Provenance Record

Well-structured News piece (Overview / What We Know / What We Don't Know / Analysis), 823 words, within the News category's 400-1200 word range. Neutral, precise language; appropriately hedges company-supplied benchmark claims and flags them as such in the 'What We Don't Know' section rather than presenting them as fact.

All 8 sources fetched successfully (HTTP 200); sha256 of each decompressed snapshot verified against sources/2026-07/moonshot-ai-releases-kimi-k3-a-28-trillion-parameter-open-weight-model-that-rattles-chinese-tech-and-chip-stocks/manifest.json before reading, no mismatches. source-0.html.gz (Fortune, Jul 16) confirms release date, DeepSeek V4 1.6T parameters, pricing figures ($15/$4.40/$0.87/$50 per million output tokens), Yutong Zhang quote, $2B/$20B funding round in May, $200M+ ARR. source-1.html.gz (Fortune, Jul 17) confirms TSMC -7%, SoftBank -9.0%, Nvidia -1.2%, Meta -2.4%+, Nasdaq 100 -1.0% at 2pm ET, Z.ai 'almost 30%' drop, Xi Jinping WAIC Shanghai remarks. source-2.html.gz (Forbes, Jul 17) confirms the 2.8T parameter figure, K3-trails-Fable-5/GPT-5.6-Sol framing with 'competitively'/'substantially outperformed' verbatim quotes, and Zhipu -28.4%/MiniMax -15.6%/Z.ai -28.4% HK trading figures. source-3.html.gz (TechCrunch, Jul 16, pre-launch FT-sourced piece) is listed in the sources array as background context but is not directly hyperlinked in the body; read and found consistent, no conflicting claims drawn from it. source-4.html.gz (TechCrunch, Jul 18) confirms Arena.ai/Vals AI independent corroboration, Xi Jinping/WAIC coincidence, Nasdaq -1% Friday, and the verbatim 'now, everything seems heightened' quote re: DeepSeek R1 (Jan 2025). source-5.html.gz (Tom's Hardware, Jul 17) confirms the 2.8T figure and 'largest open-weight AI model' framing, providing the second independent corroboration of 2.8T against Fortune's single outlier figure of 2.7T. source-6.html.gz (SCMP, Jul 20) and source-7.html.gz (PYMNTS, Jul 20) confirm the subscription pause, unaffected existing subscribers, and the underlying quoted material discussed in the quote-splicing finding above. No snapshot failed and none required an Archive.org or live WebFetch fallback.

The 2.8-trillion-parameter figure used in the headline and lead is independently corroborated by two of the eight sources (Forbes, Tom's Hardware); the article correctly avoids citing the sole outlier for that number (Fortune's Jul 16 piece states 2.7 trillion), instead using Fortune only for claims it does support (pricing, DeepSeek V4's parameter count, funding, ARR). This is a defensible, transparent resolution of a genuine cross-source numeric conflict, consistent with the submitting bot's noted resolution. All stock-move figures trace to their correctly cited sources, and Fortune's and Forbes' independently-tallied Z.ai figures ('almost 30%' vs '28.4%') are kept distinct and separately attributed rather than conflated into a false consensus. One quote-construction issue found: the SCMP-sourced Moonshot statement about pausing subscriptions is presented as one continuous quotation but is actually a splice of three non-contiguous fragments, with connecting material omitted without an ellipsis (see findings). No fabricated facts, no orphan source URLs (every body URL appears in article.sources), and no unsourced claims in the headline, summary, or lead.

Well-sourced, carefully hedged News piece that correctly resolves a real cross-source numeric conflict (2.8T vs. 2.7T parameters) and accurately attributes multiple overlapping stock-move figures to their correct sources. One minor, single-instance quote-splicing issue in a secondary supporting quote (not the headline/summary/lead) merits a public corrections note. APPROVE_WITH_CORRECTIONS.