Provenance Record

Well-structured News piece (592 words, in range). Clear Overview / What We Know / What We Don't Know / Analysis arc. Technical depth is appropriate and the article consistently frames every benchmark as company- or vendor-reported, with explicit caveats about the absence of independent verification at launch.

All three source snapshots read from disk (gunzip -c). source-0 (SCMP, status 200): confirms M3 debuted 1 Jun 2026, 1M tokens 'five times more than its predecessor, the M2.7', architecture cut computational requirements to 'as little as one-twentieth of previous levels' slashing inference costs, beat GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro, did NOT disclose model size or training infrastructure. source-1 (MarkTechPost, status 200): confirms SWE-Bench Pro 59.0% (surpasses GPT-5.5 and Gemini 3.1 Pro, approaches Opus 4.7), Terminal-Bench 2.1 66.0%, OSWorld-Verified 70.06%; per-token compute '1/20th that of the previous-generation M2 models', '>9x in the prefill stage and >15x in the decoding stage at 1M-token context'; weights+technical report open-sourced within 10 days; 'image and video input and desktop computer operation natively', 'mixed-modality training from step 0'; benchmarks tested on internal infrastructure with Claude Code scaffolding. source-2 (The Decoder, status 200): confirms 'first open-weight model to combine top-tier coding performance, a one-million-token context window, and native multimodality'; BrowseComp 83.5 vs Opus 4.7's 79.3; MSA processes selected segments not every token pair; pricing requests up to 512,000 input tokens at standard rate with longer contexts costing more, token plan $20/mo for ~1.7B tokens up to $120 for 9.8B tokens; weights+report published on Hugging Face and GitHub.

Every specific verified against the snapshots. Predecessor attribution is correctly distinguished: the 5x context expansion is attributed to the M2.7 (SCMP) while the 1/20 per-token compute and >9x prefill / >15x decode speedups are attributed to the M2 (MarkTechPost) — both attributions match their respective sources exactly. All benchmark numbers (59.0, 66.0, 70.06, 83.5/79.3) are exact. Pricing figures exact. June 1 2026 date and 10-day weights timeline correct. No direct quotes appear in the article, so no quote-verbatim risk. Headline ('tops GPT-5.5 on SWE-Bench Pro'), summary, and lead are each sourced and each correctly framed as a company claim ('that It Says', 'company benchmarks').

High-quality, well-sourced submission. Every specific (benchmark numbers, 1M context, speedup multipliers and their correct predecessor attribution, pricing, dates) verified against the on-disk snapshots, and the article consistently and correctly frames the benchmarks as vendor-reported and not independently verified. The only blemishes are two cross-citation imprecisions where a true fact is attributed to one cited source though it appears verbatim in another cited source — both URLs are in article.sources, the facts are accurate, and the provenance chain is unbroken, so this does not warrant a correction. APPROVE.