Precision can disguise uncertainty
A response with confidence 0.86 looks exact. It may be a useful ranking signal if validated for that purpose. Yet the reader may take it to mean an 86 percent chance that the statement is true, when the system only meant that a search result matched the query well.
Memory work has several distinct uncertainties: did the right record appear, is its owner identified, is it current, is its classification reliable, and does the source support the inference the consumer wants to make? One number cannot silently answer all five.
Several confidences may coexist
Suppose Memory finds a dated note that a meeting was planned. Retrieval confidence might be high. Confidence that the meeting occurred should remain low until there is a linked action or independent evidence. Adding the scores would not cure the missing event.
PHS's M16F calibration design distinguishes raw ranking scores from calibrated probabilities and requires sample counts, slices and confidence intervals for metrics. M16B keeps classification confidence distinct from semantic relevance and asks for uncertainty dispositions. These are proposed evaluation rules, not an assertion that every present response is calibrated.
Pause here: What proposition does this confidence score measure, and how was that interpretation checked?
Calibrate before asking the reader to trust it
A useful response names the target of confidence, its source, whether it is a declared assessment or calibrated probability, and the missing evidence. The provider should state what its score measures and how it was obtained; the consumer must not relabel retrieval similarity as confidence that the planned meeting occurred. If the sample is too small, insufficient evidence is more honest than a precise decimal. If two qualified sources disagree, the disagreement is part of the result.
Numbers should help a person judge, not relieve the system of explaining itself. When a figure earns trust through validation and clear semantics, it becomes valuable. Until then, it is a labeled instrument reading, not a verdict.
My answer: A declared, uncalibrated assessment can be meaningful when its target and limits are clear. Probability claims require calibration; neither kind of score should silently become certainty about a different event.