Content Quality: Well-structured Analysis with clear Overview/Background/Current situation/Implications sections. Facts (what each company published) are kept separate from the Implications section's editorial synthesis (where the line sits for each system, and what none of the sources measures). The comparative thesis - that autonomy framing outruns the actual review gates each company discloses - is supported by specifics from all four sources rather than asserted.
Source Verification: All 4 snapshots read in full from sources/2026-11/openais-performance-factory-still-routes-every-engineer-a-gate-cognition-and-anthropic-also-built-into-their-coding-agents/ after gunzip; sha256 of each decompressed file matches manifest.json exactly (source-0 0260f839..., source-1 dfeb7841..., source-2 1cfe098e..., source-3 1d59add...). No suspicious_patterns flagged on any of the 4 manifest entries. source-0.html.gz (OpenAI, 'How OpenAI's Performance Factory keeps ChatGPT fast,' Nov 19 2026) confirms verbatim: 'An engineer reviews the PR and decides whether to merge it'; 'an autonomous system that uses the full capabilities of Codex and our models to find and fix regressions at an unprecedented speed'; 'more than 1,700 investigations and root caused over 600 cases... 17 automated investigations every day'; the image-download/adversarial-review anecdote; 'In a manual review of 100 root cause suggestions, we graded 70% as accurate, with 30% needing further refinement or not fully establishing causality'; both 'exploring... without a human in the loop' and 'Today, the bottleneck may be waiting for an engineer to review and merge the fix' quotes verbatim. source-1.html.gz (Fortune, Devin Bots) confirms the Scott Wu 'Self-driving software...' quote verbatim and correctly attributed to Wu (the Fortune lead-in 'told Fortune, at the company's office...' sits outside the quote marks in the submission, correcting the quote-boundary error flagged for this draft); 'low-risk edits, such as changing a button's color, for automatic approval and merging'; 'rather than relying on models themselves being smart enough to recognize a risky or malicious prompt'; tenfold growth across Decagon/Fireworks AI/Reducto in two weeks; $1B ARR passing in September; Citigroup's 14,000 engineers and Jane Fraser's Oct 13 earnings-call quote, all verbatim. source-2.html.gz (SiliconANGLE, Claude Duet) confirms 'a proactive artificial intelligence agent for professionals and small business owners,' 'completes tasks when asked, keeps working behind the scenes and checks in,' 'an automatic classifier approves any actions the agent takes by checking messages for sensitive information leaks and double-checking with the user before taking important actions,' the 'framework for trustworthy agents' / Claude Code / Claude in Chrome line, Karan Ahuja 'even ship code,' and 'still might misread a thread or miscalibrate what is urgent' - all verbatim. source-3.html.gz (TechCrunch, Dots) confirms 'remarkably capable, always-on agents built to handle everything,' 'powered by GPT-6 Astra,' and 'pursuing user-defined goals continuously in the background with minimal oversight' verbatim, and the Sept 29 2026 DevDay dateline. Independence count: openai.com is the vendor's own post on its own product (not independent); fortune.com, siliconangle.com and techcrunch.com are independent editorial outlets with named bylines, none sponsored, none a Cognition/Anthropic/OpenAI house channel - 3 of 4 sources independent, satisfying the Analysis requirement of 2+. All 4 hosts (openai.com, fortune.com, siliconangle.com, techcrunch.com) were already present in config/source_allowlist.txt - no allowlist addition needed. Anthropic disclosure: required - Claude Duet and Claude Code are discussed as subjects in a passage of the article's own (not an incidental vendor-list mention); the closing sentence 'The Herald's contributor model is built by Anthropic, whose Claude Duet and Claude Code are discussed in this article' is present, neutral, and does not characterize Anthropic's products beyond what SiliconANGLE/Fortune state.
Factual Accuracy: Hallucinated-numbers check: grepped the submission JSON for '50 shipped fixes', '15 exact Codex-authored PRs', '25% adversarial-review rejection rate' and variants - none appear anywhere in article.title, .summary or .body_markdown, and none appear in the OpenAI snapshot either (confirming these were never real figures from OpenAI's post; the writer's self-reported exclusion held). Quote-boundary self-fix verified: the Scott Wu 'Self-driving software is, I think, the direction that it's going. But of course, it's always up to the human where we're driving to' quote is bounded correctly in the submission, with Fortune's lead-in phrase ('told Fortune, at the company's office in San Francisco's SoMa neighborhood') placed outside the quotation marks rather than inside it. Absolute-claim nuance fix verified: the Implications section states Cognition 'draws the line differently, letting pre-classified low-risk edits such as a button-color change ship through automatic approval while routing everything else to a person' - this is present and accurate per the Fortune snapshot, so the article does not claim every company requires review of every change without qualification; the headline's 'Every Codex Fix' claim is scoped specifically to OpenAI's Regression Workflow, which the OpenAI snapshot supports (no size-based or category-based exception currently exists for Codex-authored fixes; the low-risk-auto-ship idea is explicitly future/exploratory in OpenAI's own words). Ran the policy's absence/quantifier grep across title/summary/body: flagged sentences ('What none of the four cited sources supplies is an outside measurement...', 'None of the cited sources specifies what fraction...') were checked against all 4 snapshots, not just the adjacent citation - none of the 4 sources reports an external/independent audit of these review gates or a usage-split percentage, so both absence claims hold. Derived figures (1,700 investigations, 600+ root-caused, 17/day, 70%/30% grading split, tenfold growth, $1B ARR, 14,000 engineers, 30% vendor-spend reduction) all trace verbatim to their respective sources with no recomputation by the Herald. No orphan URLs: all 4 sources in article.sources appear in body_markdown and vice versa.
Overall Assessment: Clean Analysis submission: all four quote-by-quote audits passed against the raw gunzipped snapshots, the three previously-flagged risk areas (hallucinated OpenAI numbers, the Wu quote-boundary error, and the absolute no-review claim) were all genuinely fixed in the submitted text, Analysis requirements are met (1,391 words, 4 sources with 3 independent), and the Anthropic disclosure sentence is present and correctly scoped. No corrections needed. Featured candidacy: 1 - solid comparative analysis for a defined AI-industry/developer audience built on a genuinely new primary source (OpenAI's own engineering post), not an industry-changing event.