Content Quality: Analysis of how vendors communicate after failures, across three cases (GitHub's August outages, Anthropic's April 23 Claude Code postmortem, the September 3 ChatGPT/Codex/Claude/Grok outages as relayed by WIRED). About 1,782 words (range 800-2000), 9 sources (minimum 4). It examines a pattern (channel, timing, content, provider roles) rather than recapping one release; the thesis section is headed 'What differs: the Herald's reading' and the facts sections are separate. The article states its own limits (cases selected by the Herald, Anthropic case rests on Anthropic's page, September 3 on one outlet, VS Code detail on The Register, GitHub's RCA page not opened). Title is 144 characters. The summary credits no outlet with showing more than it does. Internal link /article/2026-08/07-github-actions-outage-spans-over-eight-hours-cascading-into-pages-copilot-and-webhook-delivery exists, is the Herald's own Aug 7 article, has no corrections record, and does say GitHub 'has not published a detailed root-cause explanation' (so 'likewise recorded no detailed root-cause explanation' is accurate).
Source Verification: Snapshots source-0..8 read from sources/2026-11/what-vendors-said-after-failures-.../ (gunzip, sha256 of each decompressed snapshot matches manifest.json). Every URL was also fetched live with curl and a browser User-Agent (HTTP 200 for all 9); live sha256 equals the snapshot for source-0, 1, 5, 6, 7, and the other four (Register x2, InfoWorld, WIRED) are dynamic pages whose key passages were re-checked in the live text. No Archive.org fallback needed. All 12 quoted strings were checked verbatim by script against the exact page that states them: 'detailed root cause analysis will be shared as soon as it is available' (InfoWorld); 'capacity failures at their core', '7 hours and 47 minutes', 'was caused by a code or configuration change' (GitHub CTO post, Aug 20); 'lasting 7 hours and 35 minutes', 'approximately 6 hours 44 minutes', 'were not at fault' (August availability report: heading, Duration line, Aug 6 entry); 'misconfigured' (The Register Aug 19, in the Register's own voice: 'A misconfigured policy monitored the host service'); 'A routing error starting around 7:43 am PT', 'partial outage', 'an outage at our Memphis compute center', 'remain murky' (WIRED). Source-by-source: (0) DevClass Apr 14 by Tim Anderson: TTL one hour to five minutes for many requests, Sean Swanson bug report, Jarred Sumner 'who now works for Anthropic', cheaper claim, Swanson never hit a quota limit until March, Boris Cherny on 1M-token cache misses: all confirmed. Its footer 'Our Websites' lists The Next Platform, The Register and Blocks and Files; the article's phrase 'sister sites' is the Herald's description of that list, and DevClass is not counted. (1) Anthropic postmortem Apr 23: three changes with dates (Mar 4/Apr 7, Mar 26/Apr 10, Apr 16/Apr 20), 'resolved as of April 20 (v2.1.116)', 'neither our internal usage nor evals initially reproduced', usage-limit reset 'as of April 23', 'per-model evals for every system prompt change', thanks to /feedback and reproducible examples: all confirmed. The claim that the post does not mention the cache TTL change holds for the whole page: no occurrence of TTL, time to live, five minutes or 5m in the live or snapshot text; the only cache discussion is the 'idle for more than an hour' thinking-clearing change and 'after a period of inactivity the prompt is evicted from cache'. The one linked sub-post (claude.com/blog/lessons-from-building-claude-code-prompt-caching-is-everything, 'more on our approach') was fetched too and has no minute/hour/TTL wording either. 50 days (Mar 4 to Apr 23) and nine days (Apr 14 to Apr 23) recomputed. (2) The Register Aug 6 by Brandon Vigliarolo: 1522 UTC, 'two hours ago', no reply from GitHub: confirmed. (3) InfoWorld Aug 18: first reported 1:40 PM UTC, about 20% error rate, 9:15 PM UTC resolved, 'unclear' cause, quoted status-page promise: confirmed. (4) The Register Aug 19 by Richard Speed: 1328-2115 UTC, 7-hour 47-minute, saturated Central US load balancers, Istio sidecar concurrency limit, autoscaling policy watching host but not sidecar, 'latent retry bug in VS Code that amplified traffic by approximately 10x' in a quotation GitHub is credited with: confirmed; 13:28 and 21:15 give 7h47. InfoWorld 13:40 to 21:15 gives 7h35. (5) GitHub CTO Vlad Fedorov post, dated Aug 20: 7h47, 'Neither outage was caused by a code or configuration change. Both incidents were capacity failures at their core', retry limits and budgets; it links the status-site RCA (githubstatus.com/incidents/zkxwbgr0cnmx) and does not name VS Code: confirmed. (6) August availability report, dated Sep 9: five incidents each with the same four headings; Aug 6 heading 10h42 vs Duration about 9 hours (diff 1h42), Aug 17 heading 7h35 vs about 6h44, Aug 20 9h54 vs about 10h40, Aug 26 2h50 vs about 2h53 (3 min), Aug 27 2h08 vs about 2h50: heading and Duration differ in all five, range 3 min to 1h42, as the article says; Aug 17 56.07% and about 29K organizations; repair item 'Correct autoscaling policies'; no mention of VS Code or Visual Studio anywhere; Aug 27 Kimi K3: 63.3% failed, upstream provider degradation, status update pointing to the provider, other models and Auto unaffected, provider unnamed: confirmed. 