The deliverable is a technical note, not a paper. Two omissions are deliberate and recorded here so that their absence is not read as forgetfulness, and so the note names them as its own limits.
Every number was taken on one model, Qwen/Qwen3-4B-Instruct-2507. A mechanism cannot be called
general from one point, and the note does not do so: each claim is about this model in this run
mode. What lifts the limit: the same protocol on a model of another family or size (candidates:
Qwen2.5-1.5B-Instruct, Qwen3-1.7B, something outside Qwen). Cost: half a day and one GPU rental.
The normalized scores over the candidate letters are called normalized candidate scores, never
"the probability that the answer is right" and never "calibrated". The data for the analysis IS in
the records (p_cand, gold_pos, candidate_mass): a reliability diagram and ECE are an hour of
offline work without a single new run. Not done because it is a separate question with its own
preregistration, not an appendix to this one. What lifts the limit: a calibration preregistration
with a named ECE threshold and a separate report per candidate_mass stratum (low mass is not high
confidence).
- The shared-prefix cache (one prefill of the text, KV cache broadcast to the fields): arm
B1cache; removes the main cost loss with several fields; correctness gated by agreement with the per-field arm on the same units. - The primary comparison on the full grid; the length break-even at 2048 tokens; the honest re-run of the TypeSafe adapter form with a bounded digit grid.