Positioning: PubMed stores scientific papers. ScientificClaims.org stores the evolving answers to scientific questions. The Scientific Claim Registry (SCR) is infrastructure for the memory of science — not its truth, consensus, or recommendations. It registers, it does not arbitrate. What v0.4 adds over v0.3: v0.3 delivered the implemented, live system. v0.4 makes every piece of evidence in it independently verifiable and verified — a per-evidence trust layer that, to our knowledge, no other claim/evidence registry has. Verifying that a DOI resolves proves the source exists — not that the claim is faithful to it. v0.4 closes that gap: (1) Sentence-level provenance (R-AI-13) — every evidence record carries a verbatim quote from the source that grounds its stance, with deterministic grounding (the quote must be an exact span of the source text), catching a hallucinated or paraphrased quote with no model in the loop; each claim becomes falsifiable by inspection. (2) Adversarial dual-model verification (R-AI-14) — a second, independent model re-derives the stance and judges faithfulness against the source, at two altitudes: per-source (does this source contradict or misattribute?) and per-claim (does the statement assert a specific present in no source — i.e. fabrication, checked against the union of the claim’s sources). A record is verified only when stance and provenance are independently confirmed; otherwise it is recorded as disputed (never deleted) and flagged. (3) Two distinct confidences, never fused (R-CLM-17): GRADE (the study’s quality) is kept separate from extraction confidence (how well the source was read). Full audit of the pilot. All 423 evidence sources across 42 questions were audited and reviewed: 416 verified, 0 disputed, 7 unverified (98% verified). The process caught and corrected 13 statement fabrications (concrete specifics — sample sizes, statistics, mechanisms — present in no source, including one statement with seven invented clinimetric figures), cross-paper contamination (statements assembled from other studies’ findings), stance mislabels, and review-design errors — silent errors that DOI-resolution and single-model extraction cannot surface. Every fix is recorded in each claim’s change log. Made public and machine-first (R-SITE-17): a per-question verification rollup (“N of M sources independently verified”) on every question page and in its JSON, a registry-wide seal on the homepage and questions index, and a global figure in the static read API. Verified / disputed / unverified are shown honestly side by side. Object model (unchanged). The scientific question (e.g. SQ-LIP-000007) is the primary navigable object — neutral and stable. Underneath it, claims (SCR-LIP-000001) are structured, versioned units of evidence linked with an explicit role (supporting / contradicting / refines / context), each with PECO context, a GRADE rating, dated provenance, a change log, and now sentence-level provenance and an independent verification verdict. A claim is a graph node: one article can inform claims under several questions. Lipedema pilot (proof of concept, disease-agnostic by design): expanded to 42 versioned scientific questions and 347 evidence claims (624 question–claim links), each claim now carrying verification metadata. Prior-art honesty: SCR does not invent the claim primitive or persistent claim identity — nanopublications/Trusty URIs, micropublications, Wikidata, CIViC, ClinGen, Epistemonikos/PICO, SciFact, GRADE and MAGICapp already exist. The v0.4 contribution is adversarial, sentence-grounded verification of each evidence link, exposed machine-first — built on top of that prior art, question-centric, registering-not-arbitrating. Operated by the BIO (Biological Intelligence Observatory) research method. This version establishes the priority date of the AI-verification layer and the expanded, verified lipedema pilot dataset.
Alexandre Campos Moraes Amato (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: