matCercola18 commited on
Commit
282f14b
·
verified ·
1 Parent(s): c67b25a

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. results/baselines/bank_density_mimo.txt +59 -0
  2. results/baselines/pendulum_bank_density_diagnosis_summary.txt +59 -0
  3. results/baselines/retrieval_mimo.txt +36 -0
  4. results/baselines/siso_baselines.txt +12 -0
  5. results/furuta_v4/ablation_unetAseed1_ema999_unetAseed2_ema999_unetA_ema999_unetS_ema99.txt +8 -0
  6. results/furuta_v4/budget_curves_ddim10.txt +27 -0
  7. results/furuta_v4/budget_unetA_ema999_ddim10_n200.txt +24 -0
  8. results/furuta_v4/budget_unetAseed1_ema999_n200.txt +24 -0
  9. results/furuta_v4/ddim_unetA_ema999_20.txt +2 -0
  10. results/furuta_v4/ddim_unetAseed1_ema999_10.txt +2 -0
  11. results/furuta_v4/ddim_unetAseed1_ema999_20.txt +2 -0
  12. results/furuta_v4/ddim_unetAseed1_ema999_5.txt +2 -0
  13. results/furuta_v4/gate_mlp384k_ema999.txt +2 -0
  14. results/furuta_v4/gate_mlpB256_ema999.txt +2 -0
  15. results/furuta_v4/gate_unetA_ema9995.txt +2 -0
  16. results/furuta_v4/gate_unetAseed2_ema999.txt +2 -0
  17. results/furuta_v4/gate_unetS_ema999.txt +2 -0
  18. results/furuta_v4/memo_mlp384k_ema999.txt +7 -0
  19. results/furuta_v4/perseed.txt +13 -0
  20. results/furuta_v4/rank_mlp384k_ema999.txt +32 -0
  21. results/furuta_v4/refine_mlpA_ema999_summary.txt +12 -0
  22. results/furuta_v4/reg_attribution_partial.txt +36 -0
  23. results/furuta_v4/run_gates.sh +6 -0
  24. results/furuta_v4/scan_mlp_a.txt +19 -0
  25. results/furuta_v4/scan_unet_a.txt +9 -0
  26. results/furuta_v4/traj_gen_warmstart_v4unetA_s0.txt +17 -0
  27. results/furuta_v4/traj_gen_warmstart_v4unetA_s1.txt +17 -0
  28. results/guidance/A16x_results.txt +13 -0
  29. results/guidance/A_ddim_check.txt +1 -0
  30. results/guidance/A_results.txt +47 -0
  31. results/guidance/B_ddim_check.txt +5 -0
  32. results/guidance/B_results.txt +16 -0
  33. results/guidance/pilot_choice.json +17 -0
  34. results/icml_r3/RESUME.md +35 -0
  35. results/icml_r3/RULES.md +44 -0
  36. results/ood/furuta_ood.txt +165 -0
  37. results/ood/pendulum_ood.txt +804 -0
  38. results/ood/siso_ood.txt +165 -0
  39. results/ood/train_holes.sh +6 -0
  40. results/pendulum_full/dp_bank_rank.txt +23 -0
  41. results/pendulum_full/dp_baseline.txt +25 -0
  42. results/pendulum_full/dp_interp.txt +28 -0
  43. results/pendulum_full/dp_interp4.txt +19 -0
  44. results/pendulum_full/dp_interp_kappa.txt +57 -0
  45. results/pendulum_full/dp_target_eval.txt +24 -0
  46. results/pendulum_full/family_sweep.txt +60 -0
  47. results/refine_reference/refine_reference.png +0 -0
  48. results/timing/classical_mimo_cpu1.json +253 -0
  49. results/timing/classical_mimo_cpu1q.json +253 -0
  50. results/timing/classical_mimo_cpu8.json +260 -0
results/baselines/bank_density_mimo.txt ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === BANK-DENSITY SWEEP, MIMO (held-out draws: seeds [9999, 4242], N=[120, 120], pooled N=240) ===
2
+ MODEL SIDE IS FIXED: full-data checkpoint, NOT retrained per bank fraction (the pendulum sweep retrained; this one does not).
3
+ bank = seeded random subset of training PLANTS (all stored controllers of the chosen plants); fractions (0.05, 0.1, 0.25, 0.5, 1.0); subset seeds (0, 1, 2) (f=1.0: single bank);
4
+ model draws: sampler seeds (0, 1, 2); McNemar/regret reference = model seed 0. Verifier/x0/N as retrieval_baselines.py mimo; K=8 and K=1 (K=1 = first candidate).
5
+ oracle: retrieval_baselines.mimo_oracle (150-step batched Adam, arm-independent)
6
+ regret % = (J_best - J_oracle)/J_oracle on the COMMON success set (plants where BOTH the arm and the model seed-0 draw are covered at that K); label 'cond.'.
7
+ coverage 95% CI: bootstrap (5000) over pooled plants of subset-seed 0. McNemar: exact, paired on pooled plants vs model seed 0, reported for each subset seed (min..max p).
8
+
9
+ ----- K=8 ----- MODEL coverage 0.996 +- 0.000 (3 draws), seed-0 CI [0.988,1.000], per-sample 0.961, cond. regret med 2.27%, gen 54.28 ms/ctrl
10
+ per-draw model coverage (seed 0/1/2): seed 9999: 1.00/1.00/1.00; seed 4242: 0.99/0.99/0.99
11
+ arm=retrieval
12
+ f cov mean+-sd CI(ss0) per-samp dCov(ret-model) McNemar p(ss0..) reg_med% cond model reg% same set n_common NN dist
13
+ 0.05 0.996+-0.003 [1.000,1.000] 0.914 +0.000 1..1 2.68 2.27 238 0.1747
14
+ 0.10 0.994+-0.002 [0.988,1.000] 0.915 -0.001 1..1 2.76 2.27 238 0.1657
15
+ 0.25 0.993+-0.002 [0.988,1.000] 0.927 -0.003 1..1 2.55 2.27 238 0.1540
16
+ 0.50 0.993+-0.002 [0.988,1.000] 0.933 -0.003 1..1 2.59 2.27 238 0.1454
17
+ 1.00 0.992+-0.000 [0.979,1.000] 0.934 -0.004 1..1 2.79 2.28 238 0.1376
18
+ arm=blend
19
+ f cov mean+-sd CI(ss0) per-samp dCov(ret-model) McNemar p(ss0..) reg_med% cond model reg% same set n_common NN dist
20
+ 0.05 0.993+-0.004 [0.988,1.000] 0.916 -0.003 0.5..1 2.85 2.27 237 0.1747
21
+ 0.10 0.993+-0.004 [0.988,1.000] 0.918 -0.003 0.5..1 2.85 2.27 238 0.1657
22
+ 0.25 0.992+-0.000 [0.979,1.000] 0.930 -0.004 1..1 2.59 2.27 238 0.1540
23
+ 0.50 0.993+-0.002 [0.988,1.000] 0.936 -0.003 1..1 2.75 2.27 238 0.1454
24
+ 1.00 0.992+-0.000 [0.979,1.000] 0.940 -0.004 1..1 2.87 2.28 238 0.1376
25
+
26
+ ----- K=1 ----- MODEL coverage 0.963 +- 0.000 (3 draws), seed-0 CI [0.938,0.983], per-sample 0.963, cond. regret med 11.62%, gen 54.28 ms/ctrl
27
+ per-draw model coverage (seed 0/1/2): seed 9999: 0.94/0.94/0.94; seed 4242: 0.98/0.98/0.98
28
+ arm=retrieval (blend K=1 == retrieval K=1 by construction)
29
+ f cov mean+-sd CI(ss0) per-samp dCov(ret-model) McNemar p(ss0..) reg_med% cond model reg% same set n_common NN dist
30
+ 0.05 0.932+-0.002 [0.900,0.963] 0.932 -0.031 0.12..0.17 13.62 11.68 218 0.1747
31
+ 0.10 0.922+-0.019 [0.896,0.958] 0.922 -0.040 0.00086..0.23 15.73 11.52 217 0.1657
32
+ 0.25 0.925+-0.014 [0.892,0.958] 0.925 -0.037 0.015..0.23 15.29 11.61 217 0.1540
33
+ 0.50 0.933+-0.009 [0.912,0.971] 0.933 -0.029 0.031..0.27 14.49 11.70 220 0.1454
34
+ 1.00 0.908+-0.000 [0.871,0.942] 0.908 -0.054 0.0044..0.0044 13.81 11.48 215 0.1376
35
+
36
+ ----- retrieval coverage K=8 vs K=1 per f (mean over subset seeds; pooled draws) -----
37
+ f K=1 K=8
38
+ 0.05 0.932 0.996
39
+ 0.10 0.922 0.994
40
+ 0.25 0.925 0.993
41
+ 0.50 0.933 0.993
42
+ 1.00 0.908 0.992
43
+
44
+ ----- smallest f at which retrieval MATCHES the model -----
45
+ rule (fixed before running): mean retrieval coverage >= model coverage - 0.05 AND pooled McNemar p >= 0.05 for >= 2 of 3 subset seeds (f=1.0: its single bank); 'never' if no f qualifies.
46
+ mimo arm=retrieval K=8: f=0.05
47
+ mimo arm=blend K=8: f=0.05
48
+ mimo arm=retrieval K=1: f=0.05
49
+
50
+ ----- NN distance vs f (normalized conditioning space, mean over queries; mean over subset seeds) -----
51
+ f=0.05: draw9999: 0.1782, draw4242: 0.1713
52
+ f=0.1: draw9999: 0.1683, draw4242: 0.1631
53
+ f=0.25: draw9999: 0.1561, draw4242: 0.1519
54
+ f=0.5: draw9999: 0.1483, draw4242: 0.1424
55
+ f=1.0: draw9999: 0.1411, draw4242: 0.1342
56
+
57
+ ADDENDUM: per-held-out-draw K=8 coverage (mean over subset seeds; model = seed-0 sampler / 3-seed mean)
58
+ draw0: model 1.000; retrieval f=0.05: 0.994 f=0.1: 0.994 f=0.25: 0.994 f=0.5: 0.994 f=1.0: 0.992
59
+ draw1: model 0.992; retrieval f=0.05: 0.997 f=0.1: 0.994 f=0.25: 0.992 f=0.5: 0.992 f=1.0: 0.992
results/baselines/pendulum_bank_density_diagnosis_summary.txt ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === PENDULUM bank-density anomaly: diagnosis summary ===
2
+ Anomaly: in pendulum_bank_density.txt, BOTH retrieval and the model get WORSE (higher median
3
+ regret) as f increases (0.25 -> 0.5 -> 1.0), on both held-out draws. Since every f<1.0 bank is a
4
+ subset of the f=1.0 pool, retrieval getting worse with more candidates available is not possible
5
+ via "closer neighbour = better transfer" alone. Three checks, in order:
6
+
7
+ 1) Does nearest-neighbour DISTANCE shrink with f, as the geometry demands?
8
+ YES, monotonically and cleanly (median NN-distance in standardized (l,m,b) space):
9
+ f=0.25 (n=650): 0.2496
10
+ f=0.50 (n=1300): 0.1914
11
+ f=1.00 (n=2600): 0.1691
12
+ So the geometric premise holds: denser banks give closer matches, exactly as expected.
13
+
14
+ 2) Does retrieval regret track that shrinking distance? (diagnostic a: K=1, no best-of-K selection)
15
+ NO. K=1 median regret vs f (mean +/- sd over 3 seeds):
16
+ seed123: f=0.25: 1.71+-0.16% f=0.50: 1.63+-0.20% f=1.00: 1.84+-0.00%
17
+ seed321: f=0.25: 1.19+-0.22% f=0.50: 1.42+-0.16% f=1.00: 1.36+-0.00%
18
+ Flat/noisy, not decreasing -- so the coordinator's hypothesized mechanism ("K=1 improves with
19
+ f, best-of-8 does not, because a dense bank's 8 nearest are near-duplicate") does NOT apply:
20
+ its premise (K=1 improving with f) is not observed. Distance and regret are decoupled here.
