Content Quality: News, 1,059 words (range 400-1200), title 143 characters. Clear structure (Overview, What was released, Hardware requirements, Results and who ran them, Verified or not, What we don't know). Neutral; vendor claims attributed to Runway/The Robot Report/The Decoder; the Herald's own counts (ten repos, 21 policies, ordering of rows) are labelled as the Herald's.
Source Verification: Snapshots read from disk (all status 200, suspicious_patterns null for all eight): source-0 The Robot Report (byline Brianna Wessling, dated November 12, 2026, says Runway released Praxis-1 'yesterday'): confirms open weights under Apache 2.0, 'including for commercial use', GitHub/Hugging Face/vLLM-Omni/Reactor, runs on H100/B200/RTX 5090/RTX PRO 6000/Jetson Thor, 27B MoE diffusion transformer, 46.2% across 1,200 episodes vs 42.9/36.8/28.0, 'ranks fifth overall', and Andy Chen's 15-of-15 quote verbatim. source-1 The Decoder (Tomislav Bezmalinovic, Nov 14, 2026): 'Runway ran all the tests itself, and the results haven't been independently verified yet.' verbatim; 'behind two AI agents and two robot models with different architectures' verbatim; top of the World Action Model category; 88 vs 80 on a Franka arm; 95 hours / 80 percent vs 52.4 percent on RT-1; download from Hugging Face 'after signing up and agreeing to the access terms'; 'is still an open question'; limits paragraph. source-2 runway.com launch post dated November 11, 2026: 'outperforms every open-weight VLA', 'including for commercial use', 'Together, Runway and NVIDIA are opening frontier technology stacks' (the article quotes the substring 'Runway and NVIDIA are opening frontier technology stacks'). source-3/4 Hugging Face praxis-1-base and praxis-1-droid: tag 'License: apache-2.0', License section ('Apache License 2.0. The frozen dependencies (Wan2.2 VAE, Qwen/Qwen3.5-2B) are distributed separately under their own licenses.'), the access notice, 'so it does not run as a policy as is', 'Inference needs one 80 GB GPU.'; the conditions are visible only after login (article says so). source-5 GitHub README: 'See the license file of each set of weights for the weights.', 'Wan2.2 and Qwen3.5 have their own licenses.', quick start needs one CUDA GPU with at least 80 GB, NVFP4 'has not been run in closed loop on an SM120 GPU yet', 'SM120 GPUs such as RTX 5090', 'a 32 GB RTX 5090 should fit (the peak is not measured yet)'; latest commit 2026-11-12T22:14:39Z confirmed via the GitHub API (not shown in the snapshot text). source-6 leaderboard (read as raw HTML, data embedded as escaped JSON/HTML): 21 policy rows (ranks 1-17 with ties at 6, 7 and 11; the second table also lists 21); rank 5 Praxis-1 WAM 555 / 1200 at 10 eps/task, 46.2%; above it AMP1 86.0% and SenseNova-RoboRSI 57.1% (both Agentic LM), Lakeside 51.0% (LWM), BiMind v0.5 49.2% (VLA); Lakeside and BiMind v0.5 carry the lock 'Closed source: model weights not publicly available'; WAM/VLM+WAM rows: Praxis-1 46.2, FLUX 3 Action 42.9, Atomic-WAM 39.6, OASIS WAM 39.0, Cosmos3-Nano-Policy 36.8, DreamZero 25.7, Cosmos3-Edge-Policy 22.9, so highest as stated; Praxis-1 detail panel: Verified badge, Submission badge, 'Submitted Nov 5, 2026', tooltip 'RoboLab organizers were able to independently run this policy and/or cross-check its reported results', description 'fine-tuned on public DROID demonstrations'; no last-updated time in the page text (HTTP Last-Modified header Nov 11, 2026 15:29 GMT); page footer NVIDIA Corporation. source-7 Hugging Face list API: ten model entries under runwayml, all license:apache-2.0, createdAt Nov 3-12 2026; the org overview API gives numModels 10. The cited list URL carries NO gated field (identical JSON to the stored snapshot, sha256 57ab93a1...); per-repository calls (https://huggingface.co/api/models/runwayml/praxis-1-<name>) and the list with expand[]=gated return gated: 'auto' for all ten, so the gating claim is true but is not visible at the cited URL (clarification filed). The LICENSE file of praxis-1-base