Outcome blending#

Status: refuted

The label blended the search label with a win or loss tag from the game the board came from.

Shared vocabulary in Glossary; battle terms in Theta comparison. This page adds:

The hypothesis#

Game outcomes carry the one signal search labels lack: whether the line actually survived. Mixing a scaled outcome tag into the label should inject the survival gradient without hand-designing badness.

What happened#

Strength neutral, blocking survival worse. Per-game tags mark "positions I was losing in" as globally bad, but survival paths pass through losing-looking positions, so the tags taught the net to avoid the very states survival runs through.

The verdict#

Refuted. An honest outcome signal needs per-board winrate aggregation, which needs roughly a hundred visits per board; measured board overlap between games makes that cost far more games than the loop can afford. Search-derived labels remain the practical survival signal. The variant that survived the gauntlet is Death labels.