Normalization scale

A task's normalized score is its final result minus the task's floor, divided by the average final result across the models on the roster, so 1 is an average score. A run's final result is the best private score it reached, read off the private split on its own terms rather than through whichever submission led the public board. Both halves of the scale move as rollouts land, so they are frozen (2026-09-02) and reported here rather than recomputed on the fly. A run's whole clock counts toward its final result, whatever length its campaign ran.

The floor is the metric's own lower bound rather than any chosen baseline, and it is zero throughout; aligned SNR, the one metric with no lower bound, clamps there. The divisor is the average of each roster model's mean final result above the floor, per-model means first, so a model that ran more rollouts does not move the scale. A task that runs on more than one board carries a divisor per board, since each board is its own measurement.