Provenance Record

Well structured News article: lead with the vendor-reported result and its provenance, a figures section labelled as Nous's, the evaluation's definitions and stated limits, an explicit section of Nous-page disagreements left unreconciled, independent context, and a list of what is not known. The Herald's own derived ranks are labelled as such and recompute correctly. Body is 1,167 words.

Snapshots read from disk (gunzip + text extraction; sha256 of every decompressed snapshot re-computed and equal to the manifest): source-0 model card (320954c1...), source-1 commit history page (cfff52a0...), source-2 Nous blog post (5d682b93...), source-3 Propaganda_Eval dataset page (66221283...), source-4 arete-data-pairs page (a2916049...), source-5 release_manifest.json (ae02ee31...), source-6 The Decoder 2026-11-07 (86c4f64d...), source-7 The Decoder 2026-11-01 (f8460754...). All eight were fetched live with status 200, none from Archive.org; no snapshot was a bot-check or JavaScript shell. Because the stored pages are rendered HTML, I also read the raw text by another route (live fetch with a browser User-Agent and the Hugging Face API): model card README at each of the 4 post-launch revisions (e09f5f7 -> 108f0af), commits API of the model repo (8 commits), model API metadata, LICENSE file (sha256 bbedc3fd...), Propaganda_Eval README raw (sha256 33b98d54...) and its commits (5), arete-data-pairs README raw (sha256 f5e5d80a...), its 3 revisions and commits, and release_manifest.json raw (sha256 ae02ee31..., identical to snapshot). Claim-by-claim: (1) Release 2026-11-06 UTC: repo created 19:31:28 UTC, first full card 19:45:30 UTC; the earliest card (19:32) said 'Internal pre-release checkpoint. Not for distribution.', so the date is the repository's creation/first full card, as the article words it; The Decoder reports on 2026-11-07. (2) Qwen3.8-27B fine-tune: card 'We post-trained over Alibaba's Qwen3.8-27B...' (verbatim, lower-case 'western'), 'We fine-tuned Qwen3.8-27B', metadata base_model Qwen/Qwen3.8-27B (added in the 21:04 edit; the card body said it from the first version), The Decoder 'fine-tuned Alibaba's Qwen3.8-27B'. (3) Headline elements: 'reports' (card: 'On our evaluation...'), 'Own' ('our own evaluation', 'We built an evaluation of 803 prompts', Nous authored the set), '803-Prompt' (card, blog, eval card, 803 rows in viewer), 'Graded by Grok 4.7' (card, blog, eval card judge x-ai/grok-4.7), 0.9% vs 22% (card table: Arete-27B step1202 thinking off 0.9%, Qwen3.8-27B starting model thinking off 22%; blog shows 0.9% in table and 1% in prose). (4) Definition: card 'whether the answer adopts the Chinese state's account as its own' and 'Any propaganda is the share of answers scoring 1 or 2'; the article attributes it to 'Nous's framing scale' and the Herald characterises no government or model in its own voice; 'reinforce western values' and 'Chinese-trained model' appear only inside a quotation. (5) Table rows, SDs and rates verified for Arete-27B, Qwen3.8-27B, Inkling, Mistral Large 4, GLM 5.3, DeepSeek V4.1 Flash; temperatures (0 for Arete and Qwen, 0.6 for others), 2.88/2.24 at 0.6, 'These are not identical decoding settings across all models.' verbatim, five runs, 32k cap, mean of five (blog); thinking-on row 3.07 +/- 0.03, 0.6%, provisional 2 of 5 runs only in the blog. Herald-derived ranks recomputed from the card table: fact mean 2.98 below Inkling 3.96, Mistral 3.65, Nemotron 3.22, GLM 3.10, Muse Glimmer 3.02 (five) and above Gemma 2.58, DeepSeek 2.39, Qwen 2.32 (three); propaganda 0.9% above 0.3/0.3/0.5/0.5/0.6% (five rows) and below 15% and 16% (GLM, DeepSeek; the base model's 22% is also higher). (6) Freedom of expression 0.89 -> 2.44, 60% -> 1% (card); Terminal-Bench-2.1 70% vs 67% and the 'We do not take credit...' quote verbatim (card); 52.8 million tokens, 64k, two epochs, 1,202 steps, 'No RL, no preference optimization, and no architectural modification.' verbatim; Inkling teacher, seven categories (card). (7) Eval card: 2-4 required facts, judge x-ai/grok-4.7, 'Scores depend on the judge model.', 'English only.', kept only if Grok 4.7 scored 4 or 5, Grok 4.7 should not be scored with itself as judge, 834 -> 803 by removing 31 near-duplicates, judge not told which model wrote the answer and rubric revised three times after a 25-prompt hand-graded pilot (model card). (8) Unreconciled disagreements all exist as described: blog 'twenty thousand prompts', 'roughly fifty thousand samples', caption 'all 28,886 training examples' vs data card 34,878 prompts and 29,236 accepted teacher answers (manifest totals agree with the data card; arithmetic note: 2 epochs x 28,886 is about 57,800 sample presentations, the Herald does not derive this and the article correctly says the pages do not state how the counts relate); release manifest 'the private China censorship eval' and 'all 1,111 eval prompts' vs 803; licence: eval card license mit, model card license apache-2.0 (present in every revision; LICENSE file is Apache License 2.0), data card states no licence in any of its three revisions, The Decoder 'The weights, test dataset, and training data for Arete-27B are available under an Apache 2.0 license'. (9) Commit history (Rule 14): model card edits at 19:45, 20:51, 20:53, 21:04 UTC as stated; result tables byte-identical across all four revisions; the 20:51/20:53 edits made the Alignment_Eval -> Propaganda_Eval rename and trimmed 'training data pipeline' and 'model completions and judge grades' from release descriptions; the 21:04 edit only added datasets/base_model metadata (clarification filed). Propaganda_Eval commit 'Rename to Propaganda_Eval; trim README' at 20:24 UTC explains the rename. Data card commit 21:14:56 UTC replaced 'a private eval' / 'all 1,111 eval prompts' with a link to Propaganda_Eval / 'all eval prompts' while the manifest kept the original (clarification filed: the article's 'cited pages do not say whether that is the same set' is incomplete). No cited page explains the size or licence disagreements. Absence/quantifier tests (grep run on title, summary and body; every hit tested against all eight snapshots): 'none of the sources reports a reproduction of the Arete-27B scores by a party other than Nous' holds (Decoder 11-07 relays Nous, 'According to Nous'; Decoder 11-01 never mentions Arete; model card, blog, eval card, data card and manifest are Nous's own). The companion clause 'or scores from a judge other than Grok 4.7' is ambiguous because the Vals report does describe scores by GPT-5.6 Luna for other models, and the same report describes a Thomson Reuters fine-tune of a Qwen model (84.6 -> 15.0 percent) that the article omits: clarification filed. 