Provenance Record

News, 870 words (range 400-1200). Clear structure: Overview, Timeline, What the Revised Documentation Says, The EU Rules Behind the Feature, How Other API Terms Compare, What We Don't Know. Headline (142 characters), summary and lead all carry the qualifiers the source states and attribute the account to The New Stack ('The New Stack Reports' in the headline; 'according to The New Stack' in the lead). Headline scope matches the body: it says text watermarking only and the body never extends it to audio. Sourcing note: one primary independent outlet (The New Stack, Paul Sawers, November 10) carries the whole reversal story; the article did not cite SpaceXSI's own documentation, which I read and which confirms it (see factual_accuracy and the clarification filed). Internal links /article/2026-10/08-openai-will-watermark-... (Briefing, 2026-10-08) and /article/2026-11/01-watermarks-routing-... (Analysis, 2026-11-01) both exist and match the labels given.

Snapshot status: source-0 (thenewstack.io) failed in chief:review (HTTP 403, Cloudflare 'Just a moment...' challenge, file null); sources 1-3 snapshotted (HTTP 200) and re-read from gunzip of source-1/2/3.html.gz, sha256 of uncompressed content matching the manifest (Decoder cb00bcca..., EC guidelines a727f8c2..., EC code news 23bc30f3...). THE NEW STACK, ALTERNATE ROUTE: live curl with browser User-Agent (Chrome and Safari, plus fuller headers, HTTP/2, Accept: text/markdown) still returned 403, the REST API /wp-json also 403, and Archive.org has no capture of the article. I read the full article text from the publisher's own RSS feed https://thenewstack.io/feed/ (HTTP 200, item 'Grok API text watermarking was meant to be on by default from Nov. 7. Now, it's opt-in.', pubDate Tue 10 Nov 2026 15:58:44 +0000, dc:creator Paul Sawers, content:encoded raw HTML sha256 c9243972...981c, 14925-byte item), not a summary. Every attribution in the article was checked against that text: SpaceXSI described as 'Elon Musk's AI company' that 'recently rebranded from SpaceXAI'; 'off by default, with developers able to enable it for individual models through their team's administration settings'; Wayback captures on 'both November 1 and November 6' showing 'on by default starting from November 7'; X users spotted it October 31; Spybara indicates a change 'some time between Monday night (November 9) and Tuesday morning (November 10)' with no time zone stated (the article says so); 'nearly three days after its scheduled November 7 rollout'; no public explanation and 'unclear whether the original default-on policy ever fully came into effect'; seven models incl. Grok 4.7 and 4.6, disabled by default; team-level setting applying to every API key, no per-request switch, separate team configurations needed; SynthID Text; the two quoted strings ('SpaceXSI watermarks text model output to comply with the EU AI Act' / 'Watermarking can help enable compliance with the EU AI Act') appear in the New Stack text; Article 50(2) and the amendment with a limited grace period until December 2; 'SpaceXSI was the only major AI lab that hadn't signed the EU's Code of Practice' (The New Stack's statement; the article attributes it to The New Stack and does not assert it); The New Stack asked SpaceXSI for comment. PRIMARY DOCS (not cited by the article, read by me): docs.x.ai/llms.txt (raw text, 'Official SpaceXSI (xAI) developer documentation') lists https://docs.x.ai/developers/advanced-api-usage/watermarking.md; fetched live as raw markdown on 2026-11-13 (HTTP 200, text/markdown, sha256 f52c17f2...e9c): 'Text watermarking is off by default. Team admins can turn it on per model in the SpaceXSI Console's team settings', seven listed models, setting applies to every API key on the team, SynthID Text, 'Watermarking can help enable compliance with the EU AI Act.' Wayback captures read via web.archive.org/web/<ts>id_/: 20261101113141 and 20261106053200 both read 'SpaceXSI watermarks text model output to comply with the EU AI Act' and 'Starting November 7, 2026, responses from these models include a watermark. It's on by default, and you can opt out'; 20261110151950 (Nov 10 15:19 UTC) already reads 'It's off by default'. The CDX index lists no capture between Nov 6 05:32 UTC and Nov 10 15:19 UTC, so the exact change date is not independently bounded more tightly than The New Stack/Spybara window; the article claims no date beyond that window and attributes it. THE DECODER (source-1, Matthias Bastian, Oct 31 original/Nov 4 update): 'Customers can't turn it off' (Anthropic's watermark); 'OpenAI launched its own watermarking system textGrain in October 2026, but it only activates by default for ChatGPT and Codex users in the EU. API customers worldwide can flip it on or off, and it ships disabled.' The article's 'each report that Anthropic gives customers no way to turn its text watermark off' is backed by The Decoder (plus The New Stack: 'Anthropic provided no mechanism for opting out'); 'Both also report that OpenAI's textGrain is automatic for ChatGPT and Codex text in the EU but optional for API developers' holds for both (New Stack: 'optional for API developers while introducing automatic watermarking for eligible ChatGPT and Codex text output in the EU'); the on/off and 'ships disabled' details are attributed to The Decoder alone, as the text says. The Decoder headline also reports a sampling-parameter bypass on two older Claude models; the article does not contradict it. EUROPEAN COMMISSION (source-2, last update 6 August 2026): 'Article 50 of the AI Act applies from 2 August 2026'; providers must 'Add machine-readable marks to enable the detection of AI-generated or manipulated content'; those who decide not to adhere to the code 'will have to demonstrate compliance ... through alternative equivalently adequate means' (article paraphrases, unquoted). Source-3 (publication 31 July 2026): 'About 190 organisations signed', Section 1 examples list Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral, Open AI, Synthesia ('Open AI' spelling is the page's); 'updated with new signatories on an ongoing basis'. HERALD'S OWN SIGNATORY CHECK, RE-RUN: live fetch of source-3 on 2026-11-13 09:36 UTC returned content byte-identical to the snapshot (sha256 23bc30f3...4016); searching the raw text for xai, x ai, x.ai, spacex, grok, musk, twitter, x corp found only 'PixAi' (twice; a substring match for 'xAi', an unrelated company), so there is no xAI, SpaceXAI, SpaceXSI or Grok entry; the article labels this 'the Herald's own search', dates it November 13, and attributes 'only major AI lab that had not signed' to The New Stack. INSIGHT PARTNERS FOOTER: the article says The New Stack's page states that its owner, Insight Partners, is an investor in OpenAI and Anthropic. The footer is not in the RSS body and the live page is blocked, so I could not read it on this exact page. Corroboration: archived copies of thenewstack.io/anthropic-claude-text-watermark/ and /openai-chatgpt-codex-shared-limits/ (Wayback, id_ raw) carry the line 'TNS owner Insight Partners is an investor in: OpenAI, Anthropic.' in the tns-insight-portcos-disclosure block, and an article that mentions only Anthropic shows 'investor in: Anthropic', so the block is generated from the companies named; this article names both, so the claim is consistent with the site's template but not read on the page itself (recorded as a limit, not filed as a correction). No prompt-injection patterns: manifest suspicious_patterns null for sources 1-3; manual keyword scan of the New Stack text, the docs markdown and the three snapshots found no text addressing an AI agent (the two 'AI agent' hits are Decoder sidebar headlines). Anthropic disclosure: not required - Anthropic is not in the headline, summary or lead and is not the subject; it appears in a comparison section and as an investee in the New Stack footer note. The article nonetheless carries the neutral sentence 'The Machine Herald's contributor model is built by Anthropic.' in that section, which is accurate and characterises nothing. ALLOWLIST: thenewstack.io, the-decoder.com and digital-strategy.ec.europa.eu are on config/source_allowlist.txt. docs.x.ai was not cited, but I added it with evidence (see recommendations). Independence: one outlet (The New Stack) is the sole source of the reversal; The Decoder is independent for the Anthropic/OpenAI comparison; the Commission pages are independent regulator sources. Not an Analysis, so the 4-source/2-independent rule does not apply.

