SLWM unpublished manuscript v0.3 — Study companion pack
Diogo Antunes / valuealigned.ai and keplen.ai — November 14, 2026

Primary source: slwm_x1b_paper_v03.pdf (21 pages), preserved byte for byte.
Article: https://diogoantunes.ai/publications/signal-latent-world-models
Repository: https://github.com/ValueAlignedAI/slwm
Organization card: https://huggingface.co/valueai-org
Examples and validation guide: https://github.com/ValueAlignedAI/slwm/blob/main/docs/research/x1b_study_artifacts_and_validation.md

Research context
This work is developed at valuealigned.ai and keplen.ai.
https://valuealigned.ai
https://keplen.ai
This companion pack contains study artifacts for reference alongside an
unpublished manuscript. Public access to these artifacts is not a formal model
release or publication of the paper. The supplied PDF is preserved unchanged;
the companion citation identifies it as an unpublished draft.

Contents
- The original manuscript.
- figures/: the 12 original PNG figure panels referenced by the supplied manuscript source.
- closed_loop_replays.json: unchanged Wave 4, seed-2001 recorded trajectories.
- reported-results.csv: selected values transcribed from manuscript Tables 1, 5, 6 and 7.
- slwm-v03.bib: citation for the unpublished manuscript draft.
- model-artifacts.json: links and verification records for the study checkpoints.
- provenance.json: source mapping and SHA-256 checksums.

Interpretation
CSV values are reported summary values, not raw observations or newly computed
experiments. The seed_sd field is standard deviation across three seeds, not a
confidence interval; it is blank where a value has no reported SD. State sizes
exclude model weights, activations, training state, and planning buffers.
Wave 5's 7.31 mm xArm6 result uses trajectory-disjoint data from one workspace.
Wave 7's 31.46 mm result is for the 453M spatial model on held-out scenes.
All physical-robot results are offline predictions on recorded trajectories.

Replay provenance
The JSON is taken unchanged from repository commit
 db14b63b47379ef1fb4938896160a98986ba1447
at experiments/multimodal/x1b/spatiotemporal-jepa-1b-20261111/examples/closed_loop_replays.json.
These are individual Wave 4 examples, not Wave 5 or Wave 7 aggregate evaluations.

The figure panels and manuscript are the supplied v0.3 author artifacts.
Repository links on the article pin the inspected Wave 4 source snapshot; later-wave
results on the page are sourced to the v0.3 manuscript and its figures.
This pack does not include trained model weights, raw datasets, or a complete
seven-wave training archive. The corresponding reference checkpoints are hosted
separately on Hugging Face under the MIT license used by the GitHub project:
- https://huggingface.co/valueai-org/SLWM-X1B
- https://huggingface.co/valueai-org/SLWM-X450M-Memory
- https://huggingface.co/valueai-org/SLWM-v2-453M
MIT copyright owners for those checkpoint packages:
- Diogo Antunes <contact@diogoantunes.ai>
- ValueAI <contact@valuealigned.ai>
- Keplen <contact@keplen.ai>
