Requirements changed, the document did not
An SRS written for a system that behaves predictably can describe behaviour as a list of rules. A system that generates its own answers cannot, because the correct behaviour is a quality rather than an output. The document has to move from specifying results to specifying the conditions under which results can be trusted.
What to add
- Acceptance criteria expressed as thresholds, not examples
- The escalation path when confidence is low
- What the system must never do, stated explicitly
The sections are not new. Their content is: ranges instead of fixed values, and review instead of assumption.