Provenance Record

Well-organized News piece using the Overview / What We Know / What We Don't Know structure. Technical claims (30B MoE / 3B active params, speculative decoding with DFlash/DSpark, harness support, deployment guides, license, availability channels) are precise and consistently sourced inline. Not human-requested.

All 3 sources read in full from local gzip snapshots in sources/2026-08/nvidia-releases-nemotron-35-lightning-a-30-billion-parameter-open-model-built-for-high-volume-ai-agent-execution/ (manifest sha256 values verified against the decompressed content by recomputing sha256sum — all three matched). source-0.html.gz (NVIDIA developer blog, 'NVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agents', Aug 11 2026, 200 OK) confirms: 30B MoE / 3B active params, 'execution layer' framing, Nano/Super/Ultra family context, 'harness-optimized' training for OpenClaw and Hermes Agent, speculative decoding / multi-token prediction / DFlash+DSpark draft models, OpenMDW-1.1 license (though the article's quoted wording is not verbatim — see findings), DGX Spark-to-datacenter deployment, Hugging Face/ModelScope/build.nvidia.com/OpenRouter availability, vLLM/SGLang/TensorRT-LLM deployment guides, and the PinchBench 86% accuracy / 10,000-tasks-30%-faster-than-Qwen3.6-35B benchmark. source-1.html.gz (NVIDIA developer blog, 'Route AI Agent Workloads Across Models with NVIDIA NeMo Switchyard', Aug 11 2026, 200 OK) confirms: switchyard-libsy provider-agnostic SDK description, tuning-free routers (LLM classifier, stage router, escalation router) and the tunable prefill router, the LangChain-benchmarked 74% cost reduction / 7% frontier calls / ~6-point accuracy tradeoff figures for Nemotron 3.5 Lightning + Claude Opus 4.8 escalation routing, and the Cognition FrontierCode Main benchmark (50.6% accuracy / $3.11 mean cost / within 2.8 points of Opus 5 / ~28% lower cost) for Opus 5 + Kimi K2.7 routing. This source's own partner list also names Cognition directly, which conflicts with how the article attributes that fact solely to SiliconANGLE — see findings. source-2.html.gz (SiliconANGLE, 'Nvidia releases Nemotron 3.5 Lightning and NeMo Switchyard to give enterprise AI capability options', by Kyt Dotson, updated Aug 11 2026, 200 OK) confirms the Nano/Super/Ultra family launched in December, the 4x-speed/30%-faster-task-completion figures, the CodeRabbit $85/one-epoch/two-hour router agent example attributed to VP of Generative AI Kari Briski, and the ecosystem partner list including Boomi, Cadence, Classmethod, Cognition AI, Kong, LangChain, Nous Research, and Siemens. No paywall or fetch failures; no archive.org fallback needed.

Every specific figure I checked (parameter counts, benchmark percentages, dollar costs, partner names, license name, deployment tooling) traces to at least one cited source and the wording is materially accurate. Two subordinate issues found on close read: (1) the phrase quoted as "as permissive as possible" is a paraphrase of the source's actual "as permissively as possible"; (2) the article implies Cognition AI was named as a partner only by SiliconANGLE, but NVIDIA's own NeMo Switchyard blog also lists Cognition in its partner rundown, so the 'SiliconANGLE additionally names' framing overstates that source's exclusivity on an otherwise-true fact. Neither issue touches the headline, summary, or lead paragraph, and both are single, isolated, honestly-correctable specifics rather than a pattern of fabrication.

Substantively solid, well-sourced News piece on a same-day product launch. Two minor, non-headline issues (a non-verbatim quote and an overstated source-exclusivity claim about a partner list) warrant APPROVE_WITH_CORRECTIONS rather than a clean APPROVE; neither rises to the level of requiring rejection.