21
+
22
+ 3) Is the oracle itself introducing noise (stochastic local solver, different luck per call)?
23
+ NO. Re-ran eval_pendulum.strong_oracle TWICE, independent RNG streams, on 3 plants:
24
+ plant 0: 0.57009 vs 0.57009 (0.00% spread)
25
+ plant 5: 1.88592 vs 1.88590 (0.00% spread)
26
+ plant 10: 0.57613 vs 0.57613 (0.00% spread)
27
+ The oracle is reproducible to 5 significant figures regardless of restart-RNG seed -- ruled out
28
+ as a confound. Regret numbers are not noisy because of oracle non-determinism.
29
+
30
+ 4) So what IS retrieval's K=8 regret so much lower than K=1's, if not distance or oracle luck?
31
+ SELECTION, not native target excellence. Diagnostic (b): the SINGLE nearest training plant's
32
+ OWN stored controller, evaluated on ITS OWN plant against a fresh oracle solve of that same
33
+ plant ("native quality"), has median regret +1.40% (p90 +2.67%, max +3.58%, n=40) -- NOT the
34
+ ~0.2-0.4% the coordinator's hypothesis anticipated. Compared LIKE-FOR-LIKE (native K=1 quality
35
+ vs the SAME candidate transferred to the nearby held-out query, i.e. diagnostic-a's K=1 column):
36
+ native 1.40% vs transferred 1.71-1.84% (seed123) -- a modest, sensible ~0.3-0.4 point TRANSFER
37
+ PENALTY, not the dramatic gap the original (mismatched) comparison against K=8 suggested.
38
+ Best-of-8 retrieval's much lower regret (0.17-0.40%) is therefore explained by SELECTION: among
39
+ 8 candidates whose INDIVIDUAL native quality is unremarkable (~1.4% typical), the one that
40
+ happens to transfer best to THIS query (verified directly on it) is kept -- an order-statistic
41
+ effect of trying 8 non-identical candidates and keeping the minimum-cost one, not evidence that
42
+ the stored targets are individually near-oracle. This is consistent with, but a more precise
43
+ version of, the "Iteration 3" reading (retrieval's edge is in the SELECTION machinery applied to
44
+ an ordinary bank, not in the bank being unusually good).
45
+
46
+ 5) Remaining, NOT resolved by (1)-(4): the MODEL's own regret rise (0.24-0.26% at f=0.25 to
47
+ 0.87-0.98% at f=0.5/1.0, reproducible across both held-out draws, tight sd across training
48
+ seeds) is a real, separate effect concerning model TRAINING, not retrieval geometry. A plausible
49
+ but UNTESTED factor: at a FIXED 800 epochs, f=0.25 (650 examples) receives ~4x fewer gradient
50
+ steps than f=1.0 (2600 examples) at the same batch size, an implicit early-stopping-like effect
51
+ that could make the f=0.25 model's outputs smoother / less fit to per-plant idiosyncrasies that
52
+ do not generalize to nearby held-out points. Flagged as an open question, not verified here.
53
+
54
+ 6) The residual small ordering wobble in retrieval's own regret across f (e.g. seed123 K=8:
55
+ 0.17% -> 0.30% -> 0.40%) is within noise of averaging only 3 INDEPENDENT (non-nested) random
56
+ bank draws per f on N=40 query plants -- given (1)-(3) rule out the distance and oracle
57
+ explanations, this residual is most plausibly small-sample composition noise across 3 seeds,
58
+ not a systematic density effect. It does not change the headline: the model never overtakes
59
+ retrieval's median regret at any swept f, on either held-out draw (pendulum_bank_density.txt).
results/baselines/retrieval_mimo.txt ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === MIMO retrieval/interpolation baselines (N=120, seed=9999) ===
2
+ verifier: simulate_closed_loop_mimo_batched rollout (n_ic=4, dt=0.03, duration=12.0, energy_thresh=0.3)
3
+ oracle: eval_mimo_regret.py's oracle protocol (fixed generic init 0.3*cap, 150-step batched Adam lr=0.08 through the differentiable MIMO sim, sector-projected) -- arm-independent, computed ONCE, shared by every arm (1900.0 ms/plant)
4
+ K_std=8, diffusion seeds=(0, 1, 2)
5
+
6
+ --- K=8 ---
7
+ arm coverage persamp reg_med% reg_p90% reg_max% ms/plant McNemar(n10,n01,p) Wilcoxon(n,p)
8
+ diffusion 1.00 0.95 2.32 8.89 31.76 305.42 (reference) (reference)
9
+ (sd across 3 seeds: coverage 0.004, per-samp 0.004)
10
+ retrieval 0.99 0.92 2.62 11.04 33.89 1.165 (0,1,1) (119,0.87)
11
+ interpolation 0.99 0.93 2.86 12.50 36.31 1.165 (0,1,1) (119,0.89)
12
+
13
+ --- K=1 ---
14
+ arm coverage persamp reg_med% reg_p90% reg_max% ms/plant McNemar(n10,n01,p) Wilcoxon(n,p)
15
+ diffusion 0.94 0.94 9.83 70.78 819.17 305.42 (reference) (reference)
16
+ (sd across 3 seeds: coverage 0.004, per-samp 0.004)
17
+ retrieval 0.86 0.86 13.66 96.49 496.91 0.338 (0,10,0.002) (103,0.17)
18
+ interpolation 0.86 0.86 13.66 96.49 496.91 0.338 (0,10,0.002) (103,0.17)
19
+
20
+
21
+ --- K=8 bootstrap 95% CIs (addendum, computed post-hoc from saved npz; not re-simulated) ---
22
+ diffusion (seed[0]): coverage 95% CI [1.000,1.000] regret_med 95% CI [+1.63,+2.99]%
23
+ retrieval: coverage 95% CI [0.975,1.000] regret_med 95% CI [+1.45,+3.45]%
24
+ interpolation: coverage 95% CI [0.975,1.000] regret_med 95% CI [+1.78,+3.58]%
25
+ --- K=1 bootstrap 95% CIs (addendum, computed post-hoc from saved npz; not re-simulated) ---
26
+ diffusion (seed[0]): coverage 95% CI [0.892,0.983] regret_med 95% CI [+6.14,+12.50]%
27
+ retrieval: coverage 95% CI [0.792,0.917] regret_med 95% CI [+10.22,+19.86]%
28
+ interpolation: coverage 95% CI [0.792,0.917] regret_med 95% CI [+10.22,+19.86]%
29
+
30
+ --- addendum: confirming the success set used by the Wilcoxon cost test above (coordinator follow-up #3) ---
31
+ K=8: diffusion(seed0) covers 120/120 plants
32
+ retrieval: covers 119/120; common-with-diffusion=119 (= a strict subset of diffusion coverage)
33
+ interpolation: covers 119/120; common-with-diffusion=119 (= a strict subset of diffusion coverage)
34
+ K=1: diffusion(seed0) covers 113/120 plants
35
+ retrieval: covers 103/120; common-with-diffusion=103 (= a strict subset of diffusion coverage)
36
+ interpolation: covers 103/120; common-with-diffusion=103 (= a strict subset of diffusion coverage)
results/baselines/siso_baselines.txt ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ SISO why-generative baseline — pooled over 5 seeds × 60 plants (N=300), K=8
2
+
3
+ regressor : plant-stabilisable 0.72 [0.67,0.77] per-sample-stable 0.72
4
+ diff_1shot : plant-stabilisable 0.61 [0.56,0.67] per-sample-stable 0.61
5
+ diff_Kshot : plant-stabilisable 0.75 [0.70,0.80] per-sample-stable 0.61
6
+
7
+ diffusion K-shot vs regressor (paired McNemar-ish sign test): p = 3.09e-02
8
+ controller diversity (per-point std across K diffusion samples): 3.4 (regressor = 0, a single point)
9
+
10
+ TAKEAWAY: the regressor (conditional mean) stabilises fewer plants than diffusion sampled
11
+ K times; the generative DISTRIBUTION is what lets test-time selection/repair find a
12
+ stabilising controller a point estimate misses.
results/furuta_v4/ablation_unetAseed1_ema999_unetAseed2_ema999_unetA_ema999_unetS_ema99.txt ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ model K=1 (draw4242/draw9001, pooled, paired diff to nearest+P) // K=4
2
+ unetAseed1_ema999 K=1: 92.8/93.9 pooled 93.3 vs retr 54.2 | diff +39.2 [+30.0,+47.8] p=0.0000 * K=4: 93.3/94.4 pooled 93.9 vs retr 82.5 | diff +11.4 [+5.6,+17.5] p=0.0006 *
3
+ unetAseed2_ema999 K=1: 90.6/93.3 pooled 91.9 vs retr 54.2 | diff +37.8 [+29.2,+46.7] p=0.0000 * K=4: 91.7/93.3 pooled 92.5 vs retr 82.5 | diff +10.0 [+4.2,+15.8] p=0.0022 *
4
+ unetA_ema999 K=1: 90.6/93.3 pooled 91.9 vs retr 54.2 | diff +37.8 [+28.6,+46.9] p=0.0000 * K=4: 92.8/93.9 pooled 93.3 vs retr 82.5 | diff +10.8 [+5.0,+17.2] p=0.0006 *
5
+ unetS_ema999 K=1: 88.3/92.2 pooled 90.3 vs retr 54.2 | diff +36.1 [+27.2,+44.7] p=0.0000 * K=4: 95.0/95.0 pooled 95.0 vs retr 82.5 | diff +12.5 [+6.7,+18.3] p=0.0001 *
6
+ unetSXD_ema999 K=1: 90.0/91.1 pooled 90.6 vs retr 67.5 | diff +23.1 [+15.0,+31.4] p=0.0000 * K=4: 95.0/95.6 pooled 95.3 vs retr 90.0 | diff +5.3 [+0.8,+10.0] p=0.0380 *
7
+ mlpB256_ema999 K=1: 74.4/75.0 pooled 74.7 vs retr 54.2 | diff +20.6 [+10.3,+30.3] p=0.0002 * K=4: 88.3/91.7 pooled 90.0 vs retr 82.5 | diff +7.5 [+1.9,+13.3] p=0.0113 *
8
+ mlpXD_ema999 K=1: 75.6/77.8 pooled 76.7 vs retr 67.5 | diff +9.2 [+0.3,+18.1] p=0.0500 K=4: 89.4/93.3 pooled 91.4 vs retr 90.0 | diff +1.4 [-2.2,+5.6] p=0.5806
results/furuta_v4/budget_curves_ddim10.txt ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # success (%) vs per-plant single-thread wall-clock budget; verify times as measured, costs from budget_timing_cpu.txt; refine n=200
