Confidence answers one question: how much do we actually know?
It does not measure how positive the signals are, and it does not measure how much the signals agree with each other. A profile with consistently negative evidence can carry high confidence — Perscivo is confident about a discouraging read. A profile with a few glowing signals and nothing else carries low confidence, because there simply is not much there.
Confidence bands
- 75% and above — strong evidentiary base.
- 60–74% — moderate. Enough to reason about, not enough for a maximal call.
- Below 60% — thin. Treat any conclusion as provisional.
The sufficiency gate. Confidence cannot exceed 74% unless at least two high-quality observations support the read. One strong signal, however striking, is an anecdote — not a basis for certainty.
Confidence caps recommendations. Confidence and recommendation are never calibrated independently. If the score alone would justify "Proceed" but confidence is moderate, the recommendation is lowered to match what the evidence can actually carry. Confidence can pull a recommendation down; it can never push one up.
What raises confidence. More material, and more varied material. A profile plus a real conversation beats either one alone. Longer conversations beat snippets. Specific, concrete content beats generic content.