Notes are numbered, dated and revised in place. Each one states the measurement first and the interpretation second. The through-line is behaviour over time: what the model keeps, how it reads the exchange, and what it does when the exchange goes badly.
Facts are extracted per turn and written to a persistent store rather than re-read from context. We report recall at 7, 30 and 90 days, the rate of false memories, and what happens when a user contradicts something they said a month earlier.
We estimate the state of the exchange on every turn and route it into latency, verbosity and willingness to volunteer. A model that stays even-tempered through an argument is the clearest tell that it is not a person.
Negative exchanges have consequences: the character replies later, then in fewer words, then not at all. We describe the thresholds, the recovery path, and why an always-available model reads as a service rather than an acquaintance.
Our first note. 240ms before an answer reads as listening. 900ms reads as thinking. The same gap before a greeting reads as a broken machine, so a single global latency target is the wrong objective.