2
+ # cost model: sample1 0.03s sample4 0.07s regfwd1 3.2ms refine(40 steps) 0.70s (flat in batch<=4)
3
+ method 0.1s 0.25s 0.5s 1s 2s 3s 5s 10s 60s best config at each budget
4
+ retrieval alone 0.0 0.0 48.3 69.2 77.5 83.3 90.0 93.3 93.3 - | - | S=16,m=1 | S=16,m=3 | S=16,m=6 | S=16,m=6 | S=16,m=7 | S=16,m=11 | S=16,m=11
5
+ retrieval + refine 0.0 0.0 48.3 83.3 87.5 93.3 97.5 97.5 97.5 - | - | S=4,m=1 | S=4,m=1,I=20 | S=4,m=2,I=40 | S=4,m=2,I=40 | S=4,m=3,I=40 | S=4,m=3,I=40 | S=4,m=3,I=40
6
+ generated (U-Net, DDIM-10) alone 0.0 0.0 80.1 81.4 81.8 93.1 93.8 94.4 94.4 - | - | S=1,m=1 | S=4,m=3 | S=4,m=4 | S=8,m=8 | S=16,m=11 | S=16,m=15 | S=16,m=15
7
+ generated (U-Net, DDIM-10) + refine 0.0 0.0 80.1 84.4 86.0 97.3 98.1 98.3 98.3 - | - | S=1,m=1 | S=1,m=1,I=20 | S=4,m=2,I=40 | S=4,m=3,I=40 | S=4,m=4,I=40 | S=4,m=4,I=40 | S=4,m=4,I=40
8
+ regression twin (deterministic) alone 0.0 0.8 76.9 76.9 76.9 85.6 85.6 85.6 85.6 - | S=1,m=1 | S=1,m=1 | S=1,m=1 | S=1,m=1 | S=1,m=1 | S=1,m=1 | S=1,m=1 | S=1,m=1
9
+ regression twin (deterministic) + refine 0.0 0.8 76.9 83.3 84.7 93.3 95.6 95.6 95.6 - | S=1,m=1 | S=1,m=1 | S=1,m=1,I=20 | S=1,m=1,I=40 | S=1,m=1,I=40 | S=1,m=1,I=40 | S=1,m=1,I=40 | S=1,m=1,I=40
10
+
11
+ # paired difference A - B in pp at each budget [95% CI] p (best config per method per budget; 120 plants)
12
+ generated (U-Net, DDIM-10) alone - retrieval alone
13
+ +0.0 [ +0.0, +0.0] p=1.000 +0.0 [ +0.0, +0.0] p=1.000 +31.8 [+22.9,+40.9] p=0.000 +12.2 [ +4.7,+19.6] p=0.002 +4.3 [ -3.1,+11.8] p=0.276 +9.7 [ +4.4,+15.6] p=0.001 +3.8 [ +0.3, +8.1] p=0.086 +1.0 [ -2.1, +4.6] p=0.627 +1.1 [ -2.2, +4.8] p=0.622
14
+ generated (U-Net, DDIM-10) + refine - retrieval + refine
15
+ +0.0 [ +0.0, +0.0] p=1.000 +0.0 [ +0.0, +0.0] p=1.000 +31.8 [+22.6,+40.9] p=0.000 +1.1 [ -5.7, +8.1] p=0.776 -1.5 [ -8.3, +5.6] p=0.708 +4.0 [ +0.5, +8.1] p=0.039 +0.6 [ -0.6, +2.6] p=0.877 +0.8 [ -0.2, +2.7] p=0.750 +0.8 [ -0.2, +2.6] p=0.747
16
+ generated (U-Net, DDIM-10) alone - retrieval + refine
17
+ +0.0 [ +0.0, +0.0] p=1.000 +0.0 [ +0.0, +0.0] p=1.000 +31.8 [+22.9,+40.5] p=0.000 -1.9 [ -8.7, +5.0] p=0.605 -5.7 [-13.4, +1.4] p=0.142 -0.3 [ -4.8, +4.1] p=0.934 -3.7 [ -7.9, +0.0] p=0.065 -3.1 [ -7.1, +0.4] p=0.099 -3.1 [ -6.9, +0.4] p=0.141
18
+ generated (U-Net, DDIM-10) + refine - generated (U-Net, DDIM-10) alone
19
+ +0.0 [ +0.0, +0.0] p=1.000 +0.0 [ +0.0, +0.0] p=1.000 +0.0 [ +0.0, +0.0] p=1.000 +3.1 [ +1.0, +5.6] p=0.004 +4.3 [ +1.4, +7.7] p=0.003 +4.3 [ +1.5, +7.9] p=0.002 +4.4 [ +1.4, +7.7] p=0.005 +4.0 [ +1.0, +7.6] p=0.008 +3.9 [ +1.0, +7.4] p=0.015
20
+ regression twin (deterministic) alone - retrieval alone
21
+ +0.0 [ +0.0, +0.0] p=1.000 +0.8 [ +0.0, +2.5] p=1.000 +28.6 [+19.4,+38.1] p=0.000 +7.8 [ -0.0,+15.3] p=0.058 -0.6 [ -8.3, +7.2] p=0.943 +2.2 [ -3.6, +8.1] p=0.523 -4.4 [ -9.2, -0.0] p=0.085 -7.8 [-12.8, -3.1] p=0.002 -7.8 [-13.1, -3.1] p=0.002
22
+ regression twin (deterministic) alone - generated (U-Net, DDIM-10) alone
23
+ +0.0 [ +0.0, +0.0] p=1.000 +0.8 [ +0.0, +2.5] p=1.000 -3.1 [ -7.9, +1.1] p=0.189 -4.4 [ -9.5, +0.4] p=0.080 -4.8 [ -9.9, +0.0] p=0.066 -7.5 [-11.9, -3.5] p=0.000 -8.2 [-13.1, -4.1] p=0.000 -8.8 [-13.5, -4.5] p=0.000 -8.9 [-13.6, -4.6] p=0.000
24
+ regression twin (deterministic) + refine - retrieval + refine
25
+ +0.0 [ +0.0, +0.0] p=1.000 +0.8 [ +0.0, +2.5] p=1.000 +28.6 [+19.2,+38.1] p=0.000 -0.0 [ -7.2, +7.2] p=1.000 -2.8 [ -9.4, +4.2] p=0.479 -0.0 [ -3.9, +3.9] p=1.000 -1.9 [ -5.3, +1.1] p=0.381 -1.9 [ -5.6, +1.1] p=0.366 -1.9 [ -5.3, +1.1] p=0.371
26
+ regression twin (deterministic) + refine - generated (U-Net, DDIM-10) + refine
27
+ +0.0 [ +0.0, +0.0] p=1.000 +0.8 [ +0.0, +2.5] p=1.000 -3.1 [ -7.7, +1.1] p=0.186 -1.1 [ -5.5, +3.2] p=0.650 -1.3 [ -5.8, +3.1] p=0.606 -4.0 [ -7.5, -1.0] p=0.016 -2.6 [ -5.7, -0.2] p=0.111 -2.8 [ -5.8, -0.5] p=0.096 -2.8 [ -5.9, -0.5] p=0.093
results/furuta_v4/budget_unetA_ema999_ddim10_n200.txt ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ seed4242 gen_s0 n_sub=200
2
+ [gen_s0] SDP+verify 66s plain ok 0.900; 24 failed-certified in 8 plants
3
+ [gen_s0] refine done 86s
4
+ [gen_s0] refined ok at (5, 10, 20, 40): [0.917, 0.917, 0.95, 0.983] (K=1) (91s)
5
+ seed4242 gen_s1 n_sub=200
6
+ [gen_s1] SDP+verify 65s plain ok 0.903; 25 failed-certified in 9 plants
7
+ [gen_s1] refine done 86s
8
+ [gen_s1] refined ok at (5, 10, 20, 40): [0.95, 0.95, 0.983, 0.983] (K=1) (90s)
9
+ seed4242 gen_s2 n_sub=200
10
+ [gen_s2] SDP+verify 65s plain ok 0.901; 27 failed-certified in 8 plants
11
+ [gen_s2] refine done 85s
12
+ [gen_s2] refined ok at (5, 10, 20, 40): [0.933, 0.95, 0.967, 0.983] (K=1) (89s)
13
+ seed9001 gen_s0 n_sub=200
14
+ [gen_s0] SDP+verify 57s plain ok 0.934; 16 failed-certified in 4 plants
15
+ [gen_s0] refine done 72s
16
+ [gen_s0] refined ok at (5, 10, 20, 40): [0.967, 0.967, 0.95, 0.983] (K=1) (76s)
17
+ seed9001 gen_s1 n_sub=200
18
+ [gen_s1] SDP+verify 58s plain ok 0.934; 15 failed-certified in 4 plants
19
+ [gen_s1] refine done 72s
20
+ [gen_s1] refined ok at (5, 10, 20, 40): [0.967, 0.967, 0.95, 0.983] (K=1) (76s)
21
+ seed9001 gen_s2 n_sub=200
22
+ [gen_s2] SDP+verify 59s plain ok 0.934; 16 failed-certified in 4 plants
23
+ [gen_s2] refine done 73s
24
+ [gen_s2] refined ok at (5, 10, 20, 40): [0.967, 0.967, 0.95, 0.983] (K=1) (77s)
results/furuta_v4/budget_unetAseed1_ema999_n200.txt ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ seed4242 gen_s0 n_sub=200
2
+ [gen_s0] SDP+verify 100s plain ok 0.924; 17 failed-certified in 5 plants
3
+ [gen_s0] refine done 122s
4
+ [gen_s0] refined ok at (5, 10, 20, 40): [0.933, 0.933, 0.967, 0.95] (K=1) (129s)
5
+ seed4242 gen_s1 n_sub=200
6
+ [gen_s1] SDP+verify 106s plain ok 0.915; 21 failed-certified in 7 plants
7
+ [gen_s1] refine done 133s
8
+ [gen_s1] refined ok at (5, 10, 20, 40): [0.933, 0.95, 0.967, 0.95] (K=1) (139s)
9
+ seed4242 gen_s2 n_sub=200
10
+ [gen_s2] SDP+verify 107s plain ok 0.911; 21 failed-certified in 8 plants
11
+ [gen_s2] refine done 137s
12
+ [gen_s2] refined ok at (5, 10, 20, 40): [0.933, 0.95, 0.967, 0.967] (K=1) (143s)
13
+ seed9001 gen_s0 n_sub=200
14
+ [gen_s0] SDP+verify 100s plain ok 0.938; 15 failed-certified in 4 plants
15
+ [gen_s0] refine done 120s
16
+ [gen_s0] refined ok at (5, 10, 20, 40): [0.967, 0.967, 0.95, 0.983] (K=1) (126s)
17
+ seed9001 gen_s1 n_sub=200
18
+ [gen_s1] SDP+verify 101s plain ok 0.934; 15 failed-certified in 4 plants
19
+ [gen_s1] refine done 121s
20
+ [gen_s1] refined ok at (5, 10, 20, 40): [0.967, 0.967, 0.95, 0.983] (K=1) (126s)
21
+ seed9001 gen_s2 n_sub=200
22
+ [gen_s2] SDP+verify 95s plain ok 0.936; 16 failed-certified in 4 plants
23
+ [gen_s2] refine done 115s
24
+ [gen_s2] refined ok at (5, 10, 20, 40): [0.95, 0.967, 0.95, 1.0] (K=1) (121s)
results/furuta_v4/ddim_unetA_ema999_20.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ unetA_ema999 DDIM 20
2
+ K=1: 91.9% K=4: 93.3% K=8: 93.3% K=16: 94.7%
results/furuta_v4/ddim_unetAseed1_ema999_10.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ unetAseed1_ema999 DDIM 10
2
+ K=1: 91.7% K=4: 93.6% K=8: 93.9% K=16: 94.2%
results/furuta_v4/ddim_unetAseed1_ema999_20.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ unetAseed1_ema999 DDIM 20
2
+ K=1: 91.9% K=4: 93.6% K=8: 93.9% K=16: 93.9%
results/furuta_v4/ddim_unetAseed1_ema999_5.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ unetAseed1_ema999 DDIM 5
2
+ K=1: 90.3% K=4: 92.5% K=8: 93.3% K=16: 93.9%
results/furuta_v4/gate_mlp384k_ema999.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ mlp384k_ema999 ema999 models/v4/mlp_384k_s384000.pth
2
+ K=1: generated+P 73.3% nearest+P 54.2% K=4: generated+P 89.4% nearest+P 82.5% K=8: generated+P 92.8% nearest+P 90.0% K=16: generated+P 95.3% nearest+P 93.3%
results/furuta_v4/gate_mlpB256_ema999.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ mlpB256_ema999 ema999 models/v4/mlp_b256_s80000.pth
2
+ K=1: generated+P 74.7% nearest+P 54.2% K=4: generated+P 90.0% nearest+P 82.5% K=8: generated+P 92.2% nearest+P 90.0% K=16: generated+P 94.4% nearest+P 93.3%
results/furuta_v4/gate_unetA_ema9995.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ unetA_ema9995 ema9995 models/v4/unet_a_s96000.pth
2
