jwz8819/everfold-bench

Everfold Bench

Frames, judge verdicts, scores and human labels behind the paper Everfold: Do Generative World Models Give Two Players the Same World? (Wanzhen Jiang, 2026).

• Code: https://github.com/jwz8819/everfold
• Paper: arXiv link to come (a PDF is in the code repository under paper/).

Everfold is an open multiplayer world with an authoritative game server. Five probes script bench players through it and ask a renderer for one frame per step. A blind vision-language judge (Claude Opus 5.5) marks each ground-truth claim as present, absent or unclear. The probes are cross-view agreement, revisit retention, cross-player state propagation, identity persistence and shared-structure state change. This dataset holds every judged frame of the paper's runs, so that the scores can be checked, re-judged with another judge, or compared against new models.

Contents

Path	What
runs/<condition>-seed<n>/	One benchmark run: one condition, one seed. 190 runs and 1,576 frames in all.
labels/, labels-v2/, labels-v2-human2/	Human labels from two labellers (human1, human2) on sampled frames, with agreement reports.
runs/quests/, runs/quests-scripted/	Content-grounding runs: quests written by a language model (or by scripted personas), checked against the world.
runs/eden-calibration/	Controls for the game-locked Eden model (frames and the script that made them).
results-v2-wording.md	Per-condition means with 95% bootstrap intervals (seed is the unit), as reported in the paper.
results-v1-wording.md, before-after-v1-v2.md	The same under the original claim wording, and the change between the two (see below).

Inside a run folder

File	What
frames/<step id>.png	The frame the renderer returned for that step.
reference/<step id>.png	The engine's top-down render of the true state at that step (a pipeline check, not a world model).
steps.jsonl	One line per step: probe, player, pose, the prompt sent to the model, and the ground-truth claims.
verdicts.jsonl	The judge's answer for each claim of each frame: present, absent or unclear.
pair.json	Cross-view consistency: the pair judge's view of both players' frames (same, viewpoint, why).
metrics.json	Probe scores for the run, plus renderer cost and latency under renderer and judge settings under judge.
prompts.md, report.md	The prompts as text, and a readable report of the run.
v1/	Present where a run was re-judged: the verdicts and metrics under the original claim wording.

Conditions

Folder prefix	Condition in the paper	Seeds
abot-image-drive, lingbot-image-drive	Hosted realtime world model (ABot-World-0, LingBot-World-2 on fal), one session driven between poses	20, 12
abot-image-drive-x0.5, -x1.5	Sensitivity check: drive timing halved or increased by half	5 each
abot-image, lingbot-image	Same models steered by prompts alone	8 each
abot-image-2p, lingbot-image-2p	One session per player (cross-view only)	20 each
eden-image, eden-image-drive	Eden Live 14B Game on fal (game-locked)	1, 10
flux-oracle	FLUX.1 schnell drawing each frame from the true-state prompt (stateless)	20
lucy-greybox, lucy-greybox-2p	Decart Lucy Restyle restyling the engine's greybox clip, one clip per probe family or per player	20 each
greybox	The engine's own 3D render	20
reference	Top-down reference render (pipeline check)	1

Seeds are 42, then 1, 2, 3, and so on.

Claim wording, v1 and v2

After the first multi-seed round, three claims were loosened, based only on the stateless oracle failing them on frames drawn from the true state: "a tree stump" instead of "a fresh tree stump"; the shrine as its platform alone; people by outfit only. Every saved frame was re-judged. The paper reports v2. The v1/ folders and results-v1-wording.md keep the originals. No conclusion changes between the two.

Human labels

Two labellers answered the judge's question blind on sampled frames. Neither saw the judge's answers. human1 knew the study design; human2 worked from the one-page brief in labels-v2-human2/README.md. On 50 frames (319 claims), present-vs-not Cohen's kappa was 0.82 between the humans and 0.65–0.66 between each human and the judge.

Provenance and terms

Frames come from third-party models called through their APIs in November 2026: ABot-World-0, LingBot-World-2, Eden Live 14B Game, Decart Lucy Restyle and FLUX.1 schnell (all via fal.ai), plus the engine's own renders. The judge is Claude Opus 5.5. This dataset's annotations, scores and engine renders are released under CC BY 4.0. Model outputs may also be subject to the respective providers' terms. Scenes are procedurally generated game worlds and contain no real people or personal data.

Citation

@misc{jiang2026everfold,
  title  = {Everfold: Do Generative World Models Give Two Players the Same World?},
  author = {Jiang, Wanzhen},
  year   = {2026},
  note   = {arXiv preprint, to appear}
}