NNUE experiment protocol#

Status: standing

The net is not part of the shipped engine. This directory records the research line: what each arm tried, what it measured, and why it ended adopted, parked, or refuted.

Conventions every arm in this directory follows, learned mostly by paying for their absence.

Shared vocabulary in Glossary; battle statistics in Theta comparison; tuning terms in CMA-ES.

Readings#

• Fix the search budget everywhere before comparing anything; the budget is part of the arm.
• One hundred games screens an arm, three hundred adopts one; a 100-game reading carries about five points of standard error.
• Head-to-head duels between arms follow the same seed discipline: more games over more distinct seeds, never mirror rematches.

Fairness#

• Never handicap an arm. An early handicap experiment, zeroing part of one seat's tuning vector, cost about thirty-nine points of measured strength before it was caught.
• Collect with no jitter. The deployment condition is the only condition that matters; jittered collection once flattered the net by weakening its opponent.

Labels#

• Labels always come from the production hand eval. The net's own recorded outputs are never targets; the previous net contributes state distribution only.
• Every new arm passes the same gauntlet: duel readings plus a host test for blocking survival before adoption.

The ladder#

The 300-game column shows the pattern: fresher pools play cleaner but score no higher against production. The loop plateaued with the current recipe, which motivates Blending.