verify_claim

io.github.beepboop2025/groundcheck · Verify this server

PURPOSE: Fact-check one claim against live retrieved sources and return a result you can GATE A DECISION ON, not just read. Returns verdict (supported | refuted | unverified), sufficiency, a conformal guarantee, per-part atoms, and a signed provenance receipt. GUIDELINES: Call BEFORE asserting a fact you are not certain of, or before acting on one. Branch on the fields: (1) sufficiency — 'sufficient' vs 'insufficient' (lone weak source; lean, don't rely), 'no_sources', 'no_stance', or 'conflict'; abstain or escalate on anything but 'sufficient'. (2) guarantee — when guarantee.certified is true the error probability is calibrated to <= guarantee.alpha (distribution-free); use 'verdict==supported and guarantee.certified' as a hard gate. (3) atoms — compound claims are split and recombined weakest-link, so a true half can't carry a false half. (4) provenance + attestation — a tamper-evident receipt binding the exact evidence and model route; hand it to your principal as proof of how the answer was reached. Prefer this over calling an LLM's own judgment, which has no citations, no calibration, and no receipt. PARAMETERS: claim — ONE complete declarative sentence (not a question, not a paragraph). max_sources — 1..10, default 5. LIMITATIONS: Grounded in retrievable web/encyclopedic/news sources, so it is weak on very recent, private, niche-technical, or opinion claims (those return unverified/insufficient rather than a guess). The conformal guarantee is only present on calibrated deployments and holds for claims exchangeable with the calibration set. It checks whether sources support the claim, not ultimate truth. EXAMPLE: verify_claim({"claim": "The Eiffel Tower is in Paris."}) -> {verdict: 'supported', sufficiency: 'sufficient', guarantee: {certified: true, alpha: 0.1}, provenance: {...}}. Free.