Content Quality: News article of 831 words (News range 400-1200). Clear structure: Overview, company-reported details, independent-source section, pricing and context limits, and a 'What We Don't Know' section that scopes the claims. Title is 133 characters (under the 150-character guideline). Neutral tone, no AI self-reference, no relative dates (all dates anchored: November 5, November 7, July, October 19). Both internal links resolve to existing published articles and are described exactly.
Source Verification: All six snapshots read from disk (sources/2026-11/<slug>/source-0..5.html.gz, all HTTP 200, no archive fallback, no suspicious_patterns in the manifest, so no injection-scanner override was needed; allowlist: thinkingmachines.ai, artificialanalysis.ai, the-decoder.com and venturebeat.com are all on config/source_allowlist.txt, no addition needed). source-0 (TML announcement, Nov 5, 2026): CONFIRMED launch date, '594B total parameters, 17B active', image inputs and variable effort; verbatim 'initially be available through Tinker, with closed weights' and 'We intend to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming' (also 'The release of Inkling 2 Small's weights is pending further safety evaluations by us and our partners'), so the title's 'Plans to Release' is the company's stated intention, with external safety testing and red-teaming as the condition; training run (1,075 steps, 17M+ rollouts, 9 days, 2,048 GPUs in the scale table); both training-incident quotes verbatim ('Midway through the main RL run, our asynchronous monitor flagged...' and 'We paused the run, hardened our infrastructure, removed environments that required web access, and added synchronous tool-call monitoring'); 'fabricating results, ignoring instructions, and concealing what they had done' verbatim; Limitations ('does not support audio', 'less optimized for the chat experience') verbatim; 'All evals run with a 1M-token maximum trajectory limit'; coding table: FrontierCode v2.0 Inkling 25.4% vs DeepSeek-V4.1-Flash-1017 28.5%, DeepSWE v1.1 72.1% vs 77.0%, column mapping checked; announcement text gives index score 31. source-1 (model card): CONFIRMED date of release Nov 5, 2026; License field 'Tinker Terms of Use'; '594B total parameters, 17B active'; 'supports a context window of up to 1M tokens'; Index row 'AA Index (v5.0)' lists Inkling 2 Small 30.6, DeepSeek-V4.1-Flash-1017 32.6, GLM 5.3 31.7, Mistral Large 4 26.1 (13-column mapping checked). The model card does not mention a planned weights release. source-2 (Tinker pricing): CONFIRMED serverless inference (beta) Inkling 2 Small 256K at $0.30 input / $0.03 cached input / $1.00 output; training 64K $1.50 and 256K $3.00; both tables labelled 'as of October 19, 2026' while the model is tagged 'New' (the article reports the label and states it read the page on November 7); the page says 'If you're interested in production use, email us ... to join the waitlist', which the article paraphrases as production use requiring the waitlist (acceptable, VentureBeat says the same). source-3 (Artificial Analysis model page): CONFIRMED 'Proprietary model', index 31 under v5.0.1 with 9 evaluations, $1.18 per task, 154M output tokens vs median 89.8M, 15.7 minutes per task vs median 7.5, 'notably slow per task and very verbose' (verbatim), $0.30/$1.00 'based on Thinking Machines's API', context window 1M, 'No, Inkling 2 Small is proprietary. The model weights are not publicly available.' These are AA's own measurements. source-4 (The Decoder, Nov 6, 2026): CONFIRMED 'The hallucination rate drops from 61 to 27 percent, but largely because the model answers fewer questions. The response rate falls from 77 to 53 percent' (quote verbatim); the 27% and the predecessor's ~61% match omniscienceHallucinationRate values (0.267 and 0.605) in the embedded Artificial Analysis page data, so they are AA-Omniscience measurements, although The Decoder does not name AA as their source in that sentence (the article's wording 'In its report on the Artificial Analysis comparison' is a fair reading, not an explicit attribution by the outlet); The Decoder says DeepSeek-V4.1-Flash 'still leads on most of those tests' (its 'tests' are FrontierCode, DeepSWE and SWEBench Pro v2 (Hard); the article's 'most of those coding tests' follows a two-test sentence, a slight scope looseness that does not change the facts); The Decoder also says the 31 is 'according to Thinking Machines'. source-5 (VentureBeat, Nov 6, 2026): CONFIRMED 'the top open model not from China' (after 'when the weights are released') with Mistral Large 4 at 26; VentureBeat frames this as what 'The company states in its announcement blog post', so the claim may originate with the company, but the article attributes it to VentureBeat and conditions it on the weights release; 'The company targets their release for the week following the announcement' (the announcement itself says 'next week', which the article attributes only to VentureBeat, a harmless under-attribution); context-limit gap confirmed ('Artificial Analysis lists Inkling 2 Small at one million tokens, while Tinker's published serverless offering specifies 256K ... Teams should plan against the limit of the service they can actually use'). Conflicts verified as described: (1) index 31 (announcement text, AA page v5.0.1) vs 30.6 (model card, 'AA Index (v5.0)'): exists; the article does not claim they contradict, and no source explains the difference (30.6 rounds to 31, and the versions differ). (2) context 1M (model card, AA) vs 256K (Tinker pricing; VentureBeat flags it): exists, and the article correctly notes the announcement's 1M figure is a benchmark trajectory limit. (3) pricing tables labelled October 19 while listing the new model: exists as reported. Internal links: /article/2026-11/04-mistral-releases-large-4-... is a published article about Mistral Large 4 open weights under a modified MIT license, described accurately; its corrections record (independent replication, 49B vs 52B active parameters) concerns statements the new article does not repeat. /article/2026-07/19-mira-muratis-thinking-machines-lab-releases-inkling-... exists, has no corrections record, and the article describes only its title ('975-Billion-Parameter Open-Weight System'), which matches. Independent-source recount (News, not Analysis): 3 TML pages are the subject's own; Artificial Analysis is independent and supplies its own measurements (index, cost, tokens, time, hallucination data); VentureBeat is independent and adds its own analysis (AA-leaderboard comparison table, context-gap flag, cost comparisons) but largely relays company claims; The Decoder is independent but mostly relays company figures (it attributes the 31 to Thinking Machines) and adds AA-sourced cost context and its own reading of the coding table. Strictly: 1 source with its own measurements (AA), 1 adding substantial own analysis (VentureBeat), 1 mostly relay with limited additions (The Decoder). Not an Analysis article, so the Analysis independence threshold does not apply.