An interface that only offers retry is an interface with no control

What it is

The primary button says Try again. There is no way to put hands on the artifact, and no way to put hands on a quantity. All a person can do is draw another lot. Retry-only as absence of control judges that surface: it looks like interaction and is gambling. Control requires the next step to shrink the possibility set, not to take another point from the same distribution. Retry does not shrink. It swaps a point.

Direct edit of the output is clerical control. This entry diagnoses the poorer case: not even the clerical band, the only lever is sampling. It does not discuss how fine the lever should be to match an expert.

Why it happens

In cybernetics, control is feedback driving state toward a goal. Retry’s feedback is “another one,” with error unmapped onto any adjustable dimension. The error might be length, object, tone, one false fact; retry mixes every error into one act. People cannot make “an intervention for this error,” only “an intervention on the luck of the whole piece.” Learning breaks too: the next round does not know what to change, because the only changeable thing is “press again.”

Products often wrap this as exploration. Exploration holds when the distribution itself is the goal (a few style images). Task generation’s goal is to approach one particular artifact; exploration is stalling.

Studying it

One task, three lever sets: retry only; prompt editable but artifact not; artifact or parameters editable. Dependent variables: rounds to an acceptable artifact, mid-way abandon, whether talk attributes the miss to “luck” or to “what I can tune.” Independent variables: whether retry keeps old results side by side, whether the task is exploratory or specified. Add: whether people start clicking retry with no strategy (operationalize as consecutive retries with the prompt unchanged).

Unstrategic consecutive clicks are the behavioral signature. When control exists, two adjacent acts should be distinguishable on some dimension.

Where it stops holding

Lottery-style creative tools (random palettes, random melodies) treat retry as the feature; no control is honest. When time is extremely tight and people will “take a usable draw and go,” retry can exist as an extra lever, not the only one. If the model is pinned to a closed answer (a calculator), retry is meaningless and control should present as unavailable, not as try-again. High-stakes tasks that leave retry as the only exit are putting consequences on luck. This entry does not treat whether the output can be typed into — it assumes it cannot; if it can, it is already not retry-only.

Applying it

• Do not make Generate again the only act on a generated result. Offer at least one more band: edit the artifact, or edit a visible parameter.
• If sampling is truly all you have for now, keep each result side by side so people choose among samples rather than overwrite. Choice is still extremely weak control, stronger than overwrite.
• After more than two retries with the prompt unchanged, prompt the next band of control (point to what is wrong, open document edit) to interrupt unstrategic sampling.
• Check: a specified task (a summary with right/wrong, a mail with required fields). If unchanged-prompt retry remains the main path, and there is no hand on artifact or parameter, the UI is gambling. Log retries before abandon — a high count with unchanged strategy is absence of control.

Related

• Same group: L2.06.1 Output must be directly editable · L2.06.3 Grain should match the user’s expertise
• Nearby: L2.13 Control Granularity and Editability · L1.01 Mismatch Between Stochastic Output and Deterministic UI · L2.05 Iterative Revision
• Search terms: retry-only as absence of control · resample without a lever · control versus lottery