In shortAn app on a phone that listens to a conversation at several speeds and speaks only at the right moment. Most of the time it stays silent, and every silence is a choice it counts.
An app on a phone with one control. Pressing it wakes a shell of AI cells that listen to the people in the room at several speeds at once, each keeping its own place in the talk, until it is pressed again. It speaks aloud only at the middle point: after a speaker's thought has landed in words, never while it is still forming; when nobody else is about to say it, and then as a nudge that keeps both sides rather than a verdict; while the floor is open and the topic still live. Everywhere else it holds, and each hold is a choice it counts. It listens only among people who know it is listening, and it is judged by when it speaks before what it says: a good line at the wrong moment counts as a miss.
The middle point
Every choice is made from one position: after a parent, beside a sibling, before the options. In live talk the parent is what has just been said, the sibling is the unsaid, and the options are what could be said next. To speak is to turn a held unsaid into said, and what is said frames everyone's next moment, so the default is to hold. Siblings can nudge; parents dictate.
David's practice
The timing rule is David's own practice made into a rule: interrupt rendering, never interrupt emergence, and nudge at the collapse from emergence into words. Rendering is a speaker repeating their own known pattern; emergence is a speaker finding something new, halting as they go.
Why timing, not content
The work so far kept sliding back to content and scores, because content is what models and their harnesses are built to make. The relay scored choices against fixed values, and the first run of twenty cells asked each for a line at a moment picked in advance. Neither touched timing, and timing is what this proposal is about.
The aim
In David's terms, to help a conversation move from C, leaning on the parent and rendering history, through B, facing the options beside siblings, to A, siblings converging on a shared objective from different positions without copying one another.
What it is not
Not an assistant that answers at every pause, not a recorder, and not a judge of who is right. Enlivening it is closer to waking keel's shell than to hatching a new agent: the identity already exists, and enlivening gives it a present. Once enlivened it is mostly silent; stopped, what remains is a record, not a presence.
Is this the app we build: silent by default, speaking only at the middle point, and judged by when it speaks before what it says?
To shape it differently, answer CHANGE and say how; the work goes ahead and your suggestion gets an answer.