Token-by-token invites users to start judging before the result is complete, so decisions rest on a half-product

What it is

The first sentence looks off-topic; a hand is already reaching for stop or a rewritten prompt. The rest might have pulled it back. Deciding on a prefix is streaming moving the evaluation timetable to before completion. The evidence set for the decision is smaller than the final result, and biased toward the opening.

Shorter felt wait is streaming’s gain. Earlier evaluation is the same mechanism’s cost.

Why it happens

Reading is online. A word that has appeared is already inside a judgement being formed. Openings also enjoy primacy: the first two sentences set the key, and later text must work harder to reverse it. Generation’s openings are often boilerplate or a wrong frame, with the real answer later. People press stop during the boilerplate, and the later part never happens.

The UI also offers acts that can be pressed before completion: stop, copy, send, regenerate. The acts are legal in time and illegal in evidence. Once an early act has changed the world (sent, overwrote the previous version), the later part arriving is too late.

Studying it

The same final text, streamed versus visible only when complete. Mid-way stop/send allowed. Dependent variables: rate of mid-way acts, whether those acts would still have been taken after completion, gap versus rating of the full text. Independent variables: whether the opening is written to look off-topic on purpose, whether completed state is visually marked.

The opening manipulation is the crux. Without it, premature judgement you measure may only be that there was already enough to see.

Where it stops holding

When the user’s goal is to kill a direction fast (“don’t make this a poem”), deciding on a prefix is right — provided stop really stops, and the emitted part is not sent as a completed state. Short output (one sentence) has no meaningful prefix. Professional follow-along editing (proof as it comes) is another skill and must not be assumed of default users. This entry does not treat the model taking back a sentence people already saw struck through.

Applying it

• Before completion, stop may exist; copy/send/overwrite are off by default, or clearly marked not finished.
• Visually split incomplete from complete: if a caret is still blinking, it is not an outbound object.
• Keep opening boilerplate short, or emit a structure outline before the flesh, so the prefix reads as a direction, not a verdict.
• Check: take a text that only pulls the direction back in the later part, and stream it. If many people stop in the first part because “off topic,” and people who saw the full text do not judge that way, premature judgement is hurting. Then see whether anyone sent the unfinished text.

Related

• Same group: L1.12.1 Generation latency sits in the felt-wait band · L1.12.2 Streaming shortens TTFT, not duration · L1.12.4 Mid-stream self-correction shows overturned content · L1.12.5 Long background jobs should not stream
• Nearby: L3.11 Streaming the generation process · L3.12 Editing and taking over generated content · L2.15 Ambiguity and clarifying follow-ups
• Search terms: premature judgement · deciding on a prefix · streaming first-impression