Handle ambiguity in captured evidence

Natural-language answers rarely arrive in the exact shape expected by a rules engine. The conversation layer can propose a mapping, with an explicit path for ambiguous answers.

For closed questions, present the available options and allow the respondent to explain. The model may suggest the closest option based on that explanation, then ask for confirmation. Store both the explanation and the confirmed option code. If none fits, capture "other" or "unable to determine" according to the methodology rather than forcing the nearest favourable answer.

For evidence questions, ask for the fact the rule needs. A question about whether a process is documented may need the document's owner, review date and scope, not a general statement that documentation exists. The system should distinguish an assertion from a supplied artefact. It should also distinguish a link or file reference from verification of that material.

Set follow-up boundaries per question. The conversation may clarify a date, resolve a contradiction or ask for an example of an existing practice. It should not coach the respondent towards a higher-scoring answer or invent evidence from contextual hints. Prompt instructions help, but server-side limits should control how many follow-ups occur and which answer fields can be changed.

Store provenance at answer level:

• the exact question version shown;
• the respondent's raw reply;
• any uploaded or linked evidence reference;
• the structured value proposed by the model;
• the confirmation or correction made by the respondent;
• any reviewer amendment and its reason;
• timestamps and actor identifiers appropriate to the assessment's privacy rules.

Personal and commercially sensitive material should have retention and access rules before collection begins. The audit trail does not justify keeping every piece of conversational detail forever. Record what is needed to explain the result, then apply the approved retention schedule to raw material and derived records.