Exit math

// TP* (all lanes)

S(1.25x) = median(min(M', 1.25x)) = 1.00x
EV'(1.25x) = 15% * 1.25x + 85% * 1.00x = 1.04x

// S(t) surface

Median payoff vs t. Bright bar = TP*. Dim rows fail the min-hit gate.

// full grid (all lanes)

S = median payoff (winsorized). mean_raw = fragile mean on uncapped peaks.

// lane split

Same robust solver, Selective vs Velocity.

// why raw EV gets owned by outliers

If you use raw hit rate, one moonshot makes p_hat(20x)=1/n. With n=10 that is a 10% "chance" of 20x from a single fluke, and EV(20x) can beat sane TPs. So we do not crown TP* off raw means or raw high-t hit rates.

// fragile (do not optimize this)
EV_raw(t) = p_hat_raw(t) * t + (1 - p_hat_raw(t)) * median(miss)
// one 100x => p_hat_raw(20x) = 1/n  =>  nonsense TP*

// robust model

Winsorize peaks, then score median payoff:

W = clamp(P90(M), 5x, 10x)          // current W = 5.00x
M'_i = min(M_i, W)                   // 100x becomes W, not gospel

payoff_i(t) = min(M'_i, t)           // hit => t, miss => capped peak

S(t) = median_i(payoff_i(t))         // PRIMARY (outlier-proof)

EV'(t) = p_hat'(t)*t + (1-p_hat'(t))*median(M' | miss)   // secondary

eligible(t) <=> hits(t) >= max(1, ceil(0.15*n))   // need 124 hit(s) here

TP* = argmax S(t) over eligible t

Mean payoff on raw peaks is shown only as a caution column. Miss credit is still an upper-bound style mark (observed peak), not a forced zero. Not trade advice.