Context
Section 23 rolls 32 tokens from patterned weights. Section 24 teaches the toy cycle but only locks eight greedy ids and does not persist a checkpoint for that longer check. Companion 47 and Section 23 / 24 / 26 open issues ask for a trained 32-token fixture so generation acceptance exercises a taught continuation.
A new operator is unnecessary: the host already owns the train loop and the NPZ load path.
Decision
Specify the handoff in
27-storylm0-trained-fixture.md:
- Train with the Section 24 SGD recipe.
- Persist
storylm0-checkpoint-v1beside the patterned v0. - Lock the 36-id greedy cycle
[1, 2, 3, 4, 5, 6] × 6. - Keep Section 23’s patterned regression on v0.
No new .kina tokens. No reduce_mean VJP. No real corpus.
Consequences
- Visitors can run
just validate-trained-rolloutand see taught prose over 32 generated tokens. just validate-rolloutstill checks the untrained recipe.- AdamW-trained and real-corpus fixtures remain follow-on work.