CHARM
Summary
CHARM is a channel-aware foundation embedding model for multivariate time series. It combines text-conditioned temporal featurization, contextual inter-channel attention, and JEPA-style latent prediction to produce reusable time-channel embeddings.
Role In The Wiki
CHARM is the early time-series JEPA anchor in this corpus. It shows how the JEPA pattern can move from vision and world-model discussions into multivariate time-series representation learning while using channel descriptions as semantic context.
It is not an action-conditioned world model: the model encodes observed time series and supports downstream probes, but it does not expose actions, control inputs, interventions, or counterfactual rollout semantics.
Artifact Status
The verified local artifacts are the arXiv source archive, converted Markdown, rendered PDF, and figures under papers/charm-2025/. No official code repository, Hugging Face page, checkpoint, or released weights were found during ingest.
Evidence
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in charm-2025 rather than duplicating verdict rows here.
At the entity level, CHARM is the early time-series JEPA anchor in this corpus. It shows how the JEPA pattern can move from vision and world-model discussions into multivariate time-series representation learning while using channel descriptions as semantic context. This page should stay as the object card; source pages carry slot-level verdicts, evidence, and missing pieces.