Evidence methodology and corrections

Claimlight reports what attributable evidence supports, contradicts, complicates, or cannot verify. It is not an oracle, does not judge an author's motives, and should not replace professional medical, legal, financial, or safety advice.

How a claim is checked

1. Separate compound sentences into independently checkable assertions.
2. Search with neutral and counter-evidence queries.
3. Prefer primary records, official data, peer-reviewed research, legislation, court documents, and direct statements.
4. Use reputable independent reporting when primary evidence is unavailable or needs interpretation.
5. Accept evidence only when its URL is traced to the research provider's web-search metadata and the source is reachable.
6. Match every source relationship to the exact assertion and display uncertainty or disagreement.

Verdicts

Supported means reachable evidence directly supports the claim. Contradicted means reachable evidence directly conflicts with it. Mixed means material parts differ or credible evidence conflicts. Unverified means evidence is insufficient, indirect, inaccessible, or cannot be attributed safely.

Source and model safeguards

Source text is treated as untrusted data. A model-provided URL cannot validate itself, and a decisive verdict is downgraded to unverified when no matching, reachable source directly supports the relationship. Claimlight preserves the consulted-source list and uses a versioned evaluation set before changing models or prompts.

Corrections and appeals

Claimlight never changes or publishes text without explicit acceptance. Beta participants can report an incorrect verdict or source from the evidence inspector and can request review through the beta support page. Confirmed methodology or evidence errors are added to the regression suite and described in release notes.