You can't out-search the engine. So out-clock it. A premove is a commitment made before the opponent's move resolves — safe when their move is predictable, a gamble when it isn't. Predictability is exactly the entropy of the opponent's move distribution, which the relaxation engine already computes.
Opponent moves from real games, bucketed by the entropy of the engine's belief over their reply. Bars = how often the most-believed move was the one actually played (a depth-1 premove would have fired). Threshold H ≤ bits.
Expected seconds banked per move at commitment depth 0 (react), 1 (premove the reply) and 2 (Manny's last-last premove — chain two predicted plies). Depth-2 compounds prediction risk: it only wins where the line is forcing (low H).
Projected clock vs the fraction of moves you premove, at forced-grade (71%) and prep-grade (90%) accuracy. A missed premove costs react + input + penalty — more than just reacting — so cheap premoves are a trap. Where a curve dips under the red 15s line, you survive.
The engine can model your strategy as a distribution — but it never sees your committed premove,
and it can't win the plies you never spent. The human edge in ultra-bullet is a latency game, not a search game:
premove the forced, wait for the fork. Predict a specific opponent with their own style fingerprint and it sharpens —
the machine can't read your mind, but you can read theirs.
← the Logic engine (where the entropy comes from) · Fingerprints →