# ─────────────────────────────────────────────────────────────────────────────
# arms_movement_learned.txt — the LEARNED-MOVEMENT arms (state-conditional
# counted-SBC wave danger), on the FROZEN panel tools/ab/panel_movement.txt.
#
# Pre-registered in docs/movement_campaign.md, "Learned movement (SBC)", BEFORE
# any battle. REFERENCE is `strafe` — the SHIPPED champion. Every delta is
# (arm − strafe).
#
# Offline gate (common_libs/tests/learned_surfer_gate.py) says: the
# state-conditional model beats the global histogram and chance on held-out
# log-loss in 63/63 battles (sign-flip p = 5e-5) but the effect is TINY
# (top-1 3.93% vs global 3.96% vs chance 3.23%). So the live panel decides.
#
# Format: name | ENV=value ENV=value | label
# ─────────────────────────────────────────────────────────────────────────────

# 1. THE CHAMPION — the arm a challenger has to beat (round wins + hit rate).
strafe | TR_MOVEMENT=strafe | champion/reference — shipped strafe defaults

# 2. THE MODULE — state-conditional counted SBC, decay 128 learns / shift 1.
learned | TR_MOVEMENT=learned | state-conditional learned danger (decay on)

# 3. FORGETTING CONTROL — the same learner with decay disabled: isolates the
#    counted+decay memory mechanism. (Offline this arm is the BEST top-1 arm,
#    which is itself a prediction: the corpus is stationary, the live opponent
#    is not.)
learned_nodecay | TR_MOVEMENT=learned TR_LEARNED_DECAY_SHIFT=0 | same, no forgetting

# 4. INFORMATION CONTROL — the same mover, the same geometry, the same counted
#    SBC, but the state is forced to a single cell, i.e. the OLD global
#    histogram. This separates "the learned state conditioning" from "the wave
#    geometry + counted prior" and is the arm that answers "is the failure in
#    the information or in the learner?".
learned_global | TR_MOVEMENT=learned TR_LEARNED_GLOBAL=1 | same mover, state conditioning OFF