The CV harness gained a calibration experiment: per-outcome isotonic
maps (pool-adjacent-violators) fit walk-forward on the five folds and
judged leave-one-fold-out against a pre-registered bar. It cleared it:
LOFO ECE improves 16% (0.0146 → 0.0122) with RPS slightly better too
(0.1724 → 0.1722). The final maps (fit on all folds) are committed as
a research artifact, embedded into ratings.json by buildRatings when
the shipped config matches, and applied in matchMatrix by scaling the
score-matrix outcome regions — scorelines, Monte Carlo and displayed
probabilities all stay mutually consistent. Without the field, output
is bit-identical (golden tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Five two-year walk-forward folds (2014-2024), each with its own
pre-fold parameter fit; selection by mean fold RPS; staged greedy over
Team-DC decay/shrinkage/weight, fast-Elo blend, form multiplier, and
new Elo-walk axes (home advantage, K scale); untouched 2024-26 final
test. Result: the shipped v4 config holds — the staged champion gained
0.0003, under the 0.0005 ship bar, and the final test agrees. v4 is no
longer a one-window result.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
NilsBriggen