A small, structured, verifiable, CPU-trainable world model for physics identification research - the auditable counterpoint to billion-parameter video world models. Latent space decomposed into interpretable components with architectural guarantees; every claim in this repository is backed by a committed script, a reproducible number, and a regression test - including two public retractions of v3.x findings that did not survive re-verification.
Author: Cesar Augusto · v4.0.0 (2026-07-04, PRs #25 to #44)
Evolução e verificação do projeto: WIKI.md.
Papers e pesquisas de apoio do futuro agente: AI Search/Papers/ e
AI Search/Research/.
| Claim | Number |
|---|---|
| Physics is identifiable from position-only random-policy data | oracle friction R² 0.865 |
| A learned belief identifies physics, given engineered dynamics features | gravity R² +0.67 |
| Data collection matters as much as the model (active vs random) | +39% belief quality |
| The immutable-latent guarantee holds in open-loop rollouts | drift exactly 0.0 |
| The pipeline generalizes to real MuJoCo contact physics unchanged | 2.3× gap to linear ceiling |
Full ledger with evidence pointers, retractions, named findings and honest
limits: docs/RESULTS.md.
z_full (64) = [ z_static_immutable (8) | z_static_slow (8) | z_dynamic (32) | z_controllable (16) ]
| Component | Enforcement |
|---|---|
z_static_immutable |
hard passthrough in dynamics (AD-2) + optional cross-episode anchor (B9) |
z_static_slow |
drift-penalized residual; optionally conditioned on a causal physics belief (GRU over engineered dynamics features, B10 to B12) |
z_dynamic |
residual MLP or dissipative symplectic (q,p) head (B13) |
z_controllable |
action-conditioned (actions enter only here - architectural routing) |
Training: VICReg-regularized self-predictive objective + prediction
grounding + K-step self-fed rollouts (B7 to B8). Environments: CruelGridworld
(toy, 2D nonlinear gravity) and MujocoPointMass (real contact physics),
with random or information-seeking data collection (B11).
uv sync --extra dev
python scripts/generate_data.py --randomize-physics --process-noise-std 0.05 \
--episode-reset-prob 0.02 --seed 42
python scripts/split_data.py
python scripts/train.py --config configs/experiments/v3_variable_physics.yaml
python scripts/evaluate.py --checkpoint checkpoints/best_model.safetensorsAll workflows (belief training, probing, active exploration, belief
conditioning, signing, ONNX export): docs/USAGE.md.
src/atlas_wm/ installable package (models, environments, data, training, checkpointing, eval, export)
scripts/ pipeline entry points (generate, split, train, evaluate, probe, sign, export, oracle)
configs/ base.yaml + experiments/
tests/ 186 tests verified locally on Python 3.11 in the 2026-07-21 article audit
docs/ RESULTS.md · USAGE.md · MODEL_CARD.md · v4.0-ROADMAP.md · historical postmortems
archive/ quarantined legacy code (do not import)
docs/RESULTS.md- the results ledger: claims, evidence, retractions, limitsdocs/USAGE.md- every workflow and config switchdocs/MODEL_CARD.md- model card with full experimental historydocs/v4.0-ROADMAP.md- the red-team findings and 18-block rebuild planCHANGELOG.md- block-by-block ledger
MIT - see LICENSE.