Understand probability, confidence and review

Keep three concepts separate

Answer probability is a model signal over an answer. Confidence is a derived distribution signal for Choice and Score. Observed accuracy requires labeled examples; it is not a field you can read from one response.

Noul returns a probability of yes with no separate confidence. Score is a value on your rubric, not automatically a percentage.

Explore a threshold using invented records

The four records and values below are hand-authored teaching data, not API results or production recommendations. The slider changes a local rule only.

Choose thresholds empirically

Define unacceptable errors and their costs. Tune questions and thresholds on labeled development data, then evaluate coverage and error rates on a separate test set. A larger number can reduce automatic handling without eliminating mistakes under distribution shift.

Make review a real destination

Record an authorized input reference, model and question versions, answers, trigger reason and responsible reviewer. Timeouts, failures, missing evidence and consequential actions need explicit branches as well.

Common missing gates

Do not treat a model's approval as permission. Do not bypass approvals because confidence is high. Do not force a top candidate when none is adequate. Batch operations also need action limits and a stop switch.