+ K=1: generated+P 91.9% nearest+P 54.2% K=4: generated+P 93.3% nearest+P 82.5% K=8: generated+P 93.6% nearest+P 90.0% K=16: generated+P 93.9% nearest+P 93.3%
results/furuta_v4/gate_unetAseed2_ema999.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ unetAseed2_ema999 ema999 models/v4/unet_a_seed2_s96000.pth
2
+ K=1: generated+P 91.9% nearest+P 54.2% K=4: generated+P 92.5% nearest+P 82.5% K=8: generated+P 93.6% nearest+P 90.0% K=16: generated+P 94.4% nearest+P 93.3%
results/furuta_v4/gate_unetS_ema999.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ unetS_ema999 ema999 models/v4/unet_s_s40000.pth
2
+ K=1: generated+P 90.3% nearest+P 54.2% K=4: generated+P 95.0% nearest+P 82.5% K=8: generated+P 95.3% nearest+P 90.0% K=16: generated+P 96.1% nearest+P 93.3%
results/furuta_v4/memo_mlp384k_ema999.txt ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ bank rows 2801; bank-to-bank NN (leave-one-out) RMS: median 0.0444 p10 0.0201 p90 0.1061
2
+ sample diversity: mean pairwise RMS between the 16 samples of one plant (seed 0): median over plants 0.0867 (bank LOO NN median 0.0444)
3
+ generated (candidate 1, 360 = 2 draws x 60 plants x seeds): dist to nearest training row: median 0.0550 p10 0.0336 p90 0.1153
4
+ ratio median(d_gen)/median(d_loo) = 1.24; frac of candidates closer than the 10th pct of bank LOO: 0.006; frac closer than 0.25*median LOO: 0.000
5
+ distance to the same-plant retrieved row: median 0.1270
6
+ succeeded K=1: median d_gen 0.0519 (n=264); failed: 0.0719 (n=96)
7
+ copy test (threshold 0.0111 = 0.25 x median bank LOO): successes with d_gen<thr: 0.0% (0/264); successes with d to same-plant retrieved row<thr: 0.0% (0/264)
results/furuta_v4/perseed.txt ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name | draw | retrieval K1/K4 | sampler seed 0 K1/K4 | seed 1 | seed 2
2
+ unetA_ema999 | 4242 | 51.7/78.3 | 91.7/93.3 | 91.7/93.3 | 88.3/91.7
3
+ unetA_ema999 | 9001 | 56.7/86.7 | 93.3/95.0 | 93.3/93.3 | 93.3/93.3
4
+ unetAseed1_ema999 | 4242 | 51.7/78.3 | 93.3/93.3 | 93.3/93.3 | 91.7/93.3
5
+ unetAseed1_ema999 | 9001 | 56.7/86.7 | 95.0/95.0 | 93.3/95.0 | 93.3/93.3
6
+ unetAseed2_ema999 | 4242 | 51.7/78.3 | 91.7/91.7 | 88.3/91.7 | 91.7/91.7
7
+ unetAseed2_ema999 | 9001 | 56.7/86.7 | 93.3/93.3 | 93.3/93.3 | 93.3/93.3
8
+ unetS_ema999 | 4242 | 51.7/78.3 | 85.0/95.0 | 90.0/95.0 | 90.0/95.0
9
+ unetS_ema999 | 9001 | 56.7/86.7 | 90.0/95.0 | 93.3/95.0 | 93.3/95.0
10
+ unetSXD_ema999 | 4242 | 68.3/91.7 | 88.3/93.3 | 88.3/96.7 | 93.3/95.0
11
+ unetSXD_ema999 | 9001 | 66.7/88.3 | 88.3/96.7 | 91.7/95.0 | 93.3/95.0
12
+ mlp384k_ema999 | 4242 | 51.7/78.3 | 70.0/88.3 | 65.0/88.3 | 63.3/86.7
13
+ mlp384k_ema999 | 9001 | 56.7/86.7 | 80.0/91.7 | 80.0/90.0 | 81.7/91.7
results/furuta_v4/rank_mlp384k_ema999.txt ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ seed4242 gen_s0: ranking 16 candidates
2
+ [gen_s0] SDP+verify 54s; plain ok 0.900; 24 failed with a certified P -> refine
3
+ [gen_s0] refine done 54s
4
+ [gen_s0] refined ok 0.900 (54s)
5
+ seed4242 gen_s1: ranking 16 candidates
6
+ [gen_s1] SDP+verify 53s; plain ok 0.900; 24 failed with a certified P -> refine
7
+ [gen_s1] refine done 53s
8
+ [gen_s1] refined ok 0.900 (53s)
9
+ seed4242 gen_s2: ranking 16 candidates
10
+ [gen_s2] SDP+verify 48s; plain ok 0.875; 30 failed with a certified P -> refine
11
+ [gen_s2] refine done 48s
12
+ [gen_s2] refined ok 0.875 (48s)
13
+ seed4242 retr: ranking 16 candidates
14
+ [retr] SDP+verify 62s; plain ok 0.767; 56 failed with a certified P -> refine
15
+ [retr] refine done 62s
16
+ [retr] refined ok 0.767 (62s)
17
+ seed9001 gen_s0: ranking 16 candidates
18
+ [gen_s0] SDP+verify 55s; plain ok 0.942; 14 failed with a certified P -> refine
19
+ [gen_s0] refine done 55s
20
+ [gen_s0] refined ok 0.942 (55s)
21
+ seed9001 gen_s1: ranking 16 candidates
22
+ [gen_s1] SDP+verify 61s; plain ok 0.929; 17 failed with a certified P -> refine
23
+ [gen_s1] refine done 61s
24
+ [gen_s1] refined ok 0.929 (61s)
25
+ seed9001 gen_s2: ranking 16 candidates
26
+ [gen_s2] SDP+verify 64s; plain ok 0.925; 18 failed with a certified P -> refine
27
+ [gen_s2] refine done 64s
28
+ [gen_s2] refined ok 0.925 (64s)
29
+ seed9001 retr: ranking 16 candidates
30
+ [retr] SDP+verify 62s; plain ok 0.783; 52 failed with a certified P -> refine
31
+ [retr] refine done 62s
32
+ [retr] refined ok 0.783 (62s)
results/furuta_v4/refine_mlpA_ema999_summary.txt ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ results/furuta_v4/refine_mlpA_ema999.npz
2
+ retrieval arm taken from results/furuta_v4/refine_mlpA_ema999.npz
3
+ K=1: generated plain 60.8 -> refined 95.0 | retrieval plain 54.2 -> refined 95.0
4
+ gen-refined minus retr-refined: -0.0 [-3.9,+4.2] p=1.0000
5
+ gen-plain minus retr-plain: +6.7 [-2.8,+16.7] p=0.2131
6
+ retr refined minus plain: +40.8 [+32.5,+50.0] p=0.0000
7
+ gen refined minus plain: +34.2 [+28.6,+40.0] p=0.0000
8
+ K=4: generated plain 87.8 -> refined 97.8 | retrieval plain 82.5 -> refined 97.5
9
+ gen-refined minus retr-refined: +0.3 [-1.7,+2.5] p=1.0000
10
+ gen-plain minus retr-plain: +5.3 [-0.6,+11.7] p=0.1138
11
+ retr refined minus plain: +15.0 [+9.2,+21.7] p=0.0000
12
+ gen refined minus plain: +10.0 [+6.1,+14.4] p=0.0000
results/furuta_v4/reg_attribution_partial.txt ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ == pooled success % (120 plants); per-training-seed values after the slash ==
2
+ method K=1 K=2 K=3 K=4 K=8
3
+ nearest+P retrieval 54.2 82.5 90.0
4
+ diffusion unet_a (9 runs; per train seed) 92.4 / 91.9/93.3/91.9 93.2 / 93.3/93.9/92.593.8 / 93.6/94.2/93.6
5
+ regression twin, deterministic (3 seeds) 83.6 / 85.0/82.5/83.3
6
+ regression ens3 (members as K candidates) 83.6 86.9 88.3
7
+ regression MC-dropout p=.1 (9 runs) 78.6 / 81.1/78.6/76.1 86.9 / 87.2/87.8/85.888.1 / 88.3/88.3/87.8
8
+ same net, dropout OFF (K=1) 83.9 / 84.2/85.8/81.7
9
+
10
+ == paired tests, per-plant seed-averaged difference A-B (pp), bootstrap 95% CI, sign-flip p ==
11
+ K=1: diffusion - retrieval +38.2 [+29.2,+46.9] p=0.0000
12
+ K=4: diffusion - retrieval +10.7 [ +5.0,+16.7] p=0.0006
13
+ K=8: diffusion - retrieval +3.8 [ +0.0, +8.0] p=0.0858
14
+ K=1: regression det - retrieval +29.4 [+20.8,+38.3] p=0.0000
15
+ K=1: diffusion - regression det +8.8 [ +4.9,+13.1] p=0.0001
16
+ K=1: diffusion - MC-dropout net +13.8 [ +9.5,+18.3] p=0.0000
17
+ K=1: regression det - MC-dropout net (K=1) +5.0 [ +2.6, +7.5] p=0.0000
18
+ K=4: diffusion - MC-dropout +6.3 [ +2.9,+10.5] p=0.0001
19
+ K=4: MC-dropout - retrieval +4.4 [ -0.6, +9.8] p=0.1119
20
+ K=8: diffusion - MC-dropout +5.6 [ +2.3, +9.8] p=0.0008
21
+ K=8: MC-dropout - retrieval -1.9 [ -6.0, +2.1] p=0.4133
22
+ ens3 K=3 - diffusion K=4 (diffusion gets MORE candidates) -4.9 [ -9.1, -1.6] p=0.0155
23
+ ens3 K=3 - regression det K=1 +4.7 [ +2.2, +7.5] p=0.0008
24
+ MC-dropout K=4 - regression det K=1 +3.3 [ +1.5, +5.6] p=0.0011
25
+ diffusion K=4 - diffusion K=1 +0.8 [ +0.2, +1.8] p=0.0626
26
+
27
+ == exact McNemar, single runs on 120 plants (b = diffusion only, c = regression only) ==
28
+ train seed 0: diffusion(samp0) 92.5 vs det 85.0 b=9 c=0 p=0.004
29
+ train seed 1: diffusion(samp0) 94.2 vs det 82.5 b=14 c=0 p=0.000
30
+ train seed 2: diffusion(samp0) 92.5 vs det 83.3 b=11 c=0 p=0.001
31
+ majority over runs: diffusion 92.5 vs det 84.2 b=10 c=0 p=0.002
32
+ K=1 majority diffusion vs retrieval b=47 c=1 p=3.5e-13
33
+ K=4 majority diffusion vs retrieval b=14 c=1 p=0.00098
34
+ K=1 majority regression det vs retrieval b=40 c=4 p=1.7e-08
35
+
36
+ diversity (pairwise RMS of 8 MC-dropout samples, normalised space, mean over plants): 0.037 (diffusion unet_a 0.029, bank NN 0.044)
results/furuta_v4/run_gates.sh ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+ # usage: run_gates.sh "name ckpt weights" ... ; sequential gate evaluations (CPU)
3
+ cd /home/cercola/matte_server/DiffusionControl; export PYTHONPATH=. OMP_NUM_THREADS=1 CUDA_VISIBLE_DEVICES=""
4
+ for spec in "$@"; do set -- $spec
5
+ [ -f results/furuta_v4/gate_$1.npz ] || myenv/bin/python -u -m nonlinear.experiments.furuta_v4_eval --ckpt $2 --weights $3 --name $1 --plants gate --workers 6 ${4:+--bank $4}
6
+ done
results/furuta_v4/scan_mlp_a.txt ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ mlp_a_s32000_ema999 K1=42.5 K4=81.2 K8=85.0 nearK1=57.5 nearK4=85.0 score=61.9
