Adapter Contract
The adapter is valid only when it satisfies all of the following.
Trace identity
capture(task) returns the vectors actually reinjected at the model's latent-token positions, in recurrent order, with shape [latent_steps, hidden_size].
Supplied-trace replay
score_with_trace(task, trace) runs the same teacher-forced prompt and answer suffix while replacing only the reinjected latent vectors. Exact own-trace replay must reproduce answer logits within the frozen tolerance.
Source parity
The package includes a separate reference recurrence derived from the pinned upstream source. The adapter engine must match it exactly on a controlled PyTorch fixture.
Fixed surfaces
Across clean and intervention arms, the following remain fixed:
- checkpoint and base model;
- tokenizer and special tokens;
- prompt text for the recipient task;
- token positions;
- attention mask;
- answer suffix;
- teacher-forced label mask;
- decoding settings.
Modes
latent: admissible for causal and semantic construction.direct: admissible as a pre-seal mode control.verbalized: contains gold proof steps and is post-seal only.
Fingerprint
The adapter fingerprint commits:
- upstream repository and commit;
- source-verification receipt;
- checkpoint repository, revision, filename, and full SHA-256;
- local base-model path, revision, and tree SHA-256;
- latent-step count;
- device type.
Failure
Any replay, source-lock, cache representation, latent-position, or trace-shape mismatch returns an adapter failure. It must not be treated as evidence about the model's semantics.