The Calibration Gap

Why cross-boundary confidence should be divided by 5–7

The scatter animation is illustrative, not measured: it shows the shape of the problem, that a prediction spreads as it leaves the domain it was measured in. Two numbers are measured (KC#META-1): cross-boundary predictions succeed about 1 in 8, and confidence should be divided by 5–7. That correction is a single band. No per-boundary-type or per-crossing gradation was measured, so every crossing below carries the same band.

Within-domain prediction
0
Within domain Single crossing Double crossing Triple crossing

Example boundary types

Within Domain Physics to Biology Lab to Deployment Model A to Model B
1×
scatter drawn, illustrative
8 / 8
dots in target, illustrative
does not apply
measured correction (KC#META-1)

The Empirical Finding

Cross-boundary predictions succeed ~1 in 8. Before extrapolating: divide confidence by 5–7.

Eight experiments, thirteen specific claims, one consistent finding.

Correction factor: KC#META-1.

prediction lands inside target prediction misses target confidence-correction band