2
+ mlp_a_s32000_ema9995 K1=42.5 K4=80.0 K8=85.0 nearK1=57.5 nearK4=85.0 score=61.2
3
+ mlp_a_s40000_ema999 K1=46.2 K4=85.0 K8=88.8 nearK1=57.5 nearK4=85.0 score=65.6
4
+ mlp_a_s40000_ema9995 K1=43.8 K4=83.7 K8=88.8 nearK1=57.5 nearK4=85.0 score=63.7
5
+ mlp_a_s48000_ema999 K1=51.2 K4=85.0 K8=88.8 nearK1=57.5 nearK4=85.0 score=68.1
6
+ mlp_a_s48000_ema9995 K1=51.2 K4=86.2 K8=88.8 nearK1=57.5 nearK4=85.0 score=68.8
7
+ mlp_a_s56000_ema999 K1=60.0 K4=90.0 K8=91.3 nearK1=57.5 nearK4=85.0 score=75.0
8
+ mlp_a_s56000_ema9995 K1=57.5 K4=87.5 K8=91.3 nearK1=57.5 nearK4=85.0 score=72.5
9
+ mlp_a_s64000_ema999 K1=55.0 K4=88.8 K8=90.0 nearK1=57.5 nearK4=85.0 score=71.9
10
+ mlp_a_s64000_ema9995 K1=57.5 K4=88.8 K8=90.0 nearK1=57.5 nearK4=85.0 score=73.1
11
+ mlp_a_s72000_ema999 K1=57.5 K4=87.5 K8=90.0 nearK1=57.5 nearK4=85.0 score=72.5
12
+ mlp_a_s72000_ema9995 K1=58.8 K4=86.2 K8=88.8 nearK1=57.5 nearK4=85.0 score=72.5
13
+ mlp_a_s80000_ema999 K1=61.3 K4=88.8 K8=90.0 nearK1=57.5 nearK4=85.0 score=75.0
14
+ mlp_a_s80000_ema9995 K1=61.3 K4=90.0 K8=91.3 nearK1=57.5 nearK4=85.0 score=75.6
15
+ mlp_a_s88000_ema999 K1=62.5 K4=86.2 K8=88.8 nearK1=57.5 nearK4=85.0 score=74.4
16
+ mlp_a_s88000_ema9995 K1=61.3 K4=87.5 K8=88.8 nearK1=57.5 nearK4=85.0 score=74.4
17
+ mlp_a_s96000_ema999 K1=63.7 K4=90.0 K8=90.0 nearK1=57.5 nearK4=85.0 score=76.9
18
+ mlp_a_s96000_ema9995 K1=62.5 K4=90.0 K8=90.0 nearK1=57.5 nearK4=85.0 score=76.2
19
+ BEST mlp_a_s96000_ema999 K1=63.7 K4=90.0 K8=90.0 nearK1=57.5 nearK4=85.0 score=76.9
results/furuta_v4/scan_unet_a.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ unet_a_s40000_ema999 K1=87.5 K4=100.0 K8=100.0 nearK1=57.5 nearK4=85.0 score=93.8
2
+ unet_a_s48000_ema999 K1=86.2 K4=100.0 K8=100.0 nearK1=57.5 nearK4=85.0 score=93.1
3
+ unet_a_s56000_ema999 K1=95.0 K4=100.0 K8=100.0 nearK1=57.5 nearK4=85.0 score=97.5
4
+ unet_a_s64000_ema999 K1=97.5 K4=100.0 K8=100.0 nearK1=57.5 nearK4=85.0 score=98.8
5
+ unet_a_s72000_ema999 K1=98.8 K4=100.0 K8=100.0 nearK1=57.5 nearK4=85.0 score=99.4
6
+ unet_a_s80000_ema999 K1=95.0 K4=100.0 K8=100.0 nearK1=57.5 nearK4=85.0 score=97.5
7
+ unet_a_s88000_ema999 K1=96.2 K4=100.0 K8=100.0 nearK1=57.5 nearK4=85.0 score=98.1
8
+ unet_a_s96000_ema999 K1=97.5 K4=100.0 K8=100.0 nearK1=57.5 nearK4=85.0 score=98.8
9
+ BEST unet_a_s72000_ema999 K1=98.8 K4=100.0 K8=100.0 nearK1=57.5 nearK4=85.0 score=99.4
results/furuta_v4/traj_gen_warmstart_v4unetA_s0.txt ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === warm-start test (n=60, seed 4242, max_iter=700) ===
2
+ target certificate: largest_certified_region per plant (unmargined); verifier: furuta_traj_bank.score_solution (identical to every other arm in this doc)
3
+
4
+ condition success med iter med t med J
5
+ cold 53.3% 363 1.95s 4.06
6
+ nearest-warm 76.7% 45 0.25s 4.15
7
+ generated-warm 76.7% 42 0.25s 4.15
8
+
9
+ --- paired tests ---
10
+ generated-warm vs cold success: beats 14, loses 0, McNemar p=0.0001 * time ratio 0.152x (n=32), Wilcoxon p=0.0000 *
11
+ generated-warm vs nearest-warm success: beats 0, loses 0, McNemar p=1.0000 time ratio 0.959x (n=46), Wilcoxon p=0.3024
12
+ nearest-warm vs cold success: beats 14, loses 0, McNemar p=0.0001 * time ratio 0.157x (n=32), Wilcoxon p=0.0000 *
13
+
14
+ --- failure breakdown ---
15
+ cold ipopt: Maximum_Iterations_Exceeded:13 ipopt: Infeasible_Problem_Detected:11 rollout failed:4
16
+ nearest-warm ipopt: Infeasible_Problem_Detected:6 rollout failed:5 ipopt: Maximum_Iterations_Exceeded:3
17
+ generated-warm ipopt: Infeasible_Problem_Detected:6 ipopt: Maximum_Iterations_Exceeded:5 rollout failed:3
results/furuta_v4/traj_gen_warmstart_v4unetA_s1.txt ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === warm-start test (n=60, seed 4242, max_iter=700) ===
2
+ target certificate: largest_certified_region per plant (unmargined); verifier: furuta_traj_bank.score_solution (identical to every other arm in this doc)
3
+
4
+ condition success med iter med t med J
5
+ cold 53.3% 363 1.98s 4.06
6
+ nearest-warm 76.7% 45 0.26s 4.15
7
+ generated-warm 76.7% 46 0.26s 4.15
8
+
9
+ --- paired tests ---
10
+ generated-warm vs cold success: beats 14, loses 0, McNemar p=0.0001 * time ratio 0.159x (n=32), Wilcoxon p=0.0000 *
11
+ generated-warm vs nearest-warm success: beats 0, loses 0, McNemar p=1.0000 time ratio 1.023x (n=46), Wilcoxon p=0.5957
12
+ nearest-warm vs cold success: beats 14, loses 0, McNemar p=0.0001 * time ratio 0.161x (n=32), Wilcoxon p=0.0000 *
13
+
14
+ --- failure breakdown ---
15
+ cold ipopt: Maximum_Iterations_Exceeded:13 ipopt: Infeasible_Problem_Detected:11 rollout failed:4
16
+ nearest-warm ipopt: Infeasible_Problem_Detected:6 rollout failed:5 ipopt: Maximum_Iterations_Exceeded:3
17
+ generated-warm ipopt: Maximum_Iterations_Exceeded:6 ipopt: Infeasible_Problem_Detected:5 rollout failed:3
results/guidance/A16x_results.txt ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Task A budget-scaled sensitivity: 16x budget (evals/arm = 7688), arms a,d,e,f, 240 units
2
+ u_max=0.6mgl: a: cov 0.308 div 0.120 union 0.312 cost 16.68 wall 92s d: cov 0.524 div 0.011 union 0.524 cost 4.81 wall 21s e: cov 0.562 div 0.183 union 0.565 cost 4.94 wall 21s f: cov 0.539 div 0.165 union 0.541 cost 4.92 wall 106s
3
+ d vs e: diff -0.0387 p=1.53e-05 wins/ties/losses 3/13/24
4
+ d vs a: diff +0.2161 p=1.14e-07 wins/ties/losses 37/3/0
5
+ d vs f: diff -0.0154 p=4.89e-05 wins/ties/losses 2/17/21
6
+ e vs a: diff +0.2547 p=5.26e-08 wins/ties/losses 39/1/0
7
+ e vs f: diff +0.0233 p=0.0251 wins/ties/losses 22/6/12
8
+ u_max=0.9mgl: a: cov 0.822 div 0.218 union 0.845 cost 5.88 wall 103s d: cov 0.875 div 0.010 union 0.875 cost 3.83 wall 17s e: cov 0.911 div 0.192 union 0.916 cost 3.97 wall 15s f: cov 0.877 div 0.100 union 0.878 cost 3.93 wall 53s
9
+ d vs e: diff -0.0364 p=2.7e-05 wins/ties/losses 0/17/23
10
+ d vs a: diff +0.0527 p=2.01e-05 wins/ties/losses 28/8/4
11
+ d vs f: diff -0.0026 p=0.0408 wins/ties/losses 4/29/7
12
+ e vs a: diff +0.0891 p=1.15e-07 wins/ties/losses 37/2/1
13
+ e vs f: diff +0.0338 p=1.18e-05 wins/ties/losses 25/14/1
results/guidance/A_ddim_check.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ DDIM20 vs DDPM500 (pendulum, 10 plants, u_max=0.9mgl, 41x41 grid coverage): best-of-8 DDPM 0.887 DDIM 0.887; mean-of-8 DDPM 0.887 DDIM 0.881
results/guidance/A_results.txt ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Task A: torque-bounded pendulum, N=40 plants (random_pendulum_params seed 321), seeds (0, 1, 2), K=8, 240 units
2
+ budget: evals/arm = 3*K*S+K = 488 (S=20); backward=2 forward-equivalents; 1 eval = 1 controller x 80 ICs
3
+
4
+ ===== u_max = 0.6 * m g l (units 120) =====
5
+ arm coverage union(K) in-reg cost cert diversity wall s evals
6
+ retrieval K=8 (no search) 0.524+-0.000[0.48,0.57] 0.524 4.81 1.00 0.068 0.1 8
7
+ (a) retrieval + steering 0.524+-0.000[0.48,0.57] 0.524 4.92 0.06 0.068 5.0 488
8
+ prior K=8 (no search) 0.524+-0.000[0.48,0.57] 0.524 4.80 0.98 0.018 0.1 8
9
+ (b) prior + steering 0.524+-0.000[0.48,0.57] 0.524 4.95 0.11 0.017 4.9 488
10
+ (c) in-denoising gradient guidance 0.523+-0.000[0.47,0.57] 0.523 4.81 0.21 0.032 6.0 488
11
+ (d) SMC gradient-free guidance 0.524+-0.000[0.48,0.57] 0.524 4.80 1.00 0.008 1.2 488
12
+ (e) CEM on retrieval 0.538+-0.006[0.49,0.58] 0.539 4.91 0.20 0.198 1.2 488
13
+ prior best-of-488 0.524+-0.000[0.48,0.57] 0.524 4.82 0.72 0.046 1.5 488
14
+ paired Wilcoxon on per-plant coverage (seed-mean over 3 seeds), 40 plants; diff = arm - ref, two-sided p
15
+ b vs a: diff +0.0000 [+0.0000,+0.0000] p=1 wins/ties/losses 0/40/0
16
+ b vs e: diff -0.0141 [-0.0250,-0.0057] p=0.000155 wins/ties/losses 4/18/18
17
+ c vs a: diff -0.0007 [-0.0018,+0.0000] p=0.109 wins/ties/losses 0/37/3
18
+ c vs e: diff -0.0148 [-0.0259,-0.0062] p=6.33e-05 wins/ties/losses 3/17/20
19
+ d vs a: diff +0.0000 [+0.0000,+0.0000] p=1 wins/ties/losses 0/40/0
20
+ d vs e: diff -0.0141 [-0.0250,-0.0057] p=0.000155 wins/ties/losses 4/18/18
21
+ d vs f: diff -0.0000 p=0.317
22
+ e vs a: diff +0.0141 p=0.000155
23
+ e vs f: diff +0.0141 p=0.000166
24
+ PRE-REGISTERED (c or d beats BOTH a and e, p<0.05, paired): c=False d=False
25
+
26
+ ===== u_max = 0.9 * m g l (units 120) =====
27
+ arm coverage union(K) in-reg cost cert diversity wall s evals
