movement final: pre-register the ship criterion before the confirmation battles

2026-09-26 02:21:09 +02:00
parent 47ce244740
commit ff03e81591
2 changed files with 71 additions and 0 deletions
@@ -0,0 +1,39 @@
+# ─────────────────────────────────────────────────────────────────────────────
+# arms_movement_final.txt — FINAL CONFIRMATION of the movement champion, run at
+# HIGHER POWER than any previous batch (5 runs/arm vs 3), on the FROZEN panel
+# tools/ab/panel_movement.txt.
+#
+# Prior data this batch is built on (docs/movement_campaign.md §3–§5, Batches
+# 3+4 consolidated, MEASURED on the frozen panel across four independent
+# sessions, ~900 battles):
+#   * `TR_MOVEMENT=strafe` (current defaults) BEATS the shipped `tfil` on round
+#     wins in EVERY session: +0.38/+0.33 (B1), +0.58/+0.47 (B2), and 53.0% vs
+#     42.2% (B3), 52.6% vs 38.5% (B4) — all 95% CIs excluding 0, sign test
+#     favouring strafe.
+#   * It wins by SURVIVAL: incoming hit rate ~12% vs tfil's ~18%, ~50 less
+#     damage taken/run, NO detectable damage cost.
+#   * 8 further arms (reversal timing, spread/reach, heat-field strength, wall
+#     margin/curvature) were swept in Batches 3–4: NOTHING beats `strafe`.
+#     `wide_spread` is the only positive point estimate (+0.11 wins/run, NS,
+#     p=0.55, MDE 0.56) — a TIE, carried here as the single challenger.
+#
+# SHIP CRITERION (pre-registered in docs/movement_campaign.md before the
+# battles): ship only if `strafe` beats `tfil` head-to-head with the 95% CI on
+# wins/run EXCLUDING 0 AND the cross-opponent sign test FAVOURING `strafe`.
+#
+# REFERENCE is `tfil` — every delta below is (arm − tfil), opponent by opponent.
+#
+# Format: name | ENV=value ENV=value | label
+# ─────────────────────────────────────────────────────────────────────────────
+
+# 1. THE CHAMPION under confirmation — current strafe defaults.
+strafe | TR_MOVEMENT=strafe | champion — current strafe defaults (candidate to ship)
+
+# 2. THE ARM TO BEAT — the currently shipped default, explicitly selected.
+#    (The explicit `TR_MOVEMENT=tfil` matters: it also proves the override path
+#    stays reachable after any default flip.)
+tfil | TR_MOVEMENT=tfil | shipped baseline — explicit tfil override
+
+# 3. THE ONLY CHALLENGER with a positive wins point estimate from Batches 3–4
+#    (+0.11/run, damage-neutral, not distinguishable — expected to tie).
+wide_spread | TR_MOVEMENT=strafe TR_STRAFE_SPREAD=2 TR_STRAFE_REACH=216 | Batches 3–4 positive-point challenger (±2 tiles, 216px)