When the answer doesn't fit the question

People don't answer in the exact shape a rules engine expects, so the conversation layer proposes a mapping and you need a clear path for the ambiguous ones.

For closed questions, show the options and let the respondent explain. The model can suggest the closest option from what they've said and then ask them to confirm it. Store their explanation and the confirmed option code together. If nothing fits, record "other" or "unable to determine", whichever the methodology says, instead of nudging it to the nearest favourable answer.

For evidence questions, ask for the specific fact the rule needs. If the rule cares whether a process is documented, you probably need the document's owner, review date and scope, and "yes, we've got documentation" won't cover it. The system should be able to tell an assertion apart from a supplied document, and a link or file reference apart from someone having actually checked what's in it.

Set follow-up limits per question. The conversation can clarify a date, sort out a contradiction or ask for an example of something they say they already do. It shouldn't coach anyone towards a higher-scoring answer or make up evidence from hints in the context. Prompt instructions help, but I'd put the hard limits on the server: how many follow-ups, and which answer fields the conversation can change.

Record where each answer came from:

• the exact question version shown
• the respondent's raw reply
• any uploaded or linked evidence reference
• the structured value the model proposed
• the respondent's confirmation or correction
• any reviewer change and the reason for it
• timestamps and actor IDs, to whatever level the assessment's privacy rules allow

Personal and commercially sensitive material needs retention and access rules sorted out before you start collecting it. An audit trail isn't a licence to keep every bit of conversation forever. Keep what you need to explain the result, then apply the agreed retention schedule to the raw material and anything derived from it.