# The guesser's report card — acceptance measurement (2026-08-18)

The receipt behind the article's draft-head acceptance figures. Measured on
the same workstation as every other number on the page, by reading the model
runtime's own verbose journal during greedy (temperature 0) generation of a
fixed 128-token reply — no benchmark harness, no timing, so live-product
contention cannot touch these counts. Reproduced bit-identically across
three separate runner processes.

| measure | value |
|---|---|
| draft proposals accepted | 84 of 171 (49.1%) |
| tokens written | 128, in 43 forward passes |
| mean accepted length per pass | 2.95 tokens |
| acceptance by guess slot (1st→4th) | 0.860 · 0.558 · 0.372 · 0.163 |
| runtime | ollama 0.32.13, `--spec-type draft-mtp --spec-draft-n-max 4` |

Method notes. The runner logs each speculative verification pass; the counts
above are sums over one full generation, and the identical prompt at
temperature 0 produced the identical trace on every repeat. The
with-and-without-guesser timing comparison is a separate, still-owed
measurement: switching the draft head off requires restarting the model's
runner (the flag is process-level in this runtime version), which reloads a
production-resident model, so it waits for a declared maintenance window
rather than running beside live products.