Auto-sync: 2026-06-30 10:57:05 (part 2)

results/README.md

@@ -162,9 +162,10 @@ compatible tangent transport: safe4 reaches 33.45%, sparse no-op/wrong-gripper
 direct typed proposal generation adds clean support, but the generated support
 is still not local enough. Best-policy checkpoint ablations are also negative:
 sparse no-op/wrong-gripper drops to 32.29%, and no-op-only drops to 33.16%.
-A no-op-only proposal-head retrain is pending as train/eval/summary
-`14963500` -> `14963505` -> `14963507` to test whether the failure is due to
-multi-family proposal-head interference.
+A no-op-only proposal-head retrain improves the typed-generation row to 34.72%,
+with generated no-op selections succeeding 41.64%, but it still stays below the
+36.06% compatible tangent row. Multi-family interference is real but not the main
+proposal bottleneck.
 
 - Code supports `--retrieval-residual-scale`, defaulting to `1.0` so the
   completed 32.12% result is unchanged.
@@ -428,8 +429,8 @@ multi-family proposal-head interference.
 - Jobs `14963143` -> `14963150` and `14963147` -> `14963149` completed
   best-policy checkpoint ablations for sparse/no-op typed proposals at 32.29%
   and 33.16%.
-- Jobs `14963500` -> `14963505` -> `14963507` are pending for the no-op-only
-  proposal-head retrain.
+- Jobs `14963500` -> `14963505` -> `14963507` completed the no-op-only
+  proposal-head retrain at 34.72%.
 
 Paper table automation: