moving atoms
movingatoms.ai

Moving Atoms Research · Internal Benchmark

The strongest bare model on our internal benchmark

Our internal physics benchmark shows a video model 3 seconds of a real physical event and asks it to predict the next 5. Every system that scores above the numbers below is a best-of-N wrapper: it generates many candidates and submits the best. This page is the honest comparison: one model, one sample, one prompt. Ours is Atom 1, an undisclosed frontier stack: its forecaster watches the 3-second input as a physicist and commits to when visible motion stops, before a single frame is rendered. Every case, every forecast, and every prompt is on this page.

1stof 11 bare
Internal benchmark · verified
56.7 vs the runner-up’s 48.4: +8.3, and the entire confidence interval [50.6, 62.8] clears it.
1stof 4 bare v2v
Internal benchmark · original
61.1 vs 59.7. Ahead, within noise.
0.11pts
Scorer parity
Local scorer validated end-to-end against the benchmark's own pipeline on identical clips.
56.7= 56.8
Fresh vs tuned
46 never-touched scenarios score the same as the 20 used in development; no dev-set inflation.

Bare-model leaderboards

single sample · no reranking · no selection

Where the score comes from

four components, each capped at 1.00; matching the gap between two real takes is the maximum · means over all 66 cases

How each prediction is made

STEP 1 — INPUT
3 s of real footage
72 frames from a locked-off camera, ending at the switch frame, plus the benchmark's own scene and action text.
STEP 2 — FORECAST
Atom 1 simulates the physics
Its forecaster measures speeds frame-to-frame, identifies what drives the motion, simulates forward, and commits to when visible motion stops. It never sees the answer.
STEP 3 — PROMPT
Forecast fills a fixed template
Benchmark text verbatim, a continuation clause, then a timing clause selected and timed by the forecast.
STEP 4 — GENERATE
One sample, 5 seconds
Atom 1 renders the continuation. Single seed. Nothing is selected after the fact.
STEP 5 — SCORE
Against reality, twice
Compared to what really happened, and to the event filmed a second time, which sets the ceiling.

All 66 cases, complete

inputs · forecast · full prompt · prediction · both real takes