How should the rationale be recorded when a person accepts, modifies…

Direct answer

Every material human decision is linked to an immutable version of the exact AI recommendation reviewed. Classify the decision as ACCEPT, MODIFY, REJECT or OVERRIDE, and record the reason code, human explanation, supporting and conflicting evidence, rule applied, person's authority, scope, time and next action. An amendment must not erase the earlier recommendation. Link the final decision and later outcome to the rationale record as separate relationships.

In plain language

Imagine a teacher changing a grade suggested by AI. Writing only 'grade changed' is not enough. The old grade, new grade, answer that caused the change and rule applied should all be visible.

Why this matters

Human intervention without a rationale may correct an error or create a new one. If the previous recommendation is erased, it becomes impossible to see who changed what and who is accountable for the result.

Do not confuse

Acceptance adopts the recommendation without a material change.
Modification corrects or limits part of the recommendation through a human decision.
Rejection means the recommendation is not applied.
Override means an authorised human decision displaces an earlier decision or automated result.
A reason code supports classification; free text supplies the specific context.

What should you do?

Freeze the reviewed AI recommendation with its identity, model, version, prompt, data and time.
Classify the human decision explicitly as ACCEPT, MODIFY, REJECT or OVERRIDE.
Record both a reason code and a short, specific human explanation.
Link supporting evidence, conflicting evidence, the applied rule and uncertainty to the decision.
Preserve the decision-maker's identity, role, authority and relevant interest or conflict status.
Without deleting the earlier recommendation, relate the changed fields, final decision and performed action.
Add the later outcome and any renewed review to the same decision chain.

How do you audit it?

Can the exact AI recommendation and version be reviewed later?
Was the human decision type classified correctly and consistently?
Does the rationale explain the specific change, or is it merely a general label?
Can supporting and conflicting evidence be seen together?
Are the decision-maker's role, authority and possible conflict of interest recorded?
Was the earlier recommendation or unfavourable evidence deleted?
Do the reasoned decision, performed action and actual result match?

Limit

A rationale record need not reveal every internal thought or protected fact. Privacy, security and legal privilege may limit disclosure, but the material reason needed for audit and effective challenge cannot be made wholly invisible.