28
+ retrieval K=8 (no search) 0.875+-0.000[0.85,0.90] 0.875 3.84 1.00 0.068 0.1 8
29
+ (a) retrieval + steering 0.875+-0.000[0.85,0.90] 0.875 3.99 0.00 0.068 4.9 488
30
+ prior K=8 (no search) 0.875+-0.000[0.85,0.90] 0.875 3.84 0.97 0.019 0.1 8
31
+ (b) prior + steering 0.875+-0.000[0.85,0.90] 0.875 4.02 0.05 0.018 4.9 488
32
+ (c) in-denoising gradient guidance 0.874+-0.001[0.85,0.90] 0.874 4.01 0.00 0.045 6.1 488
33
+ (d) SMC gradient-free guidance 0.875+-0.000[0.85,0.90] 0.875 3.83 1.00 0.008 1.2 488
34
+ (e) CEM on retrieval 0.881+-0.001[0.86,0.90] 0.882 3.96 0.19 0.168 1.2 488
35
+ prior best-of-488 0.875+-0.000[0.85,0.90] 0.875 3.93 0.77 0.039 1.5 488
36
+ paired Wilcoxon on per-plant coverage (seed-mean over 3 seeds), 40 plants; diff = arm - ref, two-sided p
37
+ b vs a: diff +0.0000 [+0.0000,+0.0000] p=1 wins/ties/losses 0/40/0
38
+ b vs e: diff -0.0061 [-0.0107,-0.0024] p=0.00444 wins/ties/losses 1/29/10
39
+ c vs a: diff -0.0011 [-0.0032,+0.0000] p=0.18 wins/ties/losses 0/38/2
40
+ c vs e: diff -0.0072 [-0.0130,-0.0026] p=0.0037 wins/ties/losses 1/28/11
41
+ d vs a: diff +0.0000 [+0.0000,+0.0000] p=1 wins/ties/losses 0/40/0
42
+ d vs e: diff -0.0061 [-0.0107,-0.0024] p=0.00444 wins/ties/losses 1/29/10
43
+ d vs f: diff +0.0000 p=1
44
+ e vs a: diff +0.0061 p=0.00444
45
+ e vs f: diff +0.0061 p=0.00444
46
+ PRE-REGISTERED (c or d beats BOTH a and e, p<0.05, paired): c=False d=False
47
+
results/guidance/B_ddim_check.txt ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ SISO 30 plants K=8 (search ICs): DDPM500 best J_bb 15.67 stable-any 0.70 per-sample stable 0.542
2
+ DDIM20 eta=0.0: best J_bb 15.62 stable-any 0.67 per-sample stable 0.542
3
+ DDIM20 eta=1.0: best J_bb 17.19 stable-any 0.70 per-sample stable 0.567
4
+ DDIM50 eta=0.0: best J_bb 16.64 stable-any 0.70 per-sample stable 0.512
5
+ DDIM50 eta=1.0: best J_bb 16.10 stable-any 0.67 per-sample stable 0.533
results/guidance/B_results.txt ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Task B: SISO black-box J_bb, N=60 plants (gen_siso_plants seed 100), seeds 3, K=8, budget 200 evals/plant (spec: settle 0.05*peak, opposite-side excursion>0.5*peak -> +12.0s, fail=36.0)
2
+ arm J_bb (search ICs) J_bb held-out 16 ICs stable(held-out) evals wall s
3
+ distinct retrieval K=8 (no search) 16.046+-0.000[13.79,18.32] 16.753+-0.000[14.75,18.82] 0.483 8 0.6
4
+ distinct retrieval + CEM 15.179+-0.414[13.02,17.36] 16.855+-0.412[14.94,18.87] 0.428 200 15.4
5
+ retrieval best-of-200 (bank rows) 16.422+-0.000[14.12,18.70] 17.189+-0.000[15.21,19.20] 0.483 200 13.4
6
+ diffusion K=8 (no search) 17.020+-0.302[15.08,19.02] 18.835+-0.412[17.05,20.64] 0.533 8 0.6
7
+ diffusion best-of-200 (no guidance) 11.311+-0.079[9.35,13.37] 15.494+-0.442[13.38,17.67] 0.522 200 179.2
8
+ diffusion + SMC guidance 13.970+-0.096[12.06,16.00] 17.452+-0.123[15.43,19.54] 0.472 200 35.8
9
+ paired Wilcoxon on per-plant J_bb (seed-mean), negative diff = first arm better (lower J)
10
+ [search] ps vs rc: diff -1.209 [-2.232,-0.179] p=0.0407 wins/ties/losses 33/0/27
11
+ [search] ps vs rm: diff -2.452 [-3.508,-1.384] p=5.83e-05 wins/ties/losses 41/0/19
12
+ [search] ps vs pm: diff +2.660 [+2.270,+3.050] p=3.29e-11 wins/ties/losses 2/0/58
13
+ [held-out] ps vs rc: diff +0.597 [-0.657,+1.848] p=0.484 wins/ties/losses 29/0/31
14
+ [held-out] ps vs rm: diff +0.263 [-0.859,+1.413] p=0.546 wins/ties/losses 30/0/30
15
+ [held-out] ps vs pm: diff +1.958 [+1.245,+2.691] p=8.73e-06 wins/ties/losses 13/0/47
16
+ PRE-REGISTERED (ps beats rc, rm AND pm on search J_bb, p<0.05 paired): False
results/guidance/pilot_choice.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "lr": 0.5,
3
+ "lam": 0.03,
4
+ "means": {
5
+ "ret0": -0.5157668590545654,
6
+ "pri0": -0.5157651901245117,
7
+ "steer_ret_lr0.1": -0.5157667398452759,
8
+ "steer_pri_lr0.1": -0.5157657861709595,
9
+ "steer_ret_lr0.5": -0.518862247467041,
10
+ "steer_pri_lr0.5": -0.51885986328125,
11
+ "steer_ret_lr2.0": -0.40492570400238037,
12
+ "steer_pri_lr2.0": -0.37729185819625854,
13
+ "guide_lam0.03": -0.5188661813735962,
14
+ "guide_lam0.1": -0.5141855478286743,
15
+ "guide_lam0.3": -0.3907387852668762
16
+ }
17
+ }
results/icml_r3/RESUME.md ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Round 3 resume note (written 2026-10-13 02:45 PT, in case the orchestrator session ends)
2
+
3
+ State at writing:
4
+ - DONE on virgo (CPU): X4 verifier audit, X9 multimodality, X7 + X10 SOF baselines. Numbers are in
5
+ `paper_icml/plan/inventory/r3.md`; claim-hierarchy, experiments-needed and terminology are updated.
6
+ - RUNNING detached on andromeda (setsid/nohup, PPID 1, `timeout` ending 08:20 PT; watchdog
7
+ `~/icml_r3_pids/watchdog.sh` TERMs recorded PIDs at 08:25 PT):
8
+ - X2 (Furuta, corrected arm): 3 U-Net seeds + MC-dropout twin training. Post-training steps
9
+ are scripted as a detached driver; see `X2/RESUME.md`.
10
+ - X1S + X3 (SOF multi-hypothesis arms + seeds): `X1S/chain_grid.sh` chains; see `X1S/RESUME.md`.
11
+ - X1F (Furuta multi-hypothesis arms): `X1F/x1f_queue.py`; see `X1F/RESUME.md`.
12
+ - PIDs: `andromeda:~/icml_r3_pids/{X2,X1S,X1F}.txt`.
13
+ - andromeda use (GPU and CPU) ends 08:30 PT 2026-10-13. Never touch processes not in those PID files
14
+ (another user, and the user's own `claude` session there).
15
+
16
+ Next session, in order:
17
+ 1. After 08:30 PT: `ssh andromeda 'ps -u cercola -o pid,etime,args'`. Nothing from the PID files may be
18
+ left; check `~/icml_r3_pids/watchdog.log`.
19
+ 2. Copy results back (file reads only, no compute):
20
+ `rsync -a andromeda:~/DiffusionControl/results/icml_r3/{X2,X1S,X1F}/ results/icml_r3/<same>/`, plus
21
+ any `models/` checkpoints the per-experiment RESUME.md names.
22
+ 3. Any step a RESUME.md lists as "not done" runs on virgo CPU only (the virgo GPU is NOT authorized).
23
+ 4. Write docs/icml_r3_x2.md, icml_r3_x1_sof.md, icml_r3_x1_furuta.md (research-doc skill) if the agents
24
+ did not finish them; add the numbers to `paper_icml/plan/inventory/r3.md`.
25
+ 5. Then continue the paper plan: rebuild Fig 2/3 and Table 1 (retrieval = learned ARD metric, triage
26
+ rows 1-3), build Fig 4 (X9 index vs margins), then plan and draft Experiments via the zeroshot loop.
27
+ - X11 (idle-GPU timing) waits for the user's go. Until then, no time claims.
28
+
29
+ ## Update 2026-10-13 07:20 PT: ALL ROUND-3 COMPUTE DONE
30
+ - X2, X1S/X3 and X1F drivers finished by 06:30 PT. The orchestrator stopped the idle X1S queue runner and the
31
+ watchdog at ~07:05; andromeda has no processes from this work. Results and checkpoints (models/mhx,
32
+ models/x1f, models/v4/x2_*) were copied back to virgo.
33
+ - Numbers: paper_icml/plan/inventory/r3.md (X2a-k, X1Sa-c, X1Fa-c, X3a-b). claim-hierarchy C1/C2 updated.
34
+ - In progress (Sonnet agents): docs/icml_r3_{x2,x1_sof,x1_furuta}.md; Fig 2/3 + Table 1 rebuild + appendix X1 figures.
35
+ - Optional cleanup (ask the user first): andromeda models/x1f/*_resume.pt and models/mhx scratch.
results/icml_r3/RULES.md ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Shared rules for every agent in the ICML round-3 runs (read fully before doing anything)
2
+
3
+ Repo: /home/cercola/matte_server/DiffusionControl (branch icml-experiments; do NOT commit or push).
4
+ Paper: ICML 2027. Plan: paper_icml/plan/{claim-hierarchy,experiments-needed}.md; per-area evidence in
5
+ paper_icml/plan/inventory/*.md (read the parts relevant to you). Round-2 rules and lessons:
6
+ results/icml_r2/RULES.md (the clip bug, recipes, statistics) - read it too, but THIS file overrides it.
7
+ Make the method as strong as possible: baselines are targets, not verdicts; diagnose before concluding;
8
+ report honestly. Same backbone, data, steps, EMA and verifier for every arm you compare.
9
+
10
+ ## Compute (authorized by the user for this window only)
11
+ - GPU: ONLY andromeda's RTX 5090 (ssh andromeda). The user authorized it until 08:30 PT 2026-10-13.
12
+ Launch EVERY GPU process under a hard timeout that ends by 08:20 PT:
13
+ T=$(( $(TZ=America/Los_Angeles date -d 'today 08:20' +%s) - $(date +%s) )); setsid nohup timeout --signal=TERM $T <cmd> > <log> 2>&1 < /dev/null & disown
14
+ Checkpoint at least every ~30 min so runs can resume in a later window. No GPU launches after 08:00 PT.
