Source Verification: All eight snapshots read from disk (source-0 to source-7, sha256 of each decompressed snapshot re-computed and equal to the manifest; no suspicious_patterns on any source, so no injection judgement needed; all hosts are on config/source_allowlist.txt). SOURCE-0 Ars Technica (Kyle Orland, Oct 9, 2026): confirms Chen and Stratton, Jellyfish data, 300 million work events, more than 700,000 employees, over 700 firms, 2021 through March 2026, 'a mix of directly measured AI usage and analyses of GitHub activity', 'difference of differences', +30% lines of code / +20% commits / +23% pull requests for AI coding agents, Jira issues and epics 'did not change in a statistically significant way', +49% review process time, changes requested 'nearly doubl[ing]', comments +35%, reviewers +14%, 80 percent of firms with some AI code review by March 2026, 23.3 percent of review comments and 10.8 percent of pull requests, 'significant updates and upgrades' after the March 2026 cutoff: all verbatim or accurate. SOURCE-1 TechRadar (Oct 12): 30% attributed to 'when using AI', body gives +49% review-and-merge time and +35% comments, 7.03 to 10.48 days, and the subheading says 'Comments per pull request are up 49%': confirmed as the article describes. HARVARD PAPER (not cited, not allowlisted, read for verification only, no allowlist addition): both outlets link https://fion.ac/jellyfish.pdf. The live copy now (last-modified Nov 13, 2026, sha256 c9627dd0...faefa) is the version dated Nov 9, 2026: 445M work events, 720,542 workers, 655 firms, January 2021 to August 2026, and different estimates (agents +44.0% lines of code, +36.0% commits, +35.7% pull requests; assistants +30.4% lines of code at 12 months; review time +56.1%; comments per PR +38.6%; share with a comment +13.1%). The version live when Ars and TechRadar reported is the one dated August 4, 2026, which I read from the Internet Archive capture of Oct 12, 2026 (web.archive.org/web/20261012090249if_/https://fion.ac/jellyfish.pdf, sha256 27f8d1fe...c7c096): 300 million work events, about 700 firms and 700,000 workers (718 firms, 725,938 workers in the sample table), January 2021 to March 2026, agents 30% lines of code / 20% commits / 23% pull requests, review time +49% (baseline 7.03 days), comments per PR +35%, share of workers reviewing +14%, nearly 80% of firms using AI code review as of March 2026, 23.3% of review comments AI-generated and 10.8% of pull requests with at least one AI-generated comment. Every Harvard figure the article attributes to Ars and TechRadar therefore matches the Aug 4 version; none is contradicted by it. Against the Nov 9 version the article's Harvard figures are out of date, and its Herald-voice statement that 'The Harvard data ends in March 2026 and the survey was fielded from April to June 2026, so the two do not cover the same months' holds only for the earlier version (see corrections). The article's own disclosure ('did not cite the paper itself and did not check these figures against it or against any later revision') is accurate as to what it cites; the wording of Ars' 10.8 percent as 'AI agents ... responsible for ... 10.8 percent of pull requests' does not match the paper (corrections). SOURCE-2 The Register (Oct 21): 'Respondents estimated that a quarter of incoming pull requests' verbatim, 52 percent code-authoring estimate, curl announcement and Daniel Stenberg quote present, no sample size (no '234' on the page): confirmed; it restates the release and adds the curl item. SOURCE-3 TechRadar (Oct 22): 'Researchers found that artificial intelligence can reduce the amount of time it takes to write code by a staggering 52%' verbatim, no 'estimate' qualifier on that sentence: confirmed. SOURCE-4 Linux Foundation release (Oct 20, Red Hat sponsorship, '25% of incoming PRs are estimated to be low-quality', 'Respondents estimate average time reductions ... code authoring (52%)'): confirmed. SOURCE-5 LF report PDF (read via pdftotext): methodology page gives web survey, LF newsletters/social/partner channels, fielded April-June 2026, 234 qualified respondents, +/- 6.41% at 95%; Figure 2 'Average estimated percentage of incoming PRs that are low-quality and/or AI-generated', 25%, respondents = 111; Figure 19 time-reduction chart respondents = 96, authoring -52% (4h18 to 2h06), code review and feedback -44%; Figure 1 caption 'Number of respondents = 174' with All maintainers 52% / 17% and Large projects 64% / 7% (increased / reduced review time); report text 'of the 234 ... 