A running study shows decisions before they are record

A tick is recorded whole or not at all (ADR 0033), and every reader sees closed ticks only. That kept every view honest, but it made a running study look dead: the engine hands a tick's turns to the model as one batch and records them when the last reply lands, so a 1,500-persona concept test (one tick, about an hour of calls on a rate-limited model) showed nothing at all until it was over.

Each persona's decision is now published the moment its reply lands and parses: the model client tells a listener about every outcome as it arrives, the agent turns a parsed reply into a decision (who, which channel, which action, their words, who the item came from), and the study appends it to live/<world>.decisions.jsonl. The file is removed when its tick closes. The interface polls it and plays the open tick's decisions as they arrive, marked provisional, and swaps to the recorded tick when it closes.

What it does not change

The trace stays the only record. Decisions are not written to it, never counted in a digest, and never read by analysis; the live digest (written after each closed tick) is unchanged. Listening changes no outcome: the listener runs after the outcome is stored, and an error in it is swallowed rather than allowed to fail a turn. Within a tick every persona acts on the tick's opening state, so showing one decision before another misstates no cause.

Why provisional

A decision shown early may never become record. The guard can reject a parsed reply and send the turn back with a stricter prompt; an interrupted tick is discarded and asked again on resume, and a model that ignores seeds may answer differently the second time. The interface therefore labels the open tick as provisional and replaces it with the recorded tick, and nothing that computes a number reads the provisional file.

Considered options

Writing each turn into the trace as it lands was rejected: it would end ADR 0033's guarantee that a tick is whole or absent, and every reader would have to learn to ignore half-ticks. Splitting a tick into smaller batches was rejected: it trades throughput for visibility and still shows nothing within a batch.