge Issue with its own atomic proposition, sources, provenance, Support % ± Uncertainty and ledger history. When resolved, it may add evidence, downweight or invalidate an evidence item, reclassify evidence, modify context, split a claim, or have no material effect. The parent claim is then recalculated as a new revision.

What we refuse to do

Truth Task Force does not use repetition as corroboration, infer intent from a low support score, silently delete old assessments, manufacture a percentage for an ambiguous claim, rank candidates, or recommend how anyone should vote.