Calibration is the training step that derives each trait's three breaks from the trait's trained sample distributions. Each trait is calibrated independently against its own distributions — there is no global scale.
Calibration takes the scored positive and negative samples for a trait and computes the three break values that partition the 0–100 axis into four tiers. The break values are the quartile-derived thresholds defined by the breaks specification. Calibration runs once per trait, per training run, and produces the break values stored on the trained model.
The trait's scoring parameters, fit earlier in the training run, are applied to the positive and negative training samples. Each sample receives a score on the 0–100 axis. The collection of scored positive samples forms the positive training distribution; the collection of scored negative samples forms the negative training distribution.
The three breaks are computed from the positive and negative distributions as defined in the breaks specification:
developing = 75th percentile of the negative training distributionsolid = 25th percentile of the positive training distribution, floored to developingstrong = 75th percentile of the positive training distribution, floored to solidThe floor operations enforce monotonicity: developing ≤ solid ≤ strong. When the positive and negative distributions overlap, floors cause adjacent breaks to collapse. The collapse propagates to low confidence and null tier assignments in the overlap region.
Each trait in a model has its own positive and negative distributions and is calibrated from those distributions alone. A score of 65 may correspond to the solid break on one trait and the developing break on another because each trait's breaks are derived from a different distribution of training samples. Calibration does not impose a shared scale across traits.
Quartile estimates from very few samples are unstable. A trait trained on a handful of samples per side produces quartiles that depend heavily on the specific samples chosen, and the derived breaks may shift substantially when a new sample is added. Stability improves as the per-side sample count grows; the effect on scoring is captured by confidence, which drops in regions where the distributions fail to separate reliably.
When the positive and negative distributions overlap on the score axis, the monotonicity floor causes one or more breaks to collapse. The collapsed band has zero width — no score can fall within it — and scores in that region are returned with a null tier label and low confidence. Calibration does not attempt to resolve overlap; the overlap is surfaced as reduced confidence rather than hidden.