15
+ - The GPU is shared with other agents of this session and possibly other users. Your VRAM budget is stated
16
+ in your task; check `nvidia-smi --query-gpu=memory.free --format=csv,noheader,nounits` before each launch
17
+ and keep >= 3000 MiB free. NEVER kill/stop/pause a process you did not start. Record your PIDs in
18
+ ~/icml_r3_pids/<your agent id>.txt on the machine where they run; kill only by those PIDs.
19
+ - virgo (local): CPU ONLY. Never use the virgo GPU (export CUDA_VISIBLE_DEVICES= for every virgo job).
20
+ Another user is running jobs on virgo; respect your worker cap.
21
+ - andromeda CPU: 32 threads; respect your stated thread cap. The user's window covers ALL of andromeda (CPU too): every andromeda process of yours must finish or be stopped by 08:30 PT (use the same timeout for CPU jobs there); anything still needed after that moves to virgo CPU. andromeda disk is 99% full (~54 GB free):
22
+ keep your andromeda outputs under 8 GB, never duplicate big datasets, delete your own scratch when done.
23
+
24
+ ## Code and sync
25
+ - Local python: myenv/bin/python (PYTHONPATH=. OMP_NUM_THREADS=1). andromeda: ~/DiffusionControl (rsync copy,
26
+ not git; code synced 2026-10-13 00:15), venv ~/DiffusionControl/.venv (torch cu128, numpy 1.26: use the
27
+ np.trapezoid = np.trapz shim when a module needs it), run with WANDB_MODE=offline.
28
+ - Other agents edit code concurrently: write NEW files (or opt-in flags in files nobody else touches); never
29
+ change defaults of shared modules. Sync to andromeda ONLY the specific files/dirs you created or need
30
+ (rsync -az -e ssh <path> andromeda:~/DiffusionControl/<path>); never a whole-repo rsync, never --delete.
31
+
32
+ ## Protocol
33
+ - Before scoring any test set, write results/icml_r3/<ID>/PREREG.md: arms, n, plants (seeds), metric,
34
+ test, multiplicity correction. Choose configs on validation plants only. Held-out plants never in a bank.
35
+ - Paired tests (exact McNemar / sign-flip permutation), bootstrap CIs over plants; when several training
36
+ seeds exist, report per-seed results and treat the SEED as the resampling unit for the across-seed
37
+ claim (no pooled-over-seeds McNemar). Name the bank (size, file) next to every retrieval number.
38
+ - Outputs: results/icml_r3/<ID>/ (copy results back from andromeda), figures under figures/icml_r3/ with
39
+ NO titles or in-axes text (explanations go in captions), and a doc docs/icml_r3_<id>.md written with the
40
+ research-doc skill (~/.claude/skills/research-doc/SKILL.md).
41
+ - Every job > 2 min must survive session death (setsid nohup ... & disown) and be resumable; poll with
42
+ sleep <= 300 loops.
43
+ - Return a SHORT report (<= 25 lines): headline numbers with n/CI/p, paths, what is still running
44
+ (pid, log, expected end), what is blocked and why.
results/ood/furuta_ood.txt ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === Furuta lvl0: N=60, NN ratio 1.0x; per-plant fresh-solve success 0.950 ===
2
+ arm K=1 K=4 K=8
3
+ nn 0.567 0.850 0.983
4
+ idw 0.567 0.800 0.967
5
+ gen_v3_s0 0.333 0.800 0.867
6
+ gen_v3_s1 0.417 0.800 0.850
7
+ gen_v3_s2 0.533 0.817 0.917
8
+ gen_v3e_s0 0.433 0.850 0.883
9
+ gen_v3e_s1 0.550 0.867 0.917
10
+ gen_v3e_s2 0.483 0.833 0.900
11
+ gen_v3e_t1_s0 0.583 0.867 0.900
12
+ K=1 gen_v3_s0 vs best retrieval (nn): model-only 8 retr-only 22 p=0.0161 diff -0.233
13
+ K=1 gen_v3_s1 vs best retrieval (nn): model-only 8 retr-only 17 p=0.108 diff -0.150
14
+ K=1 gen_v3_s2 vs best retrieval (nn): model-only 9 retr-only 11 p=0.824 diff -0.033
15
+ K=1 gen_v3e_s0 vs best retrieval (nn): model-only 10 retr-only 18 p=0.185 diff -0.133
16
+ K=1 gen_v3e_s1 vs best retrieval (nn): model-only 12 retr-only 13 p=1 diff -0.017
17
+ K=1 gen_v3e_s2 vs best retrieval (nn): model-only 8 retr-only 13 p=0.383 diff -0.083
18
+ K=1 gen_v3e_t1_s0 vs best retrieval (nn): model-only 12 retr-only 11 p=1 diff +0.017
19
+ K=4 gen_v3_s0 vs best retrieval (nn): model-only 4 retr-only 7 p=0.549 diff -0.050
20
+ K=4 gen_v3_s1 vs best retrieval (nn): model-only 3 retr-only 6 p=0.508 diff -0.050
21
+ K=4 gen_v3_s2 vs best retrieval (nn): model-only 4 retr-only 6 p=0.754 diff -0.033
22
+ K=4 gen_v3e_s0 vs best retrieval (nn): model-only 5 retr-only 5 p=1 diff +0.000
23
+ K=4 gen_v3e_s1 vs best retrieval (nn): model-only 4 retr-only 3 p=1 diff +0.017
24
+ K=4 gen_v3e_s2 vs best retrieval (nn): model-only 5 retr-only 6 p=1 diff -0.017
25
+ K=4 gen_v3e_t1_s0 vs best retrieval (nn): model-only 5 retr-only 4 p=1 diff +0.017
26
+ K=8 gen_v3_s0 vs best retrieval (nn): model-only 0 retr-only 7 p=0.0156 diff -0.117
27
+ K=8 gen_v3_s1 vs best retrieval (nn): model-only 0 retr-only 8 p=0.00781 diff -0.133
28
+ K=8 gen_v3_s2 vs best retrieval (nn): model-only 0 retr-only 4 p=0.125 diff -0.067
29
+ K=8 gen_v3e_s0 vs best retrieval (nn): model-only 0 retr-only 6 p=0.0312 diff -0.100
30
+ K=8 gen_v3e_s1 vs best retrieval (nn): model-only 0 retr-only 4 p=0.125 diff -0.067
31
+ K=8 gen_v3e_s2 vs best retrieval (nn): model-only 0 retr-only 5 p=0.0625 diff -0.083
32
+ K=8 gen_v3e_t1_s0 vs best retrieval (nn): model-only 0 retr-only 5 p=0.0625 diff -0.083
33
+
34
+ === Furuta lvl1.1: N=60, NN ratio 1.3x; per-plant fresh-solve success 0.833 ===
35
+ arm K=1 K=4 K=8
36
+ nn 0.400 0.550 0.650
37
+ idw 0.400 0.533 0.617
38
+ gen_v3_s0 0.267 0.633 0.667
39
+ gen_v3_s1 0.217 0.550 0.700
40
+ gen_v3_s2 0.300 0.583 0.633
41
+ gen_v3e_s0 0.350 0.733 0.750
42
+ gen_v3e_s1 0.400 0.733 0.767
43
+ gen_v3e_s2 0.433 0.683 0.733
44
+ gen_v3e_t1_s0 0.400 0.733 0.783
45
+ K=1 gen_v3_s0 vs best retrieval (nn): model-only 8 retr-only 16 p=0.152 diff -0.133
46
+ K=1 gen_v3_s1 vs best retrieval (nn): model-only 4 retr-only 15 p=0.0192 diff -0.183
47
+ K=1 gen_v3_s2 vs best retrieval (nn): model-only 6 retr-only 12 p=0.238 diff -0.100
48
+ K=1 gen_v3e_s0 vs best retrieval (nn): model-only 10 retr-only 13 p=0.678 diff -0.050
49
+ K=1 gen_v3e_s1 vs best retrieval (nn): model-only 12 retr-only 12 p=1 diff +0.000
50
+ K=1 gen_v3e_s2 vs best retrieval (nn): model-only 10 retr-only 8 p=0.815 diff +0.033
51
+ K=1 gen_v3e_t1_s0 vs best retrieval (nn): model-only 10 retr-only 10 p=1 diff +0.000
52
+ K=4 gen_v3_s0 vs best retrieval (nn): model-only 9 retr-only 4 p=0.267 diff +0.083
53
+ K=4 gen_v3_s1 vs best retrieval (nn): model-only 9 retr-only 9 p=1 diff +0.000
54
+ K=4 gen_v3_s2 vs best retrieval (nn): model-only 6 retr-only 4 p=0.754 diff +0.033
55
+ K=4 gen_v3e_s0 vs best retrieval (nn): model-only 12 retr-only 1 p=0.00342 diff +0.183
56
+ K=4 gen_v3e_s1 vs best retrieval (nn): model-only 12 retr-only 1 p=0.00342 diff +0.183
57
+ K=4 gen_v3e_s2 vs best retrieval (nn): model-only 11 retr-only 3 p=0.0574 diff +0.133
58
+ K=4 gen_v3e_t1_s0 vs best retrieval (nn): model-only 12 retr-only 1 p=0.00342 diff +0.183
59
+ K=8 gen_v3_s0 vs best retrieval (nn): model-only 5 retr-only 4 p=1 diff +0.017
60
+ K=8 gen_v3_s1 vs best retrieval (nn): model-only 6 retr-only 3 p=0.508 diff +0.050
61
+ K=8 gen_v3_s2 vs best retrieval (nn): model-only 2 retr-only 3 p=1 diff -0.017
62
+ K=8 gen_v3e_s0 vs best retrieval (nn): model-only 7 retr-only 1 p=0.0703 diff +0.100
63
+ K=8 gen_v3e_s1 vs best retrieval (nn): model-only 8 retr-only 1 p=0.0391 diff +0.117
64
+ K=8 gen_v3e_s2 vs best retrieval (nn): model-only 8 retr-only 3 p=0.227 diff +0.083
65
+ K=8 gen_v3e_t1_s0 vs best retrieval (nn): model-only 9 retr-only 1 p=0.0215 diff +0.133
66
+
67
+ === Furuta lvl1.25: N=60, NN ratio 2.1x; per-plant fresh-solve success 0.750 ===
68
+ arm K=1 K=4 K=8
69
+ nn 0.217 0.533 0.633
70
+ idw 0.217 0.483 0.600
71
+ gen_v3_s0 0.233 0.567 0.600
72
+ gen_v3_s1 0.267 0.550 0.583
73
+ gen_v3_s2 0.317 0.483 0.617
74
+ gen_v3e_s0 0.317 0.567 0.617
75
+ gen_v3e_s1 0.217 0.500 0.567
76
+ gen_v3e_s2 0.300 0.517 0.617
77
+ gen_v3e_t1_s0 0.333 0.633 0.650
78
+ K=1 gen_v3_s0 vs best retrieval (nn): model-only 8 retr-only 7 p=1 diff +0.017
79
+ K=1 gen_v3_s1 vs best retrieval (nn): model-only 12 retr-only 9 p=0.664 diff +0.050
80
+ K=1 gen_v3_s2 vs best retrieval (nn): model-only 14 retr-only 8 p=0.286 diff +0.100
81
+ K=1 gen_v3e_s0 vs best retrieval (nn): model-only 13 retr-only 7 p=0.263 diff +0.100
82
+ K=1 gen_v3e_s1 vs best retrieval (nn): model-only 10 retr-only 10 p=1 diff +0.000
83
