gun j123 Batch 1 outcome + Batch 2 prereg: len6 is the only arm to beat pattern on wins (15 opp), but all 5 arms are wins-positive incl. the bearing-invariant rad_offset; pre-register the 33-opponent confirmation with a structural control
This commit is contained in:
@@ -0,0 +1,34 @@
|
||||
# ─────────────────────────────────────────────────────────────────────────────
|
||||
# arms_gun_b4.txt — PHASE 2 BATCH 2 of the GUN campaign: higher-power
|
||||
# confirmation of the Batch-1 `len6` wins signal, on the wider 33-opponent
|
||||
# panel, with the bearing-invariant `rad_offset` as a structural control.
|
||||
#
|
||||
# Format: name | ENV=value ENV=value | label
|
||||
#
|
||||
# ONE frozen binary (tournament_run.sh builds it from `git archive HEAD`); every
|
||||
# arm below differs ONLY by its env dict. No per-arm rebuild.
|
||||
#
|
||||
# MOVEMENT IS PINNED IN EVERY ARM (`TR_MOVEMENT=strafe`) — same rationale as
|
||||
# Batch 1 (movement-default / concurrent-job hygiene + analyzer liveness rule).
|
||||
#
|
||||
# Why `rad_offset` is here: job j99 proved exactly that the radial correction
|
||||
# changes only the aim DISTANCE and leaves the BEARING untouched, so it cannot
|
||||
# move the real shot direction; its live effect was a tie (+0.28 pp, p=0.62). In
|
||||
# Batch 1 it nonetheless showed +0.42 wins/run (sign test p=0.065) — and so did
|
||||
# ALL five arms, including this structural no-op. If it repeats the nominal wins
|
||||
# gain on 33 opponents, the Batch-1 wins signal is a session-level reference
|
||||
# anomaly rather than a gun effect.
|
||||
# ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
# 1. REFERENCE. The shipped `onlyPattern` rack, movement pinned, no shape env.
|
||||
pattern | TR_MOVEMENT=strafe | shipped onlyPattern rack (REFERENCE)
|
||||
|
||||
# 2. The Batch-1 candidate: 6-tick match key (+0.49 wins/run on 15 opponents).
|
||||
len6 | TR_MOVEMENT=strafe TR_PATTERN_LEN=6 | match key = 6 ticks (Batch-1 candidate)
|
||||
|
||||
# 3. Dose-response on the key-length axis.
|
||||
len16 | TR_MOVEMENT=strafe TR_PATTERN_LEN=16 | match key = 16 ticks
|
||||
|
||||
# 4. STRUCTURAL POSITIVE CONTROL: bearing-invariant radial knob; cannot change
|
||||
# the real aim direction (job j99), so a hits/wins gain here is an artifact.
|
||||
rad_offset | TR_MOVEMENT=strafe TR_PATTERN_RAD_OFFSET=-20 | structural control: aim short 20 px
|
||||
Reference in New Issue
Block a user