Content Quality: Clear, well-structured News piece (Overview, What We Know, Early Hands-On Reports, What We Don't Know) with appropriate technical depth for a developer audience. Neutral tone throughout; no editorializing or AI self-reference.
Source Verification: All 4 sources fetched successfully (HTTP 200) and snapshot sha256 hashes verified against manifest.json before reading. (1) source-0.html.gz (InfoQ, 'AWS Lambda Now Runs GPU Inference with 48 Concurrent Invocations Sharing One Model per GPU', Steef-Jan Wiggers, Nov 22 2026) — read in full; confirms GA regions, instance families G4dn/G5/G6/G6f/Gr6/Gr6f/G6e, 48-concurrent-invocation cap with one shared model copy in GPU memory, the 9.75%/15% management-fee split, five base images (CUDA/CuPy/PyTorch/vLLM/SGLang), both hands-on test reports (Classmethod Akaike 37/81 instance types, 66x PyTorch speedup; 'xbill' Gemma 4 E2B test, 48 concurrent requests/417 calls/130 vs 3400-3600 tok/s, 484s publish-to-active gap), Terraform/Cloud Control gap, Web Functions exclusion, and the Yan Cui scaling caveat. Every direct quote in the submission (Daniel Abib x2, 'xbill', Yan Cui) was checked verbatim character-by-character against this snapshot: all four match exactly, including the partial Abib quote ('often sit outside your AWS environment...') and the xbill quote, which the article begins at 'a model behind events...' dropping the source's leading capitalized 'For' to fit the sentence grammatically — the quoted substance is unchanged and this is standard fragment-quoting style, not a misquote. (2) source-1.html.gz (AWS docs, 'Run GPU-accelerated Event Functions on Lambda Managed Instances') — read in full; confirms the 'Limits' table (container image 32GB vs 10GB, ephemeral storage up to 32GB vs 10GB, memory 2-32GB, AcceleratorMemorySize 3-48GB valid values 3/6/12/16/24/48, PerExecutionEnvironmentMaxConcurrency default 4 max 48), the GPU instance table, the zero-vCPU-quota-by-default language, and that AWS's docs name vLLM/SGLang/PyTorch specifically and describe Lambda-managed GPU drivers. FINDING: the table gives Gr6f a single 12GB value, not a 3/6/12GB range as the submission states for both G6f and Gr6f jointly -- see findings[0]; filed as a correction. (3) source-2.html.gz (aws.amazon.com/lambda/pricing/) — read in full; located the exact sentence 'Compute management fee: A premium on the EC2 On-Demand instance price for instances provisioned and managed by Lambda, at 15% for CPU instance types and 9.75% for GPU G-series instance types' and the '$0.20 per million requests' request-charge line. This independently confirms, directly from AWS's raw pricing page rather than from the writer's self-report or InfoQ's paraphrase, that 9.75% is correctly attributed to GPU G-series and 15% to CPU instance types (not swapped) and that the $0.20/million-request charge applies in both cases as the submission states. (4) source-3.html.gz (AWS docs, 'Configure Lambda Event Function timeout') — read in full; confirms 'For functions using AWS Lambda Managed Instances, asynchronous invocations and event source mapping invocations (except Amazon MQ and Amazon DocumentDB) support a maximum allowed value of 5,400 seconds (90 minutes)' -- matches the submission's 90-minute claim and its characterization as the same ceiling reported in the Herald's September 2026 Managed Instances timeout article (src/content/articles/2026-09/21-aws-lambda-extends-timeout-to-90-minutes-on-managed-instances-six-times-the-previous-limit.md, confirmed to exist). No suspicious_patterns flagged in manifest.json for any of the 4 sources; no prompt-injection text found in any snapshot. 48-concurrent-invocations cap: independently verified against AWS docs table (PerExecutionEnvironmentMaxConcurrency default 4, maximum 48) and InfoQ's matching language -- accurately sourced. 'Share one model' framing: independently verified against InfoQ ('Concurrent invocations in one environment share a single copy of the model in GPU memory, up to 48 per environment') and AWS docs' concurrency-within-an-execution-environment description -- accurately sourced, not an overstatement. Anthropic disclosure: not required - the story subject is AWS Lambda GPU support; Anthropic/Claude is not the subject, headline, summary, lead, or credited in a passage of its own.
Factual Accuracy: All claims, quotes, figures and derived numbers in the body were checked against the raw source text (not the writer's self-report). One subordinate inaccuracy found and corrected below (Gr6f fractional-GPU memory range); every other claim checked -- including the fee percentages, concurrency cap/default, container/storage/memory/accelerator limits, five base images, quota-zero default, 90-minute timeout ceiling, GA scope, and both hands-on test figures -- traces exactly to the cited source text.
Overall Assessment: Well-sourced, accurately quoted News piece on a genuinely new AWS Lambda feature. All four sources were read in full from verified snapshots; every direct quote matches its source verbatim; the two management-fee percentages (9.75% GPU vs 15% CPU) were independently confirmed against AWS's raw pricing page text, not just the writer's explanation; the 48-concurrent-invocation cap and 'share one model' framing both trace exactly to AWS documentation and InfoQ. One subordinate factual error was found (Gr6f's fractional-GPU memory range overstated) and is addressed with a corrections record. Featured candidacy: 1 - solid, well-verified infrastructure news for a defined cloud/serverless developer audience; not a broad-audience or industry-changing story.