Questions should narrow after failures

What it is

After a miss on the same slot, the next sentence should not be a replay of the last. Progressive constraint shrinks the reply space step by step. An insurance first-notice call may start with “tell me what happened,” move on a miss to “collision, scrape, or theft,” then to “did you hit another car.” The open form gets one chance to catch a full narrative. Staying open after a miss re-opens a coverage hole that has already shown itself. What narrows is question form, not the whole dialogue being thrown out.

Why it happens

Every miss leaves information: not “nothing is known,” but “this layer of coverage did not catch that utterance.” The posterior is still wide, yet some paths are already weaker. Keeping the next question at the same width ignores the diagnostic value of the miss; the user hears “it still cannot take the kind of thing I said” and often repeats the same sentence, copying the failure.

Narrowing is entropy reduction done with the question. An open sentence has high conditional entropy and is a reasonable first shot. A choice question concentrates mass on a few mutually exclusive classes. A yes/no question compresses what remains to one bit. Each step trades expression for parsability. Jumping too fast (yes/no on the first miss) throws away narratives a middle tier might still catch; never jumping lets people hit the wall at the same width. Other repairs — change channel, correct one slot, repeat the whole utterance — can happen at the same time; narrowing specifically means the option set got smaller.

Studying it

A progressive-prompting protocol: same task, fixed width versus width that shrinks after misses. Independent variables: which miss triggers a form change (open → choice → yes/no), and how many attempts each tier allows. Dependent measures: eventual slot fill, the tier at which people abandon, and whether the narrowed answer still matches the first open narrative (when a human transcript of that first turn exists).

The control is “after a miss, rephrase without changing width” — that is a re-prompt, not narrowing. Code whether people feel forced into the wrong class after narrowing (a debrief: “it was not one of those three”). In the field, inspect whether question form in consecutive-miss sessions actually changes with count; three misses on the same open sentence means there is no protocol.

Where it stops holding

If the first sentence is already a closed choice, there is little room to narrow; the next tier is yes/no or a handoff, not a finer invented menu for the sake of having one. If the open tier already contained parsable fragments the system missed (a name, a policy number), narrowing must not drop those fragments or the person will feel erased. Yes/no strongly induces default agreement, especially when tired or driving; a high-consequence slot must not end the ladder on a casual “yes.” When the ladder has shrunk to the system’s single guess and asks “is it this,” the move has already become confirmation, not a question.

Applying it

• Write a miss ladder for every open or wide-choice slot: first miss, change the example not the width; second, a finite set of classes; third, yes/no or handoff to a person or a screen. Put it in the state machine; do not leave it to whoever is monitoring.
• Carry fragments already heard into the narrower question: “it sounded like another car — collision or scrape?” Do not re-ask “what happened” from zero.
• The end of the ladder needs an exit: handoff, park the task, or a still-open short slot via typing or spelling. A third turn must not remain the original open sentence.
• How to check: filter sessions with two or more understanding misses and plot question width against miss count. Sessions whose width never moves are defects. Sample successful narrowing and debrief whether the answer was forced by the options.

Related

• Same group: M2.02.1 Open questions raise NLU failure · M2.02.2 Closed questions constrain expression but are reliable
• Nearby: M1.06 Repair strategies · M1.11 Silence, timeouts and re-prompts · M2.04 Error-recovery wording
• Search terms: progressive constraint after failure · tapered questioning · prompt narrowing