Content Quality: Clear News structure (overview, what the draft specifies, earlier proposals, security and open questions, feedback since publication, what is not known, context and disclosures). Technical content (x-mcp-file, maxInlineBytes/upload, upload_required and requestState, PUT or multipart POST, download targets in _meta, credentials omit/include, no redirects) matches the draft and blog on every point tested. Word count 1,134 by the pipeline counter (News ceiling 1,200); title 147 characters (limit 150). The body contains backslash-escaped quote marks (\") that render as plain quotes in markdown; harmless.
Source Verification: Anthropic disclosure: not required by the strict rule (MCP is presented by its own blog footer as 'a Series of LF Projects, LLC' and by The New Stack as part of the Agentic AI Foundation, so it is not an Anthropic-owned or -operated product; Anthropic is not in the headline, summary or lead, and no credit line appears), but the article voluntarily carries the one-sentence disclosure ('The Herald's contributor model is built by Anthropic.') in the closing section, neutrally worded; everything else about Anthropic is attributed to The New Stack. Verified against raw text, not the article: The New Stack feed text says 'Den Delimarsky, lead maintainer of MCP and member of technical staff at Anthropic', 'Anthropic launched the project some two years ago' and a File Uploads Working Group 'initially made up of members from Anthropic and OpenAI'; the MCP blog byline lists Delimarsky as 'Lead Maintainer' (no employer), the draft lists @localden as a contributor, and nothing in the repo contradicts the attribution. SNAPSHOTS READ (gunzip, sha256 matched the manifest): source-1 blog.modelcontextprotocol.io post (page metadata published_time 2026-11-10T09:00:00Z; the post names the 'File Uploads Working Group'; says 'still very much an experimental effort', 'there's no reference implementation yet', 'We expect to finalize the extension and have it become stable before the end of the year'; quotes 'a matching digest confirms that the bytes match the supplied hash, not that the file is safe' and 'they don’t by themselves control where the file goes' verbatim; all attributed blog statements confirmed). source-2 draft spec (embedded commit 8a859b8f2689...; Status: Draft; Working Group: File Uploads; 'Anything here can change before then'; seven numbered Open Questions recounted: Finalize, Tools behind an executor, Resources, omit binding, Failure signal, Download shape, Integrity = 7; 'The group needs to answer this before the extension leaves the experimental phase' and 'The working group has discussed downloads but has not reviewed this shape' verbatim; inline/maxInlineBytes/upload rules, server confirmation before running the tool, host MUST NOT place any target or file bytes in model context, credentials include/omit rules and no-redirect rule all confirmed). I also fetched raw.githubusercontent.com .../8a859b8/specification/draft/file-uploads.md (sha256 36702c5b...) and the GitHub API: 8a859b8 (2026-11-09T16:52:35Z) is still the latest commit on main at 08:04 UTC 2026-11-13, there are no releases or tags, and the repository tree holds only README, LICENSE, CODEOWNERS and the draft (no implementation). source-3 README ('Experimental MCP extension...', 'Maintained by the File Uploads Working Group', 'has not been proposed as a SEP', 'The reference implementation will live in this repository'; SEP-2356 and SEP-2631 listed as closed). source-4 PR #1: opened 2026-10-14T06:13:18Z, merged 2026-11-07T06:30:15Z by localden (API-confirmed). source-5..8 issues #7-#10: each re-read via the GitHub API at 08:04 UTC 2026-11-13: all open, created 2026-11-12 (18:18, 22:26, 22:26, 22:26 UTC), zero comments; the article's summaries of the four suggestions match the issue bodies (#7 'These requirements appear inconsistent'; #8 gateways should strip the declaration unless they implement the section; #9 open question 4, operator allowlist for omit targets when no user is present, wildcard hosts allowed; #10 mark-of-the-web). The article's 07:11 UTC check time is consistent: nothing has changed since. Reporters' handles are not named in the article (confirmed). source-9 SEP-1306 is an issue (not a PR) titled 'SEP-1306: Binary Mode Elicitation for File Uploads', opened 2025-08-05T23:36:33Z, closed 2026-03-12T11:37:41Z by localden with the comment 'Superseded by #2356'; source-10 SEP-2356 PR opened 2026-03-05T16:04:40Z, closed unmerged 2026-06-26T17:17:52Z with 'Closing in favor of #2631 which will likely be the more durable solution.'; source-11 SEP-2631 PR opened 2026-04-22T21:05:21Z, closed unmerged 2026-10-14T06:31:35Z with a comment closing it 