206 identify as contributors and 134 as maintainers' and Q29 'Number of respondents = 134'. The 174 versus 134 is as the article says (noted, not reconciled; I did not reconcile it either). The 52 percent appears for two quantities as the article says. SOURCE-6 Google post (Nov 6, 2026, Sam Saccone and Madhuri Garikapati): 13% of file-reading sessions exposed, 6% of encounters acted on, 'Over 4 in 5 of those actions (83% sessions)', 41% closed (32% fixed + 9% obsolete, plus 2% duplicate/restricted), 57% still open, regression 70% / 77% / 105% 'Per session' controlling for developer, codebase complexity, task type, model tier, number of files read, prompt length and step, 'the perfect proxy' (verbatim 'provide the perfect proxy'), 'exceeding 95% acceptance from human reviewers' verbatim, 'thousands' removed over 'the last 6 months'; no period for the traces and no named agent tools: confirmed. SOURCE-7 InfoQ (Sergio De Simone, Nov 10): restates the post, adds three social-media comments (Tulasiram, Ryan White, Awais Raza, Muhammad Zohaib), headline about technical debt, body 'they focused on TODO(bug/...) comments': confirmed. Anthropic disclosure: not required - Anthropic and Claude are not the subject, headline, summary or lead of the article and the article never mentions them; the incidental mentions in the sources (Claude Code as one measured tool and an Anthropic Institute citation in the Harvard paper, a claude.com footnote and a court case in the LF report, an unrelated related-article link on the Register page) are not carried into the article. INDEPENDENCE (recounted by me): outlets independent of the study authors and of the first-party pages: Ars Technica, TechRadar, The Register, InfoQ = 4 outlets (the two TechRadar pieces are one outlet; the LF release and report and the Google post are first-party and not counted). Per case: Harvard 2 outlets (Ars, TechRadar; both accounts of the paper, not independent measurements); LF 2 outlets (The Register restates the release and adds the curl item; TechRadar restates it and drops the 'estimate' qualifier); Google 1 outlet (InfoQ, restates the post and adds social comments, no measurement of its own). The policy minimum (at least 4 sources, at least 2 independent of the subject, counted by outlet) is met with 4 outlets, and the article describes each restating outlet as restating; the Google case alone rests on one restating outlet plus the first-party post, which the article says plainly ('reports no measurement of its own'), so I do not read it as an independent corroboration of the Google figures and neither does the article.
Factual Accuracy: Dates: earliest coverage Oct 9 (Ars), latest Nov 10 (InfoQ), Oct 12 TechRadar, Oct 20 LF release, Oct 21 Register, Oct 22 TechRadar, Nov 6 Google post: all correct and in order. Derived numbers recomputed: 7.03 to 10.48 days equals +49%; 4h18 to 2h06 equals -51.2% (report says -52%, the article gives the report's figure as the report's). Quantifier test (grep over title, summary and body): 'each one' (three reports covered by outlets), 'both outlets' (Ars and TechRadar both accounts of the study), 'neither Google's post nor InfoQ' (traces' period: true), 'neither Ars nor TechRadar reports figures for tools released after it' (true), 'only Google's describes a tool its authors built' (true), 'The Register ... does not give the sample size' (true): all hold for every cited source. Three recoverable issues, all subordinate, recorded in the corrections file: (1) the Herald-voice claim that the Harvard data ends March 2026 and does not overlap the survey's fielding holds only for the Aug 4, 2026 version, not the current Nov 9, 2026 version (data to August 2026, different sample and estimates); (2) the Ars phrasing that AI agents were 'responsible for ... 10.8 percent of pull requests' is reproduced although the paper (both versions) says 10.8% of pull requests receive at least one AI-generated review comment and 23.3% of review comments are AI-generated; (3) the Google paragraph repeats two points already filed in the corrections record for the earlier Herald Briefing on the same post (the post words the 83% as 'actions' and, in a parenthetical, 'sessions'; the post gives the 70/77/105% 'per session' with no stated comparison group), and 'Google's figures come from a regression' overstates, since only the 70/77/105% come from the regression and the 13/6/83/41/57 figures are descriptive shares. None of these is in the headline, summary or lead as a false statement, and the article states that the Harvard figures come from the coverage and were not checked against the paper.