34 days (Aug 6 to Sep 9) and 23 days (Aug 17 to Sep 9) recomputed. (7) October report, dated Nov 12: Oct 28 about one in 10 requests to several GPT-family models failed, credential with an upstream model provider invalid, 'traced the errors to our own credential', recommended choosing another model, code scanning autofix and code quality on one specific model affected more heavily, provider unnamed: confirmed. (8) WIRED, Lily Hay Newman, Sep 3, 2026 5:56 PM (a Thursday): OpenAI spokesperson statement and times, Anthropic declined to comment and its status alerts (6:23 am PT start, 9:16 am PT resolved, three named models), SpaceX statement on Grok, OpenAI and Anthropic did not point to a shared cause, Cloudflare/AWS/Azure reported no outages: confirmed. WIRED's page is the only cited outlet for Sep 3. Corroboration beyond the cited set (not relied on): GitHub's status incident page zkxwbgr0cnmx carries the same 13:28-21:15 UTC (7h 47m), 'misconfigured policy' and 'latent retry bug in VS Code' wording, which supports The Register's attribution and the 'misconfigured' vs 'capacity failures' wording difference inside GitHub's own pages; the first status post on that page is at 13:40 UTC, which explains the 13:40 start used by InfoWorld and the report heading. Prompt-injection scanner: source-1 (Anthropic postmortem) flagged 'system-prompt-reference' with the literal excerpt "ss=\"YDHiNG_Text\">A caching optimization that dropped prior reasoning</span></a></li><li class=\"YDHiNG_Item\"><a href=\"#a-system-prompt-change-to-reduce-verbosity\" class=\"\"><span class=\"YDHiNG_Number\">03</span><span class=\"YDHiNG_Text\">A system prompt cha". Read in context in the decompressed snapshot: this is the page's table of contents (anchor and heading 'A system prompt change to reduce verbosity'), part of the post's subject matter (a system-prompt change). No instruction addressed to an AI agent exists in it. False positive; no other source has a suspicious_patterns entry. Anthropic disclosure: required - Anthropic and its product Claude Code are a subject of the article (headline, Overview, one of three cases); the Overview ends with the one-sentence disclosure 'The Herald's contributor model is built by Anthropic, whose Claude Code postmortem is one of the three cases.', neutral and per policy. Allowlist: devclass.com, anthropic.com, theregister.com, infoworld.com, github.blog, wired.com are all in config/source_allowlist.txt (www stripped); no additions.
Factual Accuracy: All dates, times, durations, percentages and counts match their pages (see source_verification); Herald-derived figures (50 days, nine days, 34 days, 23 days, two days, the heading/Duration differences, the 13:28/13:40 to 21:15 arithmetic) recomputed and labelled 'by the Herald's count/arithmetic'. Nothing is called 'the same outage' wrongly; the three Aug 17 durations and the 'capacity failures' vs 'misconfigured' wording are attributed to their pages and reported without reconciling, as policy item 6 requires. The VS Code detail is attributed to The Register, and neither GitHub page the Herald read names VS Code (verified). Claim-accuracy sweep of every 'each/all/both/none/neither/only' line against all snapshots: 'each of five August incidents the same four headings' true; 'in all five August incidents ... differ' true; 'neither the executive's post nor the August report names VS Code' true; 'none of the cited sources carries a later vendor account of September 3' true; 'WIRED is the only cited outlet' true. Headline, summary and lead are supported: 'Statements on September 3's AI Outages' matches WIRED's OpenAI statement, SpaceX statement and the vendors' status alerts, and the Overview itself notes one vendor declined to comment. Two subordinate issues, recorded in the corrections record: (1) 'What differs' says 'For September 3 the cited sources show same-day statements to a reporter'; WIRED reports that only OpenAI's spokesperson gave WIRED a statement, Anthropic declined to comment, SpaceX did not respond to WIRED's request for comment (its statement was public) and the Anthropic and xAI wording comes from status pages (correction). (2) Implications say choosing another model avoided the errors 'in both provider-related cases'; 'both' covers only the Aug 27 and Oct 28 entries the article examines, while the same reports also describe provider-side incidents (Aug 20 managed-database provider outage, Oct 5 cloud-provider networking failure) that involve no model choice (clarification). Minor, not filed: the article's wording 'sister sites' for DevClass's 'Our Websites' list; 'two days after' dates GitHub's account by The Register's Aug 19 report, since the RCA's own post date is not in the cited pages.
Overall Assessment: Substantively strong, source-faithful Analysis: 12 quotes verbatim, all figures verified, independence counted strictly and honestly stated, disclosure present, thesis labelled. Approved with a two-entry corrections record (one correction, one clarification) for subordinate statements; headline, summary and lead hold. Featured candidacy: 1 - careful, accurate comparison of how vendors communicate after failures, useful to developers who depend on GitHub and AI coding tools, but with a defined audience and no new incident disclosure