Turn-taking latency

This was the first thing we measured, before memory, before mood, because it turned out to gate everything else. A model that generates the correct response instantly reads as software. The same response, delayed by an amount that matches what the exchange calls for, reads as someone who was actually listening. Getting the delay wrong in either direction breaks the effect faster than getting the words wrong.

We built the baseline from the 12.4k-turn corpus, measuring the human-to-human reply gap in each annotated exchange and bucketing it by what kind of turn preceded it: a greeting, a factual question, an emotionally loaded statement, a joke. The median gap across all categories was 240ms, which is also the number we cite elsewhere on this site, but the median obscures a spread that matters more than the center: greetings cluster tight around 150 to 300ms, while a considered answer to a hard question routinely runs past 900ms without reading as slow.

A single global target fails both ends of that spread at once. Tune the model to answer everything in 240ms and a hard question gets answered too fast to be believed. Tune it to 900ms for safety and a greeting sits in an uncomfortable silence that reads as a broken connection rather than a thoughtful pause. The fix was not a better average, it was routing: latency is now a function of turn type and the mood signal from note 03, not a fixed constant.

One failure mode we did not anticipate: latency that is contextually correct in isolation can still read as wrong if it is too consistent across turns. Real reply gaps are noisy even within the same category. We now add jitter within each category's expected range rather than hitting the median exactly every time, because a metronomic 240ms was, on its own, one of the more reliable tells we found that something on the other end was not a person.