Two steps are equivalent when they move a configured judge model into the same predictive state. From prefix K, the parent and miner each produce a step; the observer continues from both K+parent_step and K+miner_step to continuation C. The observer's per-position distribution is measured at all three points (K+parent_step, K+miner_step, K alone), producing disagreement s, parent effect d_G, and miner effect d_A.
The score rewards effect similarity (low s) and effect magnitude (low |d_A−d_G|), normalized by d_G for scale invariance. Three judges score every item: HuggingFaceTB/SmolLM2-1.7B-Instruct, microsoft/Phi-3-mini-4k-instruct, allenai/OLMo-2-1124-7B-Instruct. Each judge is reduced to its worst language/depth slice.
The crown metric is the mean of the three judges' worst-slice scores. The king is re-scored on the same fresh exam as challengers every round; a challenger dethrones with a displayed-metric lead of 0.02 in one round OR 0.01 in two consecutive rounds with the same artifact. No challenger slice may fall below the incumbent's worst slice.
After three consecutive rounds as king, both thresholds are halved. The king receives 80% of tier emission; the best challenger with strictly positive paired margin receives 20%; all others receive nothing. Previously scored bytes are skipped; one artifact per (coldkey, tier) per round.
