Auto-sync: 2026-06-27 11:17:45 (part 2)

results/paper_core_results.md

@@ -10,6 +10,11 @@ baseline is the h=16 rank-checkpoint online rollout (`29.74%`).
 | Gaussian field search | No | No | 29.10% | -0.64 pp | Field does not optimize off-manifold noise |
 | Retrieval lattice | No | Yes | 28.93% | -0.81 pp | Nearest train-state action library does not transfer |
 | Retrieval lattice, no expert | No | No | 27.13% | -2.61 pp | Conservative retrieval also fails |
+| Near-miss distillation policy | No | No | 27.48% | -2.26 pp | Imitating near-miss actions does not transfer by itself |
+| Near-miss distillation policy, BC x5 | No | No | 28.29% | -1.45 pp | Stronger BC still stays below policy baseline |
+| Near-miss proposal + field, best-policy ckpt | No | No | 26.32% | -3.42 pp | Field scoring around the BC-selected checkpoint is unstable |
+| Near-miss proposal + field, field ckpt | No | No | 30.14% | +0.41 pp | Clean proposal route begins to recover the mechanism |
+| Near-miss proposal + field, BC x5 field ckpt | No | No | 32.64% | +2.90 pp | Best deployment-clean result so far; still far below same-state lattice |
 | Lattice, no expert/no near-miss | Yes | No | 25.57% | -4.17 pp | Non-local negatives do not help |
 | Lattice, near-miss only | Yes | No | 55.94% | +26.20 pp | Local counterfactual proposals carry the gain |
 | Lattice, no expert | Yes | No | 56.99% | +27.25 pp | Reviewer-safe main result |
@@ -21,13 +26,15 @@ Suggested main-table rows:
 1. Direct h=16 policy
 2. Gaussian field search
 3. Retrieval lattice, no expert
-4. Lattice, near-miss only
-5. Lattice, no expert
-6. Lattice, full
-7. Oracle ceiling
+4. Near-miss proposal + field, BC x5 field checkpoint
+5. Lattice, near-miss only
+6. Lattice, no expert
+7. Lattice, full
+8. Oracle ceiling
 
 Suggested claim:
 
 > DoVLA-CIL is not a better behavior-cloning policy; it is a local counterfactual action
-> selection rule. The learned field only improves rollout when queried on same-state
-> intervention proposals, and the effect is isolated to near-miss counterfactuals.
+> selection rule. A deployment-clean near-miss proposal policy plus the field gives a small
+> gain, but the large effect appears only when the field is queried on same-state
+> intervention proposals, and the mechanism is isolated to near-miss counterfactuals.