How Satya checks a claim

Route, retrieve, verdict

Every submission goes through the same pipeline. First, the claim is decomposed into atomic, checkable pieces and typed (a macro statistic, a budget allocation, an attributed quote, a viral photo or video, and so on). That type decides where we look: a claim about GDP is checked against MoSPI and RBI, a claim about a viral clip is checked against fact-checkers and reverse-image search. We then retrieve evidence from a curated set of sources and ask a language model to weigh it and write a verdict — but the model never gets the final word on what counts as evidence.

Citations are verified in code, not trusted from the model

Every quote a verdict relies on is checked character-for-character against the source document before it is allowed to appear. A quote that cannot be found verbatim in a real source is dropped, and a verdict resting on no surviving evidence — or only on ordinary news coverage rather than a primary source or a fact-checker — is downgraded to "Unverifiable" automatically. We would rather say "we don't know" than publish a confident-sounding guess.

Evidence tiers

• Primary source — government data, official records, court filings.
• Fact-checker — IFCN-certified fact-checking organisations.
• News media — credible outlets with independent reporting. On its own, this tier can add colour but cannot anchor a substantive verdict.

The ten verdicts

We don't force every claim into "true" or "false". Verdicts include True, Substantially true, Missing context, Cherry-picked, Misattributed, Outdated, Misleading, False, Disputed data (when authoritative sources genuinely conflict — this happens more than you'd think with Indian official statistics), and Unverifiable.

Verdicts attach to claims, never people

Satya evaluates statements, not the character of the people who made or shared them. A verdict that a claim is false is not a verdict about the person who said it. Where a verdict is severe (False, Misleading, or Misattributed) and concerns a named individual, it is held for a human reviewer before it is published — this is a deliberate, conservative safeguard, not an accuracy signal.

No engagement bait, no political lean

We check claims from every party the same way, and we don't dress up a verdict to be more shareable than the evidence supports.