About
Last updated 2026-09-23
Every week a number goes viral: $1M ARR in 17 days, 10x cheaper in orbit, #2 on the leaderboard. WeirdTrue puts each one on trial. Anyone can bring evidence. An AI judge grades it in public, level by level, citation by citation. New evidence can overturn the verdict, and the whole history stays on the record.
How a case works
- Claim: a statement someone actually made, quoted verbatim, with who said it, when, and the source.
- Evidence: a source URL plus the exact passage from it. We fetch the source, and the passage has to appear in it verbatim, or the submission is rejected.
- Verdict: for each claim, a support level, the reasoning, and citations to specific evidence. Every verdict records the model and prompt version that produced it.
- Timeline: who added what, when, and which verdicts were overturned by which evidence.
Support levels
A verdict grades how well the citable evidence supports a specific claim. It is not a true/false call on the claim itself.
- Well supported: direct evidence from at least two independent sources agrees with the claim and nothing solid contradicts it.
- Partially supported: direct evidence backs the core of the claim but not its scope, degree, or conditions; or only one independent source exists.
- Insufficient evidence: the evidence can neither support nor refute the claim; or it is second-hand; or the only support is the interested party's own statement.
- Partially contradicted: direct evidence conflicts with a major part of the claim, but the claim could still hold.
- Contradicted: direct evidence from at least two independent sources conflicts with the claim and nothing solid supports it.
- Not adjudicable: the claim contains no checkable factual assertion (opinion, prediction, value judgment, satire), or none of the evidence has a source.
Rules that hold every time
- Opinions never enter a verdict. Only sourced evidence does.
- If every supporting citation comes from an interested party, the level cannot exceed “Insufficient evidence”.
- Independent means different organisations with their own access to the facts. Reprints and rewrites of one press release count as one source.
- Predictions and feasibility claims are broken into premises. Each premise is a hard constraint, an evidenced fact, or a labelled assumption.
- A re-adjudication starts from scratch under the current rules. When the level changes, the verdict says whether new evidence or a rule change caused it.
Who decides
The verdict is produced by a large language model (currently hosted on Alibaba Cloud Model Studio) following a public method: the prompt, the level definitions, and the per-domain constraint packs are versioned, and each verdict shows which versions it used. Earlier verdicts stay visible alongside the current one.
The people who matter are the ones who bring evidence. Anyone can read. Signed-in members can open cases, add evidence, and request a verdict.