Context
Companion 47’s first generation acceptance test is a
32-token greedy rollout from a frozen tiny checkpoint, plus
a seeded sample run and a causal-leak check. Sections 20–22
stop at three generated tokens. while cannot grow S.
In-place writes and polymorphic S are unspecified.
The existing causal forward already type-checks at any
static T. The host can hold a length-T id buffer and
re-call that forward.
Decision
Specify the acceptance test as the existing composition in
23-storylm0-rollout.md:
one T = 36 forward, host-sliced last-live logits, Section
17 argmax or Section 18 Gumbel-max, and a versioned NPZ /
JSON fixture. No new operator. Do not grow S inside
while. Do not add an in-place cache write. Weights stay
the patterned StoryLM-0 recipe, not a trained model.
Consequences
storylm0_rollout.kinaand a.sonumeric check land without a new compiler operator.- Visitors can load the fixture and roll 32 tokens; they still cannot train a useful language model or parse a mesh.
- The toy host train loop is
Section 24.
Prefill-
Tfamilies, in-place writes, and top-k / top-p remain follow-on work.