What do MASS (authoritative learned state) and WorldCast (distributed clients exchanging state) show about moving the shared state out of the video generator, and what do their evaluation protocols leave untested?

MASS shows, on a matched Snake benchmark at 128x128 and 128 ticks, that a typed learned state recovers more objects (parser 0.764 versus at most 0.128 for video baselines) and has zero cross-view disagreement by construction, but its structural-validity score is worse than independent per-view baselines and its local prediction decays when other players inputs are missing. WorldCast shows on Counter-Strike 2 that exchanging camera-aligned player state raises the rate at which other players are drawn (0.817 versus 0.056 and 0.074 for retrained baselines), but evaluates lockstep clients without network delay or faults and with ground-truth states for most whole-round results. Neither releases code or weights that were found. Curator correction 2026-10-24 (discovery audit record C3-B1085-01, curator-reviewed C3R-0691; excerpts re-located by independent curator GET 2026-10-24): Correction. C3-B115-W23-06 ends 'Neither releases code or weights that were found.' That is wrong for WorldCast and was already wrong when the record was checked on 2026-10-22. A public code repository (Apache-2.0 licence file, worldcast/ package, tools, configs) has a first commit dated 2026-10-10T13:39:09Z, and the Hugging Face model repository ZiyangYe/WorldCast holds the four-step generator, state model and depth head from its first commit dated 2026-10-10T13:26:04Z. Three training checkpoints were added on 2026-10-22 (stage2 15:31Z, stage2s 18:14Z, stage3 21:01Z) and a demo, data preparation and the paper's metrics on 2026-10-23T08:58:16Z, so the training and evaluation parts post-date the record's check, but inference code and weights did not. The arXiv paper and the project page do not link them (the project page Code and Demo buttons still point to #bibtex on 2026-10-24), which explains the miss. The MASS part still holds on 2026-10-24: the only repository found under the MASS authors' organisation is a project-page repository described as 'Project page for MASS'. The record's other claims (parser 0.764 versus 0.128, 128x128 over 128 ticks, WorldCast 0.817 versus 0.056 and 0.074, no network-fault study) were re-read against the paper HTML and match. Corrected statement (independent audit W4-270, 2026-10-24) to use in place of the record's wording: MASS: no code or weights found (only a project-page repository, checked 2026-10-24). WorldCast: a public GitHub repository (Apache-2.0 licence file) and Hugging Face model repository ZiyangYe/WorldCast hold first commits dated 2026-10-10, before this record's 2026-10-22 check, with the generator weights in the first Hugging Face commit; the paper and project page (Code button points to #bibtex) do not link them. Training checkpoints (2026-10-22) and the demo and metrics (2026-10-23) came later. Inference: public visibility on 2026-10-22 is inferred from commit dates; weights were not run and model-card terms not read. Curator qualifications: Excerpted evidence located (one table-row excerpt mismatched by formatting; one excerpt-less row); MASS leads 6 of 7 metrics but not 'Invalid' (structural validity) - kept as the salient caveat. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://arxiv.org/html/2608.06257v2",
      "note": "candidate evidence excerpts"
    },
    {
      "url": "https://arxiv.org/html/2610.12412v1",
      "note": "candidate evidence excerpts"
    },
    {
      "url": "https://ziyang-ye.github.io/WorldCast-Page",
      "note": "candidate evidence excerpts"
    }
  ]
}