SOMA
SOMA means Sharded Optimization Mixture of Assemblies. It trains domain-specialized LSTM experts using separate SPSA loss signals and combines selected predictions at inference. Independent objectives reduce cross-expert estimator noise, at the cost of jointly learned representations.
Contract And Limits
The main small-expert recipe has a shared frozen tied head after seed training; the wide fixed-size controls update separate heads. The updated browser demo uses a different learned combiner by default and is not interchangeable with the paper configuration.
See the source evidence ledger for dates, status, pinned artifacts and numerical comparisons.
Official Artifacts
Relation To Foundation TSFM Agenda
This is adjacent training/evaluation context, not direct numeric time-series or action-conditioned world-model evidence. Forward-Only Sequence Training compares the mechanisms and defines the long-memory tests still needed.