31 lines
2.2 KiB
Plaintext
31 lines
2.2 KiB
Plaintext
# ─────────────────────────────────────────────────────────────────────────────
|
|
# arms_gun_b2.txt — BATCH 2 of the GUN campaign: a WIDER-PANEL calibration of the
|
|
# two arms that carried the only nominal positive signal in Batch 1.
|
|
#
|
|
# Batch 1 (docs/gun_campaign.md, 15-opponent frozen panel, 270 battles) found NO
|
|
# arm beating the shipped `pattern` beyond the MDE, but two arms sat on the
|
|
# positive side of damage with a flat win count:
|
|
# BitBrain Δdmg +10.3 (11/15 opponents, sign test p=0.1185, sign-flip
|
|
# p=0.0325), Δwins +0.11 (MDE 0.35)
|
|
# TMHorizon Δdmg +8.7 (11/15, p=0.1185, sign-flip p=0.0992), Δwins +0.13
|
|
# Both damage effects are BELOW the Batch-1 MDE (~12-14 dmg/run), so rule 3 says
|
|
# they are not detectable — they must not be called a win OR a null. Resolving
|
|
# them needs more OPPONENTS, not more runs, so this batch re-runs exactly those
|
|
# two arms plus the reference on tools/ab/panel_gun_b2.txt (33 opponents).
|
|
#
|
|
# Format: name | ENV=value ENV=value | label
|
|
# MOVEMENT IS PINNED (`TR_MOVEMENT=strafe`) in EVERY arm — same rationale as
|
|
# Batch 1: a default flip may be landing from j120, and the pin makes every
|
|
# delta a pure gun delta and satisfies the analyzer liveness rule.
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
|
|
# 1. REFERENCE — shipped onlyPattern rack, movement pinned.
|
|
pattern | TR_MOVEMENT=strafe | shipped onlyPattern rack (REFERENCE)
|
|
|
|
# 2. BitBrain learned-gain corrector (the arm with the largest nominal damage
|
|
# edge in Batch 1: +10.3 dmg/run, 11/15).
|
|
bitbrain | TR_MOVEMENT=strafe TR_RACK_PATTERN=off TR_RACK_BITBRAIN=both TR_BITBRAIN_GAINS=1.0,1.25,1.5,2.0 TR_BITBRAIN_MEM=decay | BitBrain-only, learned gain (gains 1.0..2.0, decay)
|
|
|
|
# 3. TMHorizon corrector (Batch 1: +8.7 dmg/run, +0.13 wins/run, 11/15).
|
|
tmhorizon | TR_MOVEMENT=strafe TR_RACK_PATTERN=off TR_RACK_TMHORIZON=both | TMHorizon-only rack
|