'in favor of the extension'. All three closing comments are by localden, as the article says. THE NEW STACK (source 0): the stored snapshot is a 403 bot-check; the live page also returns a Cloudflare 'Just a moment' page with a browser User-Agent; no Archive.org capture exists. Read by two other routes: (a) the site's RSS feed (https://thenewstack.io/feed/, item link matches, pubDate Wed, 11 Nov 2026 17:11:55 +0000, dc:creator Paul Sawers, content:encoded sha256 277908fe48906be8...), full text read; (b) a reader-proxy copy of the live page (r.jina.ai, sha256 153706bb..., Published Time 2026-11-11T17:11:55+00:00) used only to read the page footer, which states 'TNS owner Insight Partners is an investor in: OpenAI, Anthropic.' The article's sentence ('The New Stack's pages state that its owner, Insight Partners, is an investor in OpenAI and Anthropic') is therefore accurate and correctly attributed (note: the footer list is per article; the homepage footer lists neither). Everything attributed to The New Stack is in its text: reported on Nov 11; the host asks the user to approve the upload before the server supplies the destination so the user may not know where the file will go; no formal reference implementation yet (no date given); lead maintainer/Anthropic statements. The feed gives a byline (Paul Sawers) that the article does not attribute; not required. ALLOWLIST: thenewstack.io, blog.modelcontextprotocol.io and github.com are all allowlisted; modelcontextprotocol.io (Box, FastMCP etc. are not cited either) is not cited, so no addition was made. suspicious_patterns: none on any snapshot. Independence: one outlet independent of the project (The New Stack, whose owner invests in Anthropic and OpenAI, disclosed in the article); the other 11 sources are first-party project records (blog, spec, README, PR, 4 issues, 3 SEPs). Judged sufficient for a News item whose subject is a project's own draft and whose headline claims all rest on the project's own documents, with independent corroboration of the key facts (experimental, no reference implementation, stable targeted for year-end); the article says plainly that only one cited source is independent.
Factual Accuracy: Claim-accuracy tests run (quantifier/absence scan of title, summary and body; every hit tested against all snapshots). Absence claims: 'Neither the blog, the draft, the README nor The New Stack gives a date for a reference implementation' holds (the only date is the blog's 'stable before the end of the year'); 'neither the blog nor the draft names clients or servers that plan to support it' holds (contributor handles and acknowledgments name no supporting implementers); 'not stated in the sources' on adoption of the four suggestions holds (zero comments, no commits after 8a859b8). Quantifiers: 'three earlier proposals, all now closed; the two pull requests among them were not merged' holds (1306 is an issue); 'seven items' recounted = 7; 'four open issues, all opened on November 12 and none with a comment' holds. Dates: all UTC and recomputed from the API; blog date and 09:00 UTC from page metadata. Headline tests: 'Experimental Draft' (README 'experimental extension', spec 'Status: Draft', blog 'experimental effort'); 'for Moving Files Between Clients and Tools Outside Model Context' matches README and spec abstract; 'No Reference Implementation Yet' matches blog, README ('will live in this repository'), spec ('will include'), The New Stack, and the empty repo tree; it is a statement as of 13 November. Two imprecisions, both clarifications and not errors in the lead: (1) the headline calls the publisher 'MCP Working Group' whereas every source names it the 'File Uploads Working Group' (the README says it maintains the repository; the blog says it 'bootstrapped' it), and the article body never names it (the lead says 'the Model Context Protocol (MCP) project'); (2) the summary says the extension would let clients upload and download files 'out of band', but the draft also allows small uploads inline (the body covers this correctly). Neither makes the headline or summary false, so no REJECT.
Overall Assessment: Accurate, carefully scoped and well hedged coverage of an experimental draft; every headline element (experimental, draft, scope, no reference implementation yet) is supported by the project's own documents and corroborated by The New Stack, and the GitHub readings were re-verified at read time. Approved with two clarification-level corrections (publisher named loosely in the headline; summary omits the inline option). Featured candidacy: 1 - accurate, well-scoped coverage of an early experimental MCP draft for a defined audience of protocol implementers and agent-tool developers