gun j123 Task A+B prereg: expose Pattern's TR_PATTERN_LEN/TR_PATTERN_DEPTH (default parity) and pre-register the 6-arm match-parameter sweep

2026-09-26 04:01:51 +02:00
parent 467e07a6a4
commit 2a98aba91b
4 changed files with 231 additions and 10 deletions
@@ -0,0 +1,53 @@
+# ─────────────────────────────────────────────────────────────────────────────
+# arms_gun_b3.txt — PHASE 2 BATCH 1 of the GUN campaign: does tuning the
+# incumbent's OWN match-shape parameters beat the shipped `Pattern`?
+#
+# 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`). Reason (hard rule):
+# a parallel job (j124) may be fighting, and the shipped movement default may
+# move; pinning the engine in EVERY arm removes movement as a confound, makes all
+# arms share one movement, and keeps the comparison a pure GUN comparison. It
+# also satisfies the analyzer liveness rule (every declared token must appear in
+# the bot's own raw-env report, and an undeclared TR_MOVEMENT is fatal).
+#
+# Task A (this job) exposed Pattern's own match-shape parameters as runtime env
+# knobs, defaults byte-identical to the pre-knob gun (proven by the parity dump
+# in the ledger):
+#   TR_PATTERN_LEN   — length of the movement segment compared (search key).
+#                      Shipped default 10.
+#   TR_PATTERN_DEPTH — how many recent history entries the matcher may scan.
+#                      Shipped default 500 (= the whole HistorySize buffer).
+# The two pre-existing radial knobs are kept as cheap CONTROLS.
+#
+# Prior data this batch is built on (taken as given; NOT re-derived):
+#   * `docs/gun_campaign.md` Batches 1-2: nothing beats the shipped `pattern`
+#     across 15 and 33 opponents; `onlyPattern` is a measured optimum.
+#   * `docs/bitbrain_campaign.md`: lead AMPLITUDE is a dead axis; lead
+#     INFORMATION is the open one.
+#   * `common_libs/tests/pattern_radial_results.md` (job j99): the radial knobs
+#     move the aim point along an UNCHANGED bearing, so under the shipped
+#     `bmPath` metric they are a structural no-op; live they were +0.28 pp
+#     (p=0.62) / −0.42 pp (p=0.46).
+# ─────────────────────────────────────────────────────────────────────────────
+
+# 1. REFERENCE. The shipped `onlyPattern` rack, movement pinned, no shape env.
+pattern | TR_MOVEMENT=strafe | shipped onlyPattern rack (REFERENCE)
+
+# 2. Shorter match key: more candidate matches, noisier replay.
+len6 | TR_MOVEMENT=strafe TR_PATTERN_LEN=6 | match key = 6 ticks
+
+# 3. Longer match key: fewer, more specific matches; more linear fallback.
+len16 | TR_MOVEMENT=strafe TR_PATTERN_LEN=16 | match key = 16 ticks
+
+# 4. Shallow history: the search never looks further back than 100 entries.
+depth100 | TR_MOVEMENT=strafe TR_PATTERN_DEPTH=100 | history search = 100 ticks
+
+# 5. CONTROL: aim 20 px short (the radial-knob header's intended use).
+rad_offset | TR_MOVEMENT=strafe TR_PATTERN_RAD_OFFSET=-20 | radial control: aim short 20 px
+
+# 6. CONTROL: scale the aim distance by 0.95.
+rad_scale | TR_MOVEMENT=strafe TR_PATTERN_RAD_SCALE=0.95 | radial control: scale 0.95