What we measure
Empirical quality
TruthX scores observable writing standards: claim clarity, specificity, attribution, evidence support, reasoning quality, factual responsibility, discussion value, and tone. The score is a context label, not a command about what to believe.
Integrity signals
The rubric looks for sourcing, logic, specificity, and proportionality. For practical writing advice, use the How to Write Well guide.
What is not measured
The system does not score whether it agrees with the underlying opinion, political side, or narrative. It scores measurable writing behavior: whether claims are specific, sourced, proportionate, logically connected, and careful about uncertainty.
Likes, follower counts, outrage, and user flagging are not used as the core ranking driver.
Model spread
When more than one scorer is configured, posts show the average score plus the low-high range across models. Opening the score panel shows each model vote and the short reasons returned by the evaluators.
Quality floor
Readers can set a sitewide quality floor. Low-scoring content is not deleted, but it can be hidden from feeds and replies until a reader chooses to lower the floor.
That is the distribution principle: people may have the right to post, but low-quality posts do not have a right to be forced into every reader's default feed.
OSINT
OSINT posts get two visible measures: post quality and verification. Quality scores discipline, source tier, information density, confidence signaling, and coherence, with redundant posts reduced after scoring.
Verification starts as awaiting verification when no pathway has raised the score. Self-attestation can create a capped baseline, source checks run over the first 24 hours, and manual sources can raise stale posts later. A score above zero is verified; zero after 24 hours is unverified. Account track record averages the last 50 resolved OSINT posts with the all-time resolved average.
AI bias
The system focuses prompts on empirical standards rather than ideology. When multiple AI scorers are configured, their spread is shown so readers can inspect disagreement instead of trusting a hidden single judge.