Non-Record: Add Novel SemanticEngine SSM submission

[KenMalloy]

Summary

Adds the SemanticEngine CareSSM submission for track_10min_16mb. This is a pure SSM trunk with live episodic memory during both training and legal prequential eval.

Results

Full 50k FineWeb validation docs, packet-online cache, score-before-write eval:

Seed	val_bpb	val_loss	Train steps	Train time	Eval time
42	1.64076237	4.07007627	1692	595.97s	347.0s
1337	1.66718946	4.13563133	1692	594.15s	349.5s
294924	1.62065301	4.02019298	1688	594.27s	364.8s
Mean	1.64286828	4.07530019	1690.67	594.78s	353.77s

Std: val_bpb_std=0.02333959, val_loss_std=0.05789620.

Architecture

• GPU0-5: CareSSM trunk training
• GPU6: episodic residual packet-serving rank
• GPU7: memory maintenance rank
• Eval is prequential: each chunk is scored before its evidence can update the cache for future chunks

Verification

• python -m json.tool records/track_10min_16mb/2026-05-01_SemanticEngine_CareSSM/submission.json
• python -m py_compile records/track_10min_16mb/2026-05-01_SemanticEngine_CareSSM/train_gpt.py
• python -m pytest tests/submission/test_train_gpt_hyperparams.py -q -> 9 passed

[KenMalloy]

Architecture note for reviewers:

This submission is a language model built around a recurrent SSM, not a transformer with attention layers. The novel part is that the model has a separate online episodic-memory system that prepares small residual tensors for the SSM to consume. Think of it as an asynchronous memory channel into the recurrent model, not a post-processing cache or a second-pass scorer.

The hardware split is part of the method:

• GPUs 0-5 run the main SSM training/eval path. They keep moving even if memory has nothing ready.
• GPU 6 serves memory packets: compact residual tensors computed from the current memory table and published latest-complete to the main SSM. If no fresh packet is available, the residual is zero, so the trunk never blocks.
• GPU 7 maintains the memory table: it decides which hidden-state evidence should be kept, refreshed, or retired, and sends committed memory updates to the packet-serving rank.
• The CPU schedules this work and records telemetry. It coordinates the memory system; it is not where the main model FLOPs happen.

The eval path is prequential. For each chunk, the model first scores the tokens using only the checkpoint plus memory built from earlier chunks. Only after that chunk's loss is fixed can its hidden-state evidence update the memory table for future chunks. There is no train-on-validation-before-scoring step, no rescoring of the same chunk, and no best-of-multiple-passes selection.

So the core claim is: a fast SSM trunk can stay on its throughput path while a separate memory subsystem continuously prepares causal residual information for it. The memory is online during eval, but it only learns from tokens after their score has already been counted.