resolves unauthenticated and is the Apache 2.0 text (11,357 bytes, same size as the GitHub LICENSE); the model README is behind login. Live re-fetch of the leaderboard and launch post matched the stored snapshots' text. HEADLINE TESTS: 'Apache 2.0 Weights' - repos' license fields, card text and LICENSE file are Apache 2.0, and the body states the frozen dependencies (Wan2.2 VAE, Qwen3.5-2B text encoder, 'not in this repo') are separately licensed, so the headline does not overstate; 'Behind a Hugging Face Access Gate' - all ten repos are gated 'auto' (accept conditions and share Hugging Face information), body describes this and says the conditions were not read; 'NVIDIA's RoboLab-120 Board Lists It Fifth of 21' - the board is an NVIDIA Research site (NVlabs/RoboLab homepage, NVIDIA Corporation footer), count 21 and rank 5 recomputed from the raw rows. ALLOWLIST: runway.com and research.nvidia.com were the only warning. Both confirmed as official sites (verified runwayml Hugging Face and GitHub organisations list runwayml.com, which redirects to runway.com, and the praxis-1-oss README links runway.com; nvidia.com/en-us/research links research.nvidia.com, the NVlabs organisation and the NVlabs/RoboLab repository link it) and read as raw text; both added to config/source_allowlist.txt with justification comments limited to first-party statements, in the same commit. DISAGREEMENT: The Decoder 'results haven't been independently verified yet' (Nov 14) versus the board's Verified badge with tooltip 'organizers were able to independently run this policy and/or cross-check'; the board offers no methodology page beyond the tooltip and legend (it also says submissions are paused), so no source explains the difference; the article reports both with attribution and notes the 'and/or' and that The Decoder's sentence sits in a report that also covers non-board tests. INDEPENDENCE: The Robot Report (WTWH Media trade outlet, staff byline) and The Decoder are independent outlets, though each draws mainly on Runway's statements/report; the leaderboard is a third-party board whose Praxis-1 row is a developer submission with an organizer 'Verified' badge. NVIDIA ties: the article's sentence is supported (Runway's post: 'Together, Runway and NVIDIA are opening frontier technology stacks'; The Robot Report lists NVIDIA hardware). HF list API as source: acceptable as a registry listing for repo count and license tags, but not for the gated flag (see above). Anthropic disclosure: not required - the article does not mention Anthropic or Claude (the Decoder page's sidebar headlines are not cited content). Claim-accuracy tests run: absence/quantifier sentences ('None of the cited sources reports an outside replication...') hold against all eight snapshots; dates consistent (post Nov 11, Robot Report Nov 12, Submitted Nov 5, README commit Nov 12 22:14 UTC); every quoted string was matched programmatically against the page it is attributed to.
Factual Accuracy: Headline, summary and lead are each backed by cited sources and recomputed from raw data. Two subordinate statements are incomplete: (1) the ten repos 'each flagged as gated' is true (gated: auto) but the cited list URL does not carry the flag; (2) 'These statements describe different configurations' is the Herald's reconciliation: neither The Robot Report nor Runway's post says which configuration their RTX 5090 statement covers, and the README says the NVFP4 path has not been run in closed loop on an SM120 GPU and the 5090 peak memory is not measured. Both filed as clarifications.
Overall Assessment: Substantively strong and carefully hedged News article; headline claims (Apache 2.0 weights, gate, fifth of 21 on NVIDIA's board) all hold against raw sources, the Verified-versus-unverified disagreement is reported fairly, and the two subordinate gaps are covered by a corrections record. Featured candidacy: 1 - accurate coverage of an open-weight robot foundation model release for a defined robotics developer audience