Board: training lane closed, no survivor - unmodified pool wins; value-anchored twins flipped every count-smuggle

agent_reports_bgkf/FRONTIER.md

@@ -35,6 +35,15 @@ to the ft lane. The WIENER increment is independently clean (additive
   is exclusion misses -> the wall is RANKING, all value sits with
   model/feature/rank lanes.
 - BLENDS (bl, complete): see validated pieces + refuted list.
+- TRAINING COMPOSITION (tr, complete, no survivor): keep the pool EXACTLY
+  as-is (full pooled closed history, synthetic grid included, no weights,
+  dies42). Every mean-negative candidate was a count-smuggle - the value-
+  anchored twin flipped ALL signs (building-balanced, drop-mega-building,
+  1/k dedup); building-robustness moves are the WORST once count-flattened
+  (the mega-building carries real transfer signal); recency harmful; label
+  variants ordering-worse. METHOD NOTE for all lanes: value-anchored twin
+  (blend=(cand,1.0)) + class-preserving weight rescaling are the two
+  count-channel controls - use them on any winner with count movement.
 
 ## REFUTED this campaign (do not revisit)
 - building-bagged/fold-ensemble fitting (+41..+44 all forms)