Source Verification: Read from the stored snapshots (sha256 re-verified against manifest.json for all 11 stored files): source-0 (AWS News Blog, Daniel Abib, 11 NOV 2026) confirms GA with no additional service fee, Rust, 127.0.0.1:9297, SigV4 for local-process mode, default anonymous telemetry that can be turned off, no ready-made Docker image, HTTP 501 for unimplemented operations, independent security agent reviewing changes adversarially, and 'It doesn't evaluate IAM policies the way AWS does'; source-1 (aboutamazon.com, Matthew Goldberg, Sr. Principal Engineer, 'Published: November 11, 2026') confirms 'The emulator was written entirely by autonomous agents. Our team hasn't written or reviewed any of the code', the dek 'Thousands of autonomous AI agents wrote every line of code in AWS Emulator', 'hundreds of agents concurrently', 438 AWS API Models, Kiro/SQS/15 minutes/'the rest of AWS' over a weekend, the Anthropic C-compiler 'lightbulb moment' passage, three teams of agents, the merge gate, and CC0 as 'a public dedication under the Creative Commons CC0 license' with the build-history insights as 'the main motivation'; source-2 (docs, What is AWS Emulator, Limitations) confirms the headline claim in the docs' own words: 'The emulator doesn't evaluate AWS Identity and Access Management (IAM) policies and resource-based policies the way AWS does' (the wording is 'doesn't evaluate', not 'partially'), plus no separation by Region or account, in-memory resources deleted on stop, and no testing for long-running use or load testing; source-3 (docs, Fidelity) confirms the verbatim quotes 'Today, the emulator matches AWS most closely for control plane APIs' and '...varies by service', HTTP 501 with error code NotImplemented, that scores are published with each release, that each release 'names the AWS Emulator commit and the version of the fidelity tests', and that tests can be added, rebuilt or retired; source-4 (GitHub page) plus the raw README and LICENSE read through the GitHub API confirm CC0-1.0 (GitHub licence key cc0-1.0), no PRs, 100 GB for target/, embedded mode, port 9297; source-5 (release v1.0.322) confirmed through the API: published_at 2026-11-11T00:32:13Z, body is a list of fixes with no fidelity-test version, asset scores.json.gz; source-6 (evaluation suite) confirms the repository and manifest.jsonl; source-7 (api-models-aws) confirmed; source-9 (InfoWorld, Anirban Ghoshal, Nov 11/12 2026) confirms the Chaturvedi, Pratt and Patel quotes verbatim with the stated titles, 'HTTP 503 errors' for operations 'not been implemented or verified' (attributed by InfoWorld to AWS, 'it said'), and 'supports all 426 AWS services' ('it added'); source-10 (SiliconANGLE, Mike Wheatley, Nov 11) confirms 'dozens of autonomous agents' and Holger Mueller's 'a digital twin of its public cloud'; source-11 (The Register, Simon Sharwood, APAC Editor, 'thu 12 Nov 2026 // 05:47 UTC') confirms 'over 1,300 items on that page' and 'appears, as ever, to encourage more use of its cloud'. The New Stack (source-8) was HTTP 403 (Cloudflare challenge) in the stored snapshot and the manifest has file=null; Archive.org has no capture (wayback availability API returned no snapshot), so the live page was read as raw text through a reader proxy (https://r.jina.ai/https://thenewstack.io/aws-emulator-ai-agents/, Published Time 2026-11-11T17:30:00+00:00, Frederic Lardinois, sha256 of the fetched markdown 4ddf87ce524f644998d7b13b0273211b6e4860e00088b0bc1c5d27189f9b529e); not a summary, and every New Stack claim in the article was checked against that raw text: Callaway quote 'The challenge is that they get out of date very quickly', 'may raise some questions about the future of LocalStack', 'entirely built by a team of agents', AWS 'wasn't sure it could actually copyright the work of these agents', '438 public AWS API models covering more than 200 services', Kiro described as 'AWS's agentic coding tool', re:Invent 'starts Nov. 30'. Prompt-injection scan: suspicious_patterns is null for every source; nothing to judge. Headline claims: (1) 'Releases ... Under CC0' holds: GitHub licence API and LICENSE are CC0-1.0, Goldberg says the software is released as a public dedication under CC0; but the README says it is 'licensed under the CC0 1.0 Universal (CC0-1.0) License, except the code in the vendor/ directory, which is under the Apache License 2.0' and does not use 'public domain', whereas the article's Overview says the source 'is dedicated to the public domain under CC0, according to the project README' (misattributed wording, vendor/ exception omitted) - subordinate, in the corrections record. The release date Nov 11 matches Goldberg's 'Published: November 11, 2026', 'Today, we released', the AWS blog's 11 NOV 2026 and InfoWorld/The New Stack 'Wednesday'; the article gives v1.0.322's time as UTC. Note the GitHub repository carries 12 earlier tagged releases from 2026-11-07 (v1.0.250 onward); the article's 'released on November 11' rests on AWS's own announcement date and is not contradicted. (2) 'Saying Agents Wrote It Entirely' is Goldberg's own sentence and AWS's own page ('100% written by autonomous AI agents'); the article attributes it to Goldberg and the AWS posts, not to the outlets. (3) 'Its Docs Say It Does Not Evaluate IAM Policies the Way AWS Does' matches the docs Limitations wording. Conflicts: 501 (AWS blog, docs) vs 503 (InfoWorld) is reported as a disagreement, attributions correct (InfoWorld itself attributes the 503 statement to AWS); surface size 438 (Goldberg, New Stack) / 426 (InfoWorld, 'according to AWS') / over 1,300 items (The Register) / 436 (Herald count) all attributed correctly; agent numbers thousands (dek) / hundreds (body) / dozens (SiliconANGLE) correct; CC0 rationale (New Stack: copyright uncertainty; Goldberg: build-history insights) correct and not reconciled. Herald-computed figures recomputed from raw data: scores.json.gz downloaded from the v1.0.322 release (top-level keys commit, schemaVersion, scoredAt, units; scoredAt 2026-11-11T00:09:58Z; no fidelity-test version field): 42,937 units; behavioral 32,402 (20,364 pass, 11,958 fail, 80 skipped; 62.85%, about 63%); wire 10,535 (3,879 pass, 6,656 fail); single-service behavioral counts S3 763/924, DynamoDB 656/801, SQS 286/304, Lambda (AWSGirApiService) 293/616, EC2 (AmazonEC2) 576/1,300, all exactly as stated. api-models-aws: git tree API (not truncated) lists 436 service-model JSON files under models/ (437 .json blobs including .scripts/overrides.json), matches. Evaluation-suite manifest.jsonl: 29,676 lines/rows, matches. The article labels all of these as the Herald's own tallies, says the docs do not define 'wire' (the fidelity page and docs/TAXONOMY.md do not define a 'wire' kind) and says the Herald did not map entries to AWS's per-service scores. Reviewer cross-check: by test identifier, 29,255 of the 32,402 behavioral scores entries match a manifest check_id, 3,147 entries (including identifiers beginning '_smoke.') have no manifest row and 421 manifest rows have no scores entry, so the 29,676-vs-32,402 comparison is between different sets of tests, not only different sizes (clarification filed). Independence: four outlets each with its own text, quotes or reporting - The New Stack (interview with AWS's Callaway; independent of the vendor's own pages though built on an AWS interview), InfoWorld (three named analysts), SiliconANGLE (Constellation Research analyst plus a second AWS post), The Register (own item count and commentary); The Register and SiliconANGLE lean heavily on AWS's own posts. No claim is attributed to an outlet that does not say it (The Register does not mention CC0, agent authorship or LocalStack, and the article does not attribute those to it). 'None of the four outlets cited reports running the fidelity tests itself' checked in all four raw texts and holds. Other quantifier/absence sentences from the grep (does not publish a ready-made image, every request signed with SigV4 in local-process mode, each release names the fidelity-test version) checked against the sources and hold. Rule 15 / contributor-model disclosure: Anthropic disclosure: required - Goldberg's post names Anthropic as the inspiration (researchers at Anthropic showing agents could build a C compiler) in a passage of its own; the one-sentence disclosure is present at the end of the Overview ('The Herald's contributor model is built by Anthropic, which Goldberg's post names in describing how the project began'), is neutral and strictly attributive, and the body's description of Anthropic's work is attributed to Goldberg and quotes only 'was a lightbulb moment'. Allowlist: all 12 hosts already listed (aws.amazon.com, aboutamazon.com, docs.aws.amazon.com, github.com, thenewstack.io, infoworld.com, siliconangle.com, theregister.com); no additions. Internal link: /article/2026-11/08-minio-localstack-boundary-and-portainer-... exists, is cited only for 'covered LocalStack' (its article does cover LocalStack), and repeats nothing from its corrections record. Word count 1,096 of the 1,200 News ceiling. Featured analysis note: not applicable (News).