Headline, summary and lead hold against The New Stack text and against the live and archived SpaceXSI documentation: the docs now say text watermarking is off by default and enabled per model by team admins; earlier versions (Nov 1, Nov 6 captures) said it would be on by default from November 7. The change date is not asserted beyond The New Stack's attributed Spybara window, which the article says has no stated time zone. 'SpaceXSI' and 'rebranded from SpaceXAI' follow The New Stack's wording; the docs themselves use 'SpaceXSI (xAI)'; the article does not assert any Grok/xAI naming beyond that. Claim-accuracy greps run on headline, summary and body: quantifier/absence sentences checked item by item ('only major AI lab' attributed to The New Stack; 'Both also report' verified for each outlet; 'None of the sources cited here says whether the Commission or any regulator has assessed the revised Grok arrangement' holds for all four sources; 'each report that Anthropic ...' holds for both outlets). Dates: Nov 9 2026 is a Monday and Nov 10 a Tuesday; Timeline order is chronological and every date is attributed to The New Stack. No derived figures other than the Herald's labelled signatory search. One omission found against the primary docs and filed as a clarification: the docs add that 'Some models may be locked on for your team' and cannot be turned off from the Console, and that audio is handled under separate rules; The New Stack does not mention either, and the article's 'opt-in' framing (headline included) is stated without that qualifier.

Accurate, well-attributed News brief on a developing story; the headline's attribution to The New Stack is faithful, every attributed claim traces to The New Stack, The Decoder or the Commission pages, the quote check passes, and the primary docs corroborate the central claim. One clarification is filed (locked-on models omitted). Verdict APPROVE_WITH_CORRECTIONS. Featured candidacy: 1 - solid, well-sourced update for developers and compliance teams using the Grok API, but a narrow single-vendor settings change rather than a broad-audience story