The 3D world is not evidence
Hand-authored motion re-created in your browser. The repo holds rendered clips only — no per-frame poses — so this is a different seed, solver, and pile. It earns no numbers.
ICLR 2026 · 2nd Workshop on World Models · accepted · corrected July 2026 · 3D companion
Pull time
backward.
The clip page watches 48 recorded frames through a window. This page hands you the world: a domino run you can orbit while you drag time in both directions. Forward, the motion spends its energy and goes still. Backward, a settled pile climbs back onto its edges and a resting ball gathers speed for nothing — dissipation running in reverse, which is the entire paper.
The world below is a re-simulated lookalike built for intuition — the repo holds rendered clips, not object poses. The recorded clips and their exact losses stay in the clip lab, and the paper’s corrections stay in the evidence receipt. Corrected paper ↗
This world needs WebGL and a connection to the Three.js CDN.
Everything measured lives in the 2D clip page, the paper, and the receipt below.
Everything measured lives in the 2D clip page, the paper, and the receipt below.
Rehearsal · re-simulated in your browser — not the recorded clip
F 01 / 48
ORDER FORWARD
CTX 8
Reversal roomorbit: drag · zoom: scroll
The run is playing. Watch where the energy goes.
9 dominoes · 48 frames @ 12 fps · hand-authored motion
Time transport← → scrub · space play · d flip
PLAYED 01 / 48 · SOURCE 01 / 48or drag the floor strip
Motion · proxy
Pile height · proxy
Geometric proxies of the staged motion — watch them refill when time runs backward. Not model outputs.
Context windowthe experiment’s shape
GIVEN 08 · PREDICTED 40context frames N
CONTEXT 14 · the awkward cell. In the corrected randomized-domino rerun, V-JEPA2’s separation flips sign exactly here — reversed becomes the easier prediction. The red band on the rail pin marks this boundary. Exact values sit in the clip lab.
Model lensqualitative only
Arrow testscore 0 / 0
A hidden direction plays in the world. Orbit around it — being able to inspect from any angle and still be unsure is the point — then call it.
CORRECT
Free play, world focused: ← → scrub · space play/pause · D flip direction · 1–4 cameras
Provenance
- What is exact
- Every loss value, context sweep, rerun mean, and withdrawal on this page is quoted from the corrected paper and the three checked-in 48-frame PyBullet sequences. The clip lab plays the real mp4 files and attaches the real numbers to them, per scene, model, and context.
- What is staged
- The 3D world above. The repository contains rendered clips only — no per-frame poses — so the dominoes you scrub are hand-authored keyframe motion: a different seed, solver, and pile than any recorded run. It demonstrates the intuition and produces no loss values. Its energy bars are geometric proxies of that staged motion, not model outputs.
Exact clip lab · the evidence
The recorded sequences,
The recorded sequences,
with their losses attached.
Six direction clips from three checked-in 48-frame PyBullet sequences. Each loss value below was read from video_000 for the selected scene and context, on the footage playing beside it. This lab is the evidence; the 3D world above is not.
bouncing-forward.mp448 frames · 12 fps · PyBullet
bouncing-reverse.mp448 frames · 12 fps · PyBullet
V-JEPA2 · BOUNCING · CONTEXT 8
+0.0000%
L forward · exact—
L reverse · exact—
SceneBOUNCING
Condition—
Values quoted from the corrected paper’s per-clip evaluation of the checked-in sequences. They describe these mp4s — not the 3D world above, not the 160-video randomized-domino rerun, and not the 220-video continuous sweep.
TRA across contextexact clip · same model · click a column
TRA = (L reverse − L forward) / L forward, on this exact clip. Positive: the reversed order costs more to predict.
Evidence receipt
Know what each number can prove.
This page mixes a playable lookalike with checked-in measurements, so the boundaries are drawn explicitly. Five scopes, no merging.
Real sequence, real loss
Six direction clips come from three checked-in 48-frame PyBullet sequences. The lab’s loss values change with scene, model, and context, read from video_000 per selection.
380 simulations
The paper evaluates 160 discrete-scene videos and 220 continuous restitution or damping videos. A clip is an example, not the claim.
Two supported models
V-JEPA2 shows the expected dissipative separation for contexts 4–12 and reverses sign at context 14. VideoMAE v1 shows the inverse pattern throughout.
Missing artifacts stay out
The corrected rerun has no matching frames checked in here; continuous-sweep videos and per-frame loss traces are also absent. The 3D world cannot regenerate any of them.
Open aggregate evidence and correction notes
SUPPORTED
V-JEPA2
Positive dissipative separation holds across contexts 4–12 in the corrected randomized-domino rerun. The aggregate sign reverses at context 14.
SUPPORTED
VideoMAE v1
The inverse pattern survives recomputation. The corrected paper fixes the checkpoint name from VideoMAE v2 to VideoMAE v1.
WITHDRAWN
MVD
The public checkpoint had no trained decoder weights. Its result cannot support a claim about distillation.
WITHDRAWN
Hiera
The evaluator used random-mask inpainting instead of symmetric future prediction. The objective comparison is invalid.
Corrected randomized-domino rerunΔ = dissipative − low-dissipation TRA
V-JEPA2
+0.1185+0.2084+0.1708+0.1207+0.1702−0.2187
VideoMAE v1
−0.1203−0.1744−0.2375−0.2868−0.3308−0.4015
Percentage points. Exact means from the corrected 160-video discrete rerun.
TRA measures a change in prediction loss. It does not prove that a model understands physics. The core evidence is synthetic, and the real-video results remain confounded by human action timing.