j127: forced-share allocator (TR_RACK_SHARE, default-off) + allocation batch pre-registration

2026-09-26 10:13:22 +02:00
parent 2e336452e6
commit 6fd5fe6328
5 changed files with 255 additions and 3 deletions
@@ -0,0 +1,42 @@
+# ─────────────────────────────────────────────────────────────────────────────
+# arms_alloc.txt — ALLOCATION BATCH: is the selector's ~66/34 share already
+# right, or is ALLOCATION a lever?
+#
+# 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`) — movement-default and
+# concurrent-job hygiene, same as every other gun batch.
+#
+# GUN_STATS_PATH is set to the SAME RELATIVE name in every arm: the runner's
+# generated bot wrapper `cd`s into a PER-RUN bot dir, so each run writes its own
+# `gun_stats_alloc.jsonl` under
+#   <outdir>/.work/<opponent>/<arm>/run<N>/bots/ModularBot/
+# That is the per-run isolation that makes the applied share and the per-gun
+# per-shot hit rates auditable per arm.
+#
+# Reference = `pattern` alone (the shipped rack). `sel_pb` is the shipped
+# selector with Pattern+BitBrain admitted, i.e. the ~66/34 OUTCOME arm; the
+# `share_*` arms FORCE a share with `TR_RACK_SHARE` (deficit round-robin over
+# dwell epochs) so `policy` is isolated from `gun choice`.
+# ─────────────────────────────────────────────────────────────────────────────
+
+# 1. REFERENCE. The shipped `onlyPattern` rack, movement pinned.
+pattern | TR_MOVEMENT=strafe GUN_STATS_PATH=gun_stats_alloc.jsonl | shipped onlyPattern rack (REFERENCE)
+
+# 2. Shipped selector, Pattern+BitBrain rack. The ~66/34 split is an OUTCOME.
+sel_pb | TR_MOVEMENT=strafe TR_RACK_BITBRAIN=both GUN_STATS_PATH=gun_stats_alloc.jsonl | selector ON, Pattern+BitBrain (shipped policy)
+
+# 3. FORCED 50/50 with BitBrain.
+share_5050_pb | TR_MOVEMENT=strafe TR_RACK_BITBRAIN=both TR_RACK_SHARE=pattern:50,bitbrain:50 GUN_STATS_PATH=gun_stats_alloc.jsonl | forced 50/50 Pattern+BitBrain
+
+# 4. FORCED 70/30 with BitBrain (close to the observed 66/34).
+share_7030_pb | TR_MOVEMENT=strafe TR_RACK_BITBRAIN=both TR_RACK_SHARE=pattern:70,bitbrain:30 GUN_STATS_PATH=gun_stats_alloc.jsonl | forced 70/30 Pattern+BitBrain
+
+# 5. FORCED 50/50 with TMHorizon (closest competitor by measured per-shot rate).
+share_5050_pt | TR_MOVEMENT=strafe TR_RACK_TMHORIZON=both TR_RACK_SHARE=pattern:50,tmhorizon:50 GUN_STATS_PATH=gun_stats_alloc.jsonl | forced 50/50 Pattern+TMHorizon
+
+# 6. FORCED 50/50 with KNN.
+share_5050_pk | TR_MOVEMENT=strafe TR_RACK_KNN=both TR_RACK_SHARE=pattern:50,knn:50 GUN_STATS_PATH=gun_stats_alloc.jsonl | forced 50/50 Pattern+KNN