{
  "id": "rather-v1-coverage-audit-20261016",
  "checked_at": "2026-10-16T02:49:18.669607+00:00",
  "scope": "Targeted primary-source review, not an exhaustive proof of novelty. Search queries included exact model names plus forward dynamics, Bridge, WBench and return benchmark.",
  "sources": [
    {
      "title": "WorldMemBench, submitted 13 October 2026",
      "url": "https://arxiv.org/html/2610.19063v1",
      "path": "worldmembench.html",
      "sha256": "557778fc8ae1c952530b569d7ad8771917f058466f571dd059600413a5b64041",
      "bytes": 597631,
      "observation": "Reports repeated revisits and camera-following results; Matrix-Game 3 uses the 3-step configuration, seed 42, and eight GPUs. Establishes existing memory benchmarking, not the planned base-versus-distilled repeated-seed single-GPU comparison."
    },
    {
      "title": "WBench author results and Hub integration",
      "url": "https://huggingface.co/datasets/meituan-longcat/WBench/discussions/9",
      "path": "wbench-discussion.html",
      "sha256": "3c939c2cfb11ca489936e211600aebdac46f804b870025b0dfde95418e312ee8",
      "bytes": 381707,
      "observation": "Authors report navigation scores for MatrixGame2 and MatrixGame3. Rather already imports WBench; neither models nor camera control are previously unevaluated. The proposed hold/return diagnostic changes the protocol and must never be labeled official WBench."
    },
    {
      "title": "NVIDIA Cosmos 3 project",
      "url": "https://research.nvidia.com/labs/cosmos-lab/cosmos3/",
      "path": "cosmos3.html",
      "sha256": "5cc16f4fde4a2b08785be953106efed37b2d199d692b140f277a4fb2e11ef978",
      "bytes": 75946,
      "observation": "Demonstrates forward dynamics, policies and inverse dynamics. This establishes existing robotics simulation capability and prior evaluation work, not independent repeated grasp-outcome agreement on the prepared simulator cohort."
    },
    {
      "title": "Cosmos 3 Nano pinned model card",
      "url": "https://huggingface.co/nvidia/Cosmos3-Nano/resolve/e59a53c25979a090fa8706c9acc0c254a6e89b92/README.md",
      "path": "cosmos3-nano.md",
      "sha256": "bbe6a5b0e3c09ae08ba1df06d35859fa2fd5726c1edc22a43788611f126caad5",
      "bytes": 43142,
      "observation": "Declares commercial availability, native WidowX action support and model benchmarks. Rather evaluates the exact released forward-dynamics configuration, with source claims kept separate."
    },
    {
      "title": "Cosmos 3 Edge pinned model card",
      "url": "https://huggingface.co/nvidia/Cosmos3-Edge/resolve/344d602b128d1bbdacb43b08d0a3626f46343e29/README.md",
      "path": "cosmos3-edge.md",
      "sha256": "eed039d1cf7a7a9c5a77b8a2abbd31ee7aed60e30f2465c573ae8afdfedbe6bd",
      "bytes": 53043,
      "observation": "Documents an August 25 checkpoint/runtime/result update. Pinning matters: results on a floating model name cannot establish the behavior of the exact released configuration used here."
    },
    {
      "title": "Independent Cosmos 3 Nano Bridge fine-tune",
      "url": "https://huggingface.co/anshks/cosmos3-nano-bridge-fd-wm/resolve/d24d36b2ae9e802042fe60e5d6e30a9e98a117c9/README.md",
      "path": "bridge-finetune.md",
      "sha256": "984f651874bde856688ba5b7708a4c525f1d2876ae2cf5f60a40aa677b0fbebe",
      "bytes": 2408,
      "observation": "Describes a post-trained multi-view Bridge forward-dynamics checkpoint, held-out image metrics and policy-ranking use. This is relevant prior functional work, with a different checkpoint, view setup and protocol."
    },
    {
      "title": "WorldFoundry Matrix-Game 3 adapter",
      "url": "https://openenvision.org/WorldFoundry/docs/guides/supported-models/matrix-game-3",
      "path": "worldfoundry.html",
      "sha256": "0cc0fdc5529f6dcf0dc3c35dccf5e5e71a9b3c70a878f312b4e20f3a27c1c7e9",
      "bytes": 108043,
      "observation": "Documents a unified action-capable runner and WorldModelBench target; the checked page labels runner parity pending. It is relevant infrastructure, not proof of absent evaluation elsewhere."
    }
  ],
  "proposed_studies": [
    {
      "id": "rather-camera-return-v1",
      "models": [
        "matrix-game3-base",
        "matrix-game3-distilled",
        "matrix-game2-universal"
      ],
      "existing_evidence": "WBench navigation results, WorldMemBench revisit results, native demonstrations and third-party runner integration already exist.",
      "gap_in_checked_sources": "A matched repeated-seed comparison of exact base/distilled/older configurations on a small transparent hold-turn-return-hold grid with a frozen baseline, action tracking coverage, all errors and measured single-GPU compute.",
      "commercial_decision": "Does the higher-cost base configuration improve controllability or return consistency enough to justify its measured operating cost under this explicit deployment setup?",
      "limits": "This diagnostic is narrower than WorldMemBench and is not a new general memory benchmark, full WBench result or proof of real-time serving."
    },
    {
      "id": "rather-widowx-lift-v1",
      "models": [
        "cosmos3-edge-robot-action",
        "cosmos3-nano-robot-action",
        "cosmos-predict25-robot-action"
      ],
      "existing_evidence": "NVIDIA forward-dynamics/policy work and an independent Bridge fine-tune with image metrics and policy-ranking use already exist.",
      "gap_in_checked_sources": "A matched repeated comparison of these exact released configurations on the freshly prepared 32-episode grasp cohort, reporting judge-estimated success/failure fidelity against simulator object-state truth, uncertain visuals and generation cost.",
      "commercial_decision": "Which released learned world preserves task outcomes along specified gripper trajectories on this explicitly simulated WidowX setup?",
      "limits": "Realized trajectory conditioning does not evaluate actuator tracking, closed-loop policy utility or real-robot transfer. Generated-video evaluator bias must be audited separately from reference calibration."
    }
  ],
  "claim_policy": "Advertise new Rather measurements and the exact comparison that was absent from the sources checked. Do not claim first-ever evaluation, no competitors, or exhaustive coverage."
}