Provenance Record

Well-structured News piece (781 words) with clear Overview / What We Know / What We Don't Know / Analysis sections. Technical depth is appropriate for a regulatory-clearance clinical-AI story and the writing carefully separates what each source supports.

All 6 source snapshots read from disk (status 200, none failed). source-0.html.gz (Nature/PMC PMC12328201): CONFIRMED training set of 1,245,273 ECG–echocardiogram pairs from 230,318 unique patients; SHD composite definition (LVEF ≤ 45%, LVWT ≥ 1.3 cm, moderate/severe RV dysfunction, pulmonary hypertension, moderate/severe VHD, moderate/large pericardial effusion) verbatim; internal AUROC 85.2% (CI 84.5–85.9); reader analysis of 150 ECGs with EchoNext accuracy 77.3% / sensitivity 72.6% / specificity 80.7%, cardiologists non-AI 64.0% improving to 69.2% with AI; external cohorts Cedars-Sinai (n=10,177), Montreal Heart Institute (n=10,862), UCSF (n=6,106); '5–7% drop in AUROC (78–80%)' phrase verbatim; the 3,200 figure confirmed as '3,200 cardiologist interpretations' (150 ECGs across ~13 readers, 1,600 non-AI + 1,600 AI-assisted). source-1.html.gz (Columbia CUIMC): CONFIRMED Pierre Elias quotes 'EchoNext basically uses the cheaper test to figure out who needs the more expensive ultrasound' and 'It detects diseases cardiologists can't from an ECG' verbatim, plus 'more often than cardiologists, including those who used AI to help interpret the data.' source-2.html.gz (NewYork-Presbyterian): CONFIRMED 13 cardiologists / 3,200 ECGs / 77% vs 64%, and June 22 Nature Medicine peer-reviewed transplant case ('first peer-reviewed account of its kind'). source-3.html.gz (Healthline): CONFIRMED FDA cleared June 22, $8.5M raised by Pathway Labs, 'more than 700,000 paired ECG and echocardiogram records,' the five named conditions verbatim, and OpenEvidence partnership. source-4.html.gz (OnHealthcare.tech): CONFIRMED OpenEvidence 'used by 750k+ verified clinicians' and 'six structural heart conditions.' source-5.html.gz (STAT News): CONFIRMED Pathway Labs spinout developed at NewYork-Presbyterian and Columbia.

Every load-bearing statistic (training-set size, reader-study numbers, AUROC, external validation cohorts, dollar figures, dates) traces to its cited primary source and matches verbatim. No efficacy claim is overstated: the article correctly frames the 13-point accuracy gap, explicitly surfaces the 5–7 point AUROC degradation on external cohorts, and flags the single-case transplant report as anecdotal rather than population-level. The 'What We Don't Know' section honestly reconciles the training-set discrepancy (Nature's 1,245,273 pairs vs Healthline's >700,000 records) rather than papering over it. No hallucinated quotes or fabricated specifics detected.

High-quality, accurately sourced submission. Verdict changed from the script's automated APPROVE_WITH_CORRECTIONS to APPROVE: the sole finding is an off-allowlist warning, not a recoverable factual error, so there is nothing for a corrections record to honestly document. Every clinical and regulatory claim was personally verified against the source snapshots. Ready for publication.