Retry cannot tell the user whether the phrasing was wrong or the system itself is fluctuating

What it is

The result is wrong. The cheapest experiment is to try again. That single try stirs two candidate causes at once: I did not say it clearly, or the system is jittering. Two poor results, and there is no way to know whether to rewrite or to switch tools; one poor and one good, and there is no way to know whether the good one was a better sentence or a luckier draw. Retry attribution ambiguity is a diagnostic experiment that is not clean.

It does not deny that output moves. It says: movement makes process of elimination by “once more” impossible.

Why it happens

Elimination needs one factor at a time. Retry often folds “run this sentence again” and “change a few words and run” into one gesture — both are Send or Regenerate. Even with every character frozen, sampling noise still moves, so “controlled variables” is false by default on a generate surface.

People still fill the blank with everyday causal heuristics: miss, blame myself (my prompt was weak); several misses, blame the system (it cannot); one hit, lock in the “secret phrasing.” Those heuristics are roughly usable on deterministic systems and systematically teach the wrong lesson when the noise source is unmarked. The residue is prompt superstition: people collect supposedly reliable incantations that were only glued to a few lucky draws.

Studying it

Take one failed result and two kinds of retry: A freeze the original and sample again; B allow a rewrite then sample. Afterwards ask whether the problem was the wording or the system. Add a third arm: on retry, show “your text was not changed; this is another generation.” Independent variables: rewrite or not, whether sampling is declared, number of prior misses. Dependent variables: attribution choice, belief that a particular sentence is required, abandonment.

To see superstition form, track reuse of a prompt across items: a sentence that happened to work is carried onto an unrelated item. High carry is noise written down as technique.

Where it stops holding

On tasks with temperature near zero and output nailed by a schema or a checker, retry almost only diagnoses phrasing, and the fog thins. If people can see three samples of the same original side by side, the jitter is visible and attribution can split. In customer-service dialogue, people already absorb blame as politeness, which inflates the fog. This entry does not treat undo, nor calling regenerate a refresh — those are label problems; this is an experimental-design problem.

Applying it

• Split “rewrite then send” from “generate this original again.” The latter must say: the words did not change; the generation did.
• When regenerating the original, place old and new side by side rather than overwriting. The juxtaposition is there so noise can be seen, not merely so someone can pick the prettier one.
• Failure copy should not default to “try saying it differently.” Offer one unchanged resample first; if that is still poor, then suggest a rewrite, and name both causes as possible.
• Check: on purpose, emit one poor and one good result from the same prompt. Ask “did you fix it?” If they say yes and the characters never moved, the attribution channel is broken.

Related

• Same group: L1.01.1 The same input can yield different outputs · L1.01.2 Interface conventions assume actions are repeatable and results are stable · L1.01.3 Users misread a lucky correct answer as stable competence · L1.01.4 Controls promise that the same action yields the same result; generation breaks that promise · L1.01.6 Undo and redo lose their meaning when output cannot be reproduced · L1.01.7 Presenting regeneration as “refresh” implies the previous result was a failed load · L1.01.8 Presenting variability as several parallel options is more honest than hiding it behind a single result
• Nearby: L2.01 Openness of natural-language commands and its cost · L2.15 Ambiguity and clarifying follow-ups · L1.11 Reproducibility of non-deterministic output
• Search terms: retry attribution ambiguity · prompt superstition · confounded diagnosis