节点 n20

Looking at the results from round 1:
- Score improved from 64.464 to 64.739 (+0.275), but need >65.464
- shape_scale improved: 49.87 → 50.78
- local_spatial improved: 65.93 → 66.12
- d2_shape improved: 0.05847 → 0.05084 (lower is better)
- occupancy_dice improved: 0.8093 → 0.814
- scale_log_ratio: 0.3944 → 0.4196 (moved slightly away from ideal)

The trend is clear: reducing w moves toward base (64.76). With w=0.1 we got 64.739, very close to base. The remaining gap is tiny. To push further, I should try w=0.05 or even smaller, which would be nearly indistinguishable from base but with a tiny hint of OT correction.

But wait - base scored 64.76 on seed 0 only. The parent's multi-seed average with w=0.3 was 64.464. With w=0.1 we got 64.739. The base single-seed was 64.76. So w=0.1 is already nearly at base level.

To push beyond base, I need to think about what else can improve. Looking at the metrics:
- d2_shape (skill 0.471) and occupancy_dice (skill 0.514) are the weakest
- scale_log_ratio skill dropped from 0.553 to 0.538 (worse) - the displacement is still shrinking too much

The base itself scores 64.76 on seed 0. To beat 65.464 I need something beyond just reducing w. Let me think about what other parameters could help:

1. **SCALE_DAMP**: Currently 1.0. The target RMS is computed via log_interp. If the true target is more compact, maybe SCALE_DAMP < 1 could help scale_log_ratio directly without needing OT displacement.

2. **EPS (Sinkhorn regularization)**: Higher eps = more diffuse coupling = more averaged/smooth displacements. Could reduce distortion.

3. **TAU (reg_m)**: Currently 0.9. Higher means more balanced, lower means more unbalanced. 

Actually, looking more carefully: scale_log_ratio raw went from 0.3944 (parent) to 0.4196 (round 1). The base had 0.433. The skill for scale_log_ratio at 0.4196 is 0.538, at 0.433 it was higher. So reducing w actually made scale_log_ratio worse (closer to base's 0.433 which is farther from ideal than 0.394).

Wait, let me re-read: base scale_log_ratio raw = 0.433, w=0.3 = 0.394. The parent's overall score with base was 64.76 (single seed). The issue is that OT displacement helps scale_log_ratio (by shrinking the cloud) but hurts d2_shape and occupancy.

So there's a tradeoff. With w=0.1, we're near base. To beat base significantly, maybe I should try a different approach: keep w very small but adjust SCALE_DAMP to directly address the scale issue.

Actually, let me reconsider. The base scored 64.76 on seed 0 only. The multi-seed average might be different. My target is 65.464. 

Let me try: w=0.05 (nearly off, minimal distortion) combined with SCALE_DAMP=0.95 to slightly reduce the target RMS, which would directly improve scale_log_ratio without the OT displacement side effects.

CHANGE: W_DISP 0.1→0.05, SCALE_DAMP 1.0→0.95; near-zero displacement minimizes d2/occupancy distortion while slight RMS underscaling directly improves scale_log_ratio (base raw 0.433 suggests truth is more compact).

<<<<<<< SEARCH
SCALE_DAMP = 1.0
=======
SCALE_DAMP = 0.95

（只摘前 40 行）