Provenance Record

Clean, well-organized News piece (888 words, within the 400-1200 range) with Overview / What We Know / What We Don't Know structure. Sentences are tightly tied to specific attributed sources throughout; no filler or padding.

All 4 source snapshots read from disk and sha256-verified against manifest.json (source-0.html.gz TechCrunch, source-1.html.gz MarkTechPost, source-2.html.gz MLQ News, source-3.html.gz OpenAI GPT-Live System Card at deploymentsafety.openai.com), all fetched at 200 status, no archive_fallback needed. Every direct quote in the article was checked verbatim against its snapshot and matched exactly (e.g. TechCrunch: 'full-duplex models, meaning they can speak and listen at the same time'; MLQ: 'GPT-Live-1 at its highest reasoning setting scored 84.2% on the GPQA scientific reasoning test, up from 45.3% for Advanced Voice Mode' and 'GPT-Live-1 scored 75.2% compared to 0.7% previously' and 'participants chose GPT-Live-1 over Advanced Voice Mode in 75.7% of comparisons'; MarkTechPost: 'the API is planned soon', 'Video, screen sharing, and full multilingual parity are not available at launch', nine remastered voices, visual cards for weather/stocks/sports; OpenAI system card: 'can listen and respond continuously instead of waiting for a clearly defined turn to end', the 0.88->0.82 and 0.97->0.95 safety score figures, and the Preparedness Framework Tracked Categories language). No hallucinated quotes or misattributions found. All 4 sources array URLs are present in the body and vice versa (no orphan/gap).

All claims sourced to the 4 cited outlets check out. One error found independently of the automated checks: the body states 'The launch comes two days before OpenAI publicly launched GPT-5.6,' but GPT-Live launched July 8, 2026 and GPT-5.6 reached general availability July 9, 2026 per the Machine Herald's own prior article (2026-07/10-openai-launches-gpt-56-to-the-public-ending-its-government-gated-preview) -- a one-day gap, not two. This claim is unsourced (no inline citation) and sits in a subordinate context paragraph, not the headline/summary/lead, so it is being handled as a corrections record rather than a rejection.

A carefully sourced, accurately quoted article with one small, isolated, and honestly-correctable chronology error in a subordinate paragraph. APPROVE_WITH_CORRECTIONS: publish with a public corrections note on the 'two days' claim.