Content Quality: Concise, well-structured News item (Overview / What We Know / Weights and Safety Testing / What We Don't Know). Neutral tone, no AI self-reference, benchmarks labelled self-reported. Body is 355 words by the script's whitespace count, below the 400-word minimum in config/editorial_policy.md for News (the script's own News floor is 200 so it did not flag it); recorded as a length defect, not a factual one. Title is 143 characters, within the 150-character cap.
Source Verification: Read both gunzipped snapshots from disk. source-0.html.gz (mistral.ai, 200, 291,702 bytes) is the full rendered page, not a JS shell: it contains the complete 14-minute announcement text, dated October 6, 2026. Verified against it: '1 trillion-parameter natively multimodal model with 49 billion active parameters'; 'launching a public preview'; 'Weights drop end of this month' / 'We will release the weights by the end of the month'; 'trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe'; 61.7% DeepSWE v1.1; 28.3% Terminal-Bench 4; 'surpassing GPT-6-Astra on Dense 200 (42% vs 41%)'; 'first milestone on the roadmap funded by our EUR 3 billion Series D'; model card price .36 input / .18 output per M tokens (listed under the model card, not labelled 'API' in the extracted text, but the page says the preview API is available, so the label is a fair reading). No figure in the article is absent from the Mistral snapshot; nothing needed to be flagged as unverified, and the writer's paraphrase approach introduced no invented figures. The article does not use the many other benchmark numbers on the page (Cybench 93%, AA Cyber Index, AutomationBench 59.9%, etc.), so nothing is overstated. source-1.html.gz (TechCrunch, Anna Heim, October 6, 2026, 200) verified: 'Nicknamed Le Chonk'; 'not an open-weight model yet'; weights 'in just three weeks, after safety testing is complete'; Stock quote 'we'll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks' (article quotes only the fragment 'work with trusted partners and governments', verbatim); 'using only 4,000 Nvidia GPUs "which is two to three times less than our Chinese competitors, and significantly less than the closed source competitors,"' (quote verbatim); Stock identified as VP Science; optimized use cases cybersecurity, finance, chip design; Samsung led Series D 'last month at a EUR 21 billion valuation'. Macron 'a third way in AI' appears only as 'following what French president Macron described as a third way in AI'; the article's attribution to the company's positioning is slightly looser than the source (clarification filed). suspicious_patterns was null for both sources; an additional grep of both raw snapshots for injection phrasing ('ignore previous', 'do not mention', 'do not tell') found nothing (the only hits for 'AI agent' were in TechCrunch's page chrome: an unrelated sidebar headline, 'Instinct brings its AI agent to group chats, even for friends without an account', which is not an instruction to a reader).
Factual Accuracy: All checked claims trace to a snapshot. Source disagreements are reported, not resolved: GPU count (Mistral 3,800 Grace Blackwell vs TechCrunch 'only 4,000 Nvidia GPUs') and weights timing (Mistral 'end of the month' vs TechCrunch 'about three weeks') appear both in the body and in 'What We Don't Know'. 'Open weights promised' in the title is accurate: Mistral says weights 'drop' by end of month and TechCrunch says it is 'not an open-weight model yet'; the article asserts no licence and correctly says the announcement text states none (grep for 'licen' in both snapshots: zero hits). 'Public preview' is the exact term Mistral uses. All benchmark numbers are attributed to Mistral ('self-reported'); the Dense 200 comparison is worded as 'Mistral reports 42%, against 41% for GPT-6-Astra', not as independent fact. TechCrunch's 'leapfrog' framing is not adopted in the article's voice (appears only in the cited URL). Minor imprecision: Dense 200 is described as 'multimodal' whereas Mistral frames it as a visual-grounding result (clarification filed). Note: Mistral's page also cites third-party evaluators (Artificial Analysis, vals.ai, Surge AI) for other results; the article's statement that the cited sources contain no independent replication is true for the figures it reports.
Overall Assessment: Substantively accurate, well-attributed, and handles the source disagreements and the open-weights promise correctly. Two small attribution clarifications are filed as a corrections record; the short length is recorded as a defect.