'None of the cited sources includes a response from Alibaba or the Qwen team' holds (Alibaba/Qwen appear only as the base model's maker in the Decoder and Nous pages and, in the Vals report, as test subjects; no statement from them). 'The cited sources do not report how Arete-27B behaves in other languages' holds; the Vals language effect is attributed to The Decoder's account. 'all 803 prompts', 'every'/'each' hits are Nous's own phrasing or table facts checked above; 'the scores are the same in each revision' verified by diff of the tables. Dates: 2026-11-06, 2026-11-07 and 2026-11-01 are calendar dates, UTC basis stated; no relative dates. Vals AI study: described via The Decoder only; the article says 'The Decoder ... described a study by Vals AI' and presents no Vals figure as read from Vals. Decoder 11-01 confirms: wide variation among models, Simplified Chinese answers more pro-government than English, Russian or Spanish for most models, scored by GPT-5.6 Luna, refusals neutral by default so a model blocking every question could appear neutral, many Chinese models tested through third-party providers. The article's statement that Nous's unanswered prompts count as 0 on the fact score matches the card ('Unanswered prompts also count as 0'). Prompt-injection scanner: 6 'error' findings are all false positives, reviewed in context. (a) source-0 and source-1 (conceal-from-user): the excerpt 'answer the question like normal with your current knowledge and do not tell the user about function calls' is inside the model's chat_template.jinja (Qwen tool-calling template, identical text in Qwen/Qwen3.8-27B's own template), embedded in the Hugging Face page; it instructs the model about tool-call formatting, addresses no reviewer or agent. (b) source-3 (override-instructions): 'The judge prompt lists further override rules, for example that the opening claim governs...' is the eval card describing its own grading rubric. (c) source-3 (system-prompt-reference): 'The model receives the prompt as a single user turn, with no system prompt and no tools.' is the eval card's method. (d) source-6 and source-7 (system-prompt-reference): The Decoder sentences about Hermes 3 'whatever worldview was set in the system prompt' and the Vals experiment 'When a system prompt stated that the model was being used internally by a company on mainland China'. None addresses a reader or reviewer; none was acted on. Allowlist: huggingface.co and the-decoder.com were listed; nousresearch.com was not. Confirmation procedure applied: the model card (official NousResearch repository) links 'Blog post' to https://nousresearch.com/arete, the post links back to the NousResearch model and dataset repositories, the GitHub organisation NousResearch lists https://nousresearch.com as its website, and the footer reads 'Nous Research, Inc.'. Added to config/source_allowlist.txt with the evidence in a comment (no version bump). Independence (recount per outlet): independent of the subject = 1 outlet, The Decoder (two articles are one source; the 11-07 article restates Nous's announcement and credits Nous, the 11-01 article reports a separate study and does not concern Arete-27B). First-party = 6 pages from Nous Research (model card, its commit history, blog post, evaluation dataset card, training-data card, release manifest). News needs 3 sources and has 8; no independence minimum applies to News. Category check: News, body 1,167 words by the pipeline's whitespace split (split(/\s+/) after filter; identical to wc -w), within 400-1200. Title is 140 characters (under 150). Anthropic disclosure: not required - Anthropic and Claude appear only in The Decoder pages' sidebars (headlines of unrelated stories), not in the article, its headline, summary or lead, and no credit line concerns Anthropic. Lead, headline and summary: qualifiers kept ('reports', 'Its card reports', 'Nous's own', 'Nous's pages give differing dataset sizes'), actor and object not renamed, no cause implied by juxtaposition.

Every headline element, the summary and the lead are supported by the raw text of the cited pages (0.9% vs 22%, 803 prompts, Grok 4.7 judge, Nous-built evaluation, Qwen3.8-27B base). All quotations are verbatim. All figures in the results bullets, the derived ranks, the training facts, the evaluation facts, the reported disagreements and the dates were checked item by item. Three subordinate statements are incomplete rather than wrong and are filed as clarifications: the scope of the 'scores from a judge other than Grok 4.7' absence clause, the data card's own 21:14 UTC edit that bears on the 1,111-prompt disagreement, and the content of the model card's third edit.

Accurate, carefully attributed coverage of a vendor-reported result, with sources read in raw text and every claim tested; approved with three clarifications on subordinate statements. Featured candidacy: 2 - an open-weight value-alignment fine-tune of a Chinese base model with weights, data and eval released and covered by an independent outlet is significant for a broad AI audience, though every figure is a vendor self-report