节点 n16

Looking at the results from round 1:

- `scale_damp=0.6` improved `shape_scale` significantly: 56.23 → 58.40 (+2.17)
- `k=15` slightly hurt `local_spatial`: 65.02 → 64.70 (-0.32)
- Net gain: +0.535 (not enough for submission, need +1.0)

The `scale_damp` direction is clearly working. Let me push it further. The parent notes the true target RMS is ~217 vs predicted ~346. With `scale_damp=0.6`, out_rms dropped to 349.4 (from ~354 at damp=1.0). The interpolation is still overshooting.

Actually, looking more carefully: with `scale_damp=1.0`, the formula gives `exp(log(354) + 0.4*(log(335)-log(354))) = exp(5.869 + 0.4*(-0.0536)) ≈ 354 * exp(-0.0214) ≈ 346.5`. With `scale_damp=0.6`, it gives `exp(5.869 + 0.6*0.4*(-0.0536)) = exp(5.869 - 0.0129) ≈ 349.4`. Wait, that's *higher* than damp=1.0. Let me re-check.

`log_interp(rms_a, rms_b, t, scale_damp)` = `exp(log(rms_a) + scale_damp * t * (log(rms_b) - log(rms_a)))`. With rms_a=354, rms_b=335, t=0.4:
- damp=1.0: exp(log(354) + 0.4*(log(335)-log(354))) = exp(5.869 - 0.0214) ≈ 346.5
- damp=0.6: exp(5.869 + 0.6*0.4*(-0.0536)) = exp(5.869 - 0.0129) ≈ 349.4

So damp=0.6 gives *higher* RMS (349.4 vs 346.5). That's moving *away* from the true target of 217! But shape_scale still improved. This means the scoring metric isn't purely about RMS matching — it's about shape similarity, and the higher RMS (closer to source stages) preserves shape better.

So the direction is: higher `scale_damp` might actually be better? No wait — damp=1.0 already gives the lowest RMS in the bracket. Higher damp would go even lower. But the true target is 217 which is far below both...

Actually, re-reading: `scale_damp` multiplies `t` in the interpolation. damp=0 means no interpolation (stay at rms_a=354), damp=1 means full interpolation toward rms_b. So damp=0.6 stays closer to the earlier stage's size. The shape improvement comes from keeping the cloud larger/closer to the source stages rather than shrinking too much.

To push shape_scale further, I should try even lower `scale_damp` (e.g., 0.3 or 0.4). But I also need to recover `local_spatial`. Let me revert k back to 25 (which was better for local_spatial) and push scale_damp down to 0.4.

CHANGE: scale_damp 0.6→0.4 to further reduce RMS interpolation (shape_scale improved with damp=0.6, pushing further), k 15→25 revert to recover local_spatial.

<<<<<<< SEARCH
PARAMS = {"align": "procrustes", "scale_damp": 0.6}
CLUSTER_PARAMS = {"k": 15, "n_clusters": 10, "seed_mode": "random",
                  "jitter_frac": 0.02, "max_layers": 4}
=======
PARAMS = {"align": "procrustes", "scale_damp": 0.4}
CLUSTER_PARAMS = {"k": 25, "n_clusters": 10, "seed_mode": "random",
                  "jitter_frac": 0.02, "max_layers": 4}
>>>>>>> REPLACE