Kokekun — findings from the completed 100-trial comparison
15 October 2026

Availability and pronunciation guidance
Main and v3.5 returned complete audio in all 100 attempts. V 3 returned 83 clips:
eight Japanese-output inputs failed on both repetitions (16 guidance failures),
and one English-output request had a temporary speech-service 503 error.
All 50 Japanese-output attempts in v3.5 included pronunciation guidance. Main
returned 33 guided Japanese clips and 17 without guidance. V 3 returned 34 guided
Japanese clips and no audio for its remaining 16 Japanese attempts. These are
observed outcomes, not a guarantee of future reliability or a listening score.

Startup: v3.5 improves on main, with a modest tradeoff against v3
Across all 100 matched v3.5/main trials, v3.5's median paired time to first
received audio bytes was 126 ms shorter. The 95% family-bootstrap interval for
that gain was 53–275 ms. Within the same 83 successful cases on all versions,
first-byte medians were main 1,131 ms, v3 863 ms and v3.5 960 ms. These medians
are not themselves paired differences. The paired v3.5–v3 difference was
+68 ms (95% interval +0.5 to +177 ms); the corresponding mean interval crossed
zero. V 3's startup advantage in successful comparisons was modest and variable.

Completion: v3.5 delivered full clips sooner
The median paired time to the final audio bytes was 802 ms shorter than main
across all 100 trials (95% gain interval 587–1,430 ms), and 1,198 ms shorter than
v3 on 83 shared successes (95% gain interval 512–1,796 ms). In the common 83,
decoded clip durations differed little in median: v3.5–main +0.026 seconds;
v3.5–v3 zero. The earlier final bytes were therefore not simply explained by
shorter clips in this batch. Audible onset, Bluetooth and listening quality
were not measured.

Where the delays occurred
V 3.5's speech startup after translation was 26 ms sooner than v3 in median
paired timing (95% gain interval 10–38 ms). Its remaining audio delivery was
1,310 ms shorter (95% gain interval 531–1,787 ms). Upstream recognition and
translation variation offset the speech-start benefit in overall startup.
The largest v3.5 first-audio delay, 6,681 ms on case 92, included 6,001 ms to
receive the transcript. This identifies the measured stage, not a proven root
cause in a provider. Four Japanese preparation measurements exceeded the 40 ms
target, reaching 335–379 ms; all four retained guidance and returned audio.

Input groups and repetition
The previously fixed input-only groups were reused; outcomes did not determine
membership. V 3.5 completed 50/50 in each group. V 3 completed 37/50 in the shorter,
fewer-conditions group and 46/50 in the longer, more-conditions group. V 3.5's
median paired full-clip gain over v3 was about 400 ms in the shorter group and
2,213 ms in the longer group. These are descriptive subgroup results.

The faster-starting version changed between repetitions for 26/50 inputs in
v3.5 versus main, and 22/41 eligible inputs in v3.5 versus v3. Two repetitions
are insufficient to choose a stable version-selection rule by input. The
clearest demonstrated improvement is preserving speech with guidance; further
speed work should be evaluated separately at recognition, translation,
pronunciation preparation and audio delivery.

Scope and verification
The 100 trials per version use 50 distinct inputs twice:40 typed inputs and 10
saved recordings, both English/Japanese directions, across 21 families. The
original 20 trials were preserved; the remaining 80 continued in the original
order after the approved shared allowance increase from 2000 to 3000. Worker
module hashes and all other settings were unchanged. Original records, fixture
identities, arithmetic and first/last confidence intervals were independently
verified. All 283 saved clips decoded, were non-silent, and had no clipped
samples. Whole-family bootstrap samples retain correlated repetitions.
Segments have different input mixes and cannot establish an effect of the
pause or allowance change. No benchmark jobs were automatically retried.

The interactive report is comparison.html. Detailed measurements and intervals
are in analysis-three.json and measurements-three.csv. The original evidence
remains in the two segment directories named in results.json.