fender70/spectra-exp-a

SPECTRA Experiment A: construct-validity dataset

Synthetic physics clips used to test whether video-world-model evaluators (DisMo/CRONOS, DINOv2, V-JEPA 2, CoTracker, RAFT, Qwen3-VL and simple pixel/state baselines) separate physically valid futures from controlled physics violations that share the same visual prefix.

The code that generates, verifies and scores this dataset lives in the (private) spectra-benchmark repository under experiments/2026-10-17-exp-a-construct-validity/. Rendering uses PhysScene at commit 70aef81fbf091eb0bb847725e9de2a341aec96fb (MuJoCo 3.16.0 physics, timestep 0.001 s, seed 1017, configs/spectra_exp_a.yaml).

Contents

file	what
dataset_full.tar	every clip with its PNG frames, video, masks, per-frame states, camera and metadata
dataset_slim.tar	the same clips without PNG frames (what the evaluators read)
dataset_manifest.json	frozen per-clip hashes (rgb, masks, trajectory) and the dataset SHA-256
clips.csv	index: one row per clip with role, base clip, violation type, severity, onset frame, fold
GATE_REPORT.md, gate_report.json	decision-gate checks run on this exact build
SHA256SUMS	checksums of the dataset files above
results/pairs_<evaluator>.csv	every clip scored against every bank reference of its scene and view
results/clips_<evaluator>.csv	reference-free scores
results/provenance_<evaluator>.json	model, options and pins behind each evaluator's scores
results/analysis/	AUROC, bank-vs-single deltas, acceptance at CRONOS and calibrated thresholds, matched reversals, severity monotonicity, CRONOS composite success

Scores cover 12 evaluators: CRONOS DisMo, DINOv2 (frames and object consistency), CoTracker3, Qwen3-VL-8B (view v0 only), bg_mse and disappearance; V-JEPA 2; RAFT and Farnebäck optical flow; pixel RMSE; and a simulator-state comparator.

Dataset SHA-256 (from the manifest): 8f59a71ba069d476a08228a371d83241988aea5a86219d78c8cb01fb905b36e9. Each clip lives at dataset/<event>/<scene>/<object>/<appearance>/<view>/<kind>/<cid>/, with the video at movies/complete.mp4.

Composition

1,182 clips: 18 physics groups drawn from events collision, fall and occlusion, scenes studio_floor and studio_table, and objects Ball, Block and Can; 3 appearances (red, orange, teal) and 2 camera views (v0, v1). Every candidate shares a matched 24-frame prefix with its reference.

role	clips
bank (valid alternative futures used as references)	84
heldout (valid futures to be accepted)	72
invalid (physics violations)	1,026

violation type	clips
teleport	232
speed_jump	178
vanish	166
gravity	108
drift	96
freeze	84
time_reversal	84
penetration	78

Violations come in graded severities; clips.csv gives the severity, injected value and onset frame per clip. Every valid and invalid label was independently re-derived from the exported states by a physics-law validator that shares no code with the generator (see the gate report).

Reproducing

The build is bit-reproducible on x86-64 hosts with AVX-512. On hosts without it, MuJoCo planning diverges and 632 clips differ from the manifest. Regenerated datasets should always be checked with python -m spectra_bench manifest --check dataset_manifest.json before scoring.

Known gaps

Gate criterion G6 reports provenance.physscene_commit missing from every clip's metadata, because PhysScene was pip-installed rather than run from a git checkout. The commit is recorded above and in the build status. The field is not part of the manifest hashes.

License

Not yet decided; to be set before any public release.