+ K=1 gen_v3e_s2 vs best retrieval (nn): model-only 13 retr-only 8 p=0.383 diff +0.083
84
+ K=1 gen_v3e_t1_s0 vs best retrieval (nn): model-only 14 retr-only 7 p=0.189 diff +0.117
85
+ K=4 gen_v3_s0 vs best retrieval (nn): model-only 8 retr-only 6 p=0.791 diff +0.033
86
+ K=4 gen_v3_s1 vs best retrieval (nn): model-only 7 retr-only 6 p=1 diff +0.017
87
+ K=4 gen_v3_s2 vs best retrieval (nn): model-only 6 retr-only 9 p=0.607 diff -0.050
88
+ K=4 gen_v3e_s0 vs best retrieval (nn): model-only 8 retr-only 6 p=0.791 diff +0.033
89
+ K=4 gen_v3e_s1 vs best retrieval (nn): model-only 5 retr-only 7 p=0.774 diff -0.033
90
+ K=4 gen_v3e_s2 vs best retrieval (nn): model-only 7 retr-only 8 p=1 diff -0.017
91
+ K=4 gen_v3e_t1_s0 vs best retrieval (nn): model-only 9 retr-only 3 p=0.146 diff +0.100
92
+ K=8 gen_v3_s0 vs best retrieval (nn): model-only 2 retr-only 4 p=0.688 diff -0.033
93
+ K=8 gen_v3_s1 vs best retrieval (nn): model-only 2 retr-only 5 p=0.453 diff -0.050
94
+ K=8 gen_v3_s2 vs best retrieval (nn): model-only 3 retr-only 4 p=1 diff -0.017
95
+ K=8 gen_v3e_s0 vs best retrieval (nn): model-only 3 retr-only 4 p=1 diff -0.017
96
+ K=8 gen_v3e_s1 vs best retrieval (nn): model-only 3 retr-only 7 p=0.344 diff -0.067
97
+ K=8 gen_v3e_s2 vs best retrieval (nn): model-only 3 retr-only 4 p=1 diff -0.017
98
+ K=8 gen_v3e_t1_s0 vs best retrieval (nn): model-only 4 retr-only 3 p=1 diff +0.017
99
+
100
+ === Furuta lvl1.5: N=60, NN ratio 3.1x; per-plant fresh-solve success 0.667 ===
101
+ arm K=1 K=4 K=8
102
+ nn 0.233 0.433 0.533
103
+ idw 0.233 0.383 0.500
104
+ gen_v3_s0 0.167 0.417 0.483
105
+ gen_v3_s1 0.117 0.367 0.483
106
+ gen_v3_s2 0.183 0.417 0.517
107
+ gen_v3e_s0 0.267 0.400 0.517
108
+ gen_v3e_s1 0.167 0.450 0.500
109
+ gen_v3e_s2 0.183 0.450 0.550
110
+ gen_v3e_t1_s0 0.300 0.483 0.517
111
+ K=1 gen_v3_s0 vs best retrieval (nn): model-only 7 retr-only 11 p=0.481 diff -0.067
112
+ K=1 gen_v3_s1 vs best retrieval (nn): model-only 6 retr-only 13 p=0.167 diff -0.117
113
+ K=1 gen_v3_s2 vs best retrieval (nn): model-only 5 retr-only 8 p=0.581 diff -0.050
114
+ K=1 gen_v3e_s0 vs best retrieval (nn): model-only 9 retr-only 7 p=0.804 diff +0.033
115
+ K=1 gen_v3e_s1 vs best retrieval (nn): model-only 8 retr-only 12 p=0.503 diff -0.067
116
+ K=1 gen_v3e_s2 vs best retrieval (nn): model-only 5 retr-only 8 p=0.581 diff -0.050
117
+ K=1 gen_v3e_t1_s0 vs best retrieval (nn): model-only 10 retr-only 6 p=0.454 diff +0.067
118
+ K=4 gen_v3_s0 vs best retrieval (nn): model-only 5 retr-only 6 p=1 diff -0.017
119
+ K=4 gen_v3_s1 vs best retrieval (nn): model-only 4 retr-only 8 p=0.388 diff -0.067
120
+ K=4 gen_v3_s2 vs best retrieval (nn): model-only 6 retr-only 7 p=1 diff -0.017
121
+ K=4 gen_v3e_s0 vs best retrieval (nn): model-only 5 retr-only 7 p=0.774 diff -0.033
122
+ K=4 gen_v3e_s1 vs best retrieval (nn): model-only 7 retr-only 6 p=1 diff +0.017
123
+ K=4 gen_v3e_s2 vs best retrieval (nn): model-only 7 retr-only 6 p=1 diff +0.017
124
+ K=4 gen_v3e_t1_s0 vs best retrieval (nn): model-only 8 retr-only 5 p=0.581 diff +0.050
125
+ K=8 gen_v3_s0 vs best retrieval (nn): model-only 3 retr-only 6 p=0.508 diff -0.050
126
+ K=8 gen_v3_s1 vs best retrieval (nn): model-only 4 retr-only 7 p=0.549 diff -0.050
127
+ K=8 gen_v3_s2 vs best retrieval (nn): model-only 2 retr-only 3 p=1 diff -0.017
128
+ K=8 gen_v3e_s0 vs best retrieval (nn): model-only 4 retr-only 5 p=1 diff -0.017
129
+ K=8 gen_v3e_s1 vs best retrieval (nn): model-only 4 retr-only 6 p=0.754 diff -0.033
130
+ K=8 gen_v3e_s2 vs best retrieval (nn): model-only 4 retr-only 3 p=1 diff +0.017
131
+ K=8 gen_v3e_t1_s0 vs best retrieval (nn): model-only 4 retr-only 5 p=1 diff -0.017
132
+
133
+ === Furuta lvl2.0: N=60, NN ratio 6.2x; per-plant fresh-solve success 0.533 ===
134
+ arm K=1 K=4 K=8
135
+ nn 0.083 0.217 0.267
136
+ idw 0.083 0.167 0.233
137
+ gen_v3_s0 0.067 0.183 0.283
138
+ gen_v3_s1 0.083 0.150 0.267
139
+ gen_v3_s2 0.100 0.217 0.317
140
+ gen_v3e_s0 0.067 0.200 0.267
141
+ gen_v3e_s1 0.067 0.133 0.183
142
+ gen_v3e_s2 0.067 0.217 0.250
143
+ gen_v3e_t1_s0 0.050 0.183 0.250
144
+ K=1 gen_v3_s0 vs best retrieval (nn): model-only 2 retr-only 3 p=1 diff -0.017
145
+ K=1 gen_v3_s1 vs best retrieval (nn): model-only 4 retr-only 4 p=1 diff +0.000
146
+ K=1 gen_v3_s2 vs best retrieval (nn): model-only 3 retr-only 2 p=1 diff +0.017
147
+ K=1 gen_v3e_s0 vs best retrieval (nn): model-only 1 retr-only 2 p=1 diff -0.017
148
+ K=1 gen_v3e_s1 vs best retrieval (nn): model-only 2 retr-only 3 p=1 diff -0.017
149
+ K=1 gen_v3e_s2 vs best retrieval (nn): model-only 1 retr-only 2 p=1 diff -0.017
150
+ K=1 gen_v3e_t1_s0 vs best retrieval (nn): model-only 2 retr-only 4 p=0.688 diff -0.033
151
+ K=4 gen_v3_s0 vs best retrieval (nn): model-only 3 retr-only 5 p=0.727 diff -0.033
152
+ K=4 gen_v3_s1 vs best retrieval (nn): model-only 3 retr-only 7 p=0.344 diff -0.067
153
+ K=4 gen_v3_s2 vs best retrieval (nn): model-only 3 retr-only 3 p=1 diff +0.000
154
+ K=4 gen_v3e_s0 vs best retrieval (nn): model-only 4 retr-only 5 p=1 diff -0.017
155
+ K=4 gen_v3e_s1 vs best retrieval (nn): model-only 1 retr-only 6 p=0.125 diff -0.083
156
+ K=4 gen_v3e_s2 vs best retrieval (nn): model-only 4 retr-only 4 p=1 diff +0.000
157
+ K=4 gen_v3e_t1_s0 vs best retrieval (nn): model-only 4 retr-only 6 p=0.754 diff -0.033
158
+ K=8 gen_v3_s0 vs best retrieval (nn): model-only 6 retr-only 5 p=1 diff +0.017
159
+ K=8 gen_v3_s1 vs best retrieval (nn): model-only 5 retr-only 5 p=1 diff +0.000
160
+ K=8 gen_v3_s2 vs best retrieval (nn): model-only 4 retr-only 1 p=0.375 diff +0.050
161
+ K=8 gen_v3e_s0 vs best retrieval (nn): model-only 4 retr-only 4 p=1 diff +0.000
162
+ K=8 gen_v3e_s1 vs best retrieval (nn): model-only 2 retr-only 7 p=0.18 diff -0.083
163
+ K=8 gen_v3e_s2 vs best retrieval (nn): model-only 4 retr-only 5 p=1 diff -0.017
164
+ K=8 gen_v3e_t1_s0 vs best retrieval (nn): model-only 5 retr-only 6 p=1 diff -0.017
165
+
results/ood/pendulum_ood.txt ADDED
@@ -0,0 +1,804 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === lvl0: N=60 NN-dist ratio test/bank 1.0x (mean NN 0.148); oracle finite 60 ===
2
+ -- no steering: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
3
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
4
+ diff_s0 | 1.000 1.000 1.44 | 1.000 1.000 1.23 | 1.000 1.000 1.09
5
+ diff_s1 | 1.000 1.000 1.49 | 1.000 1.000 1.21 | 1.000 1.000 1.15
6
+ diff_s2 | 1.000 1.000 1.39 | 1.000 1.000 1.20 | 1.000 1.000 1.07
7
+ x_s42 | 1.000 1.000 1.51 | 1.000 1.000 1.28 | 1.000 1.000 1.08
8
+ x_s43 | 1.000 1.000 1.09 | 1.000 1.000 0.73 | 1.000 1.000 0.65
9
+ x_s44 | 1.000 1.000 1.35 | 1.000 1.000 1.17 | 1.000 1.000 1.02
10
+ nn | 1.000 1.000 1.78 | 1.000 1.000 0.68 | 1.000 1.000 0.29
11
+ idw | 1.000 1.000 1.78 | 1.000 1.000 0.71 | 1.000 1.000 0.29
12
+ nns | 1.000 1.000 1.40 | 1.000 1.000 1.21 | 1.000 1.000 1.11
13
+ dls | 1.000 1.000 1.41 | 1.000 1.000 1.23 | 1.000 1.000 1.14
14
+
15
+ -- 50-step steering on every candidate: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
16
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
17
+ diff_s0 | 1.000 0.983 0.02 | 1.000 1.000 0.00 | 1.000 1.000 0.00
18
+ diff_s1 | 1.000 0.983 0.01 | 1.000 1.000 0.00 | 1.000 1.000 0.00
19
+ diff_s2 | 1.000 0.983 0.01 | 1.000 1.000 0.00 | 1.000 1.000 0.00
20
+ x_s42 | 1.000 0.967 0.03 | 1.000 1.000 0.00 | 1.000 1.000 0.00
21
+ x_s43 | 1.000 0.983 0.01 | 1.000 1.000 0.00 | 1.000 1.000 -0.00
22
+ x_s44 | 1.000 0.983 0.02 | 1.000 1.000 0.00 | 1.000 1.000 0.00
23
+ nn | 1.000 1.000 0.01 | 1.000 1.000 -0.00 | 1.000 1.000 -0.00
24
+ idw | 1.000 1.000 0.01 | 1.000 1.000 -0.00 | 1.000 1.000 -0.00
25
+ nns | 1.000 1.000 0.01 | 1.000 1.000 0.00 | 1.000 1.000 0.00
26
+ dls | 1.000 1.000 0.01 | 1.000 1.000 0.00 | 1.000 1.000 0.00
27
+
28
+ -- paired tests, diffusion (model - retrieval): [cov McNemar model-only/retr-only p | certified-cov | paired regret diff median [95% CI], Wilcoxon p]; * = best retrieval arm by median regret
29
+ K=1 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.21 [-0.48,+0.09] p=0.112
30
+ K=1 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.21 [-0.48,+0.09] p=0.112
31
+ K=1 st=0 diff_s0 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.07 [-0.01,+0.22] p=0.11
32
+ K=1 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.14 [-0.27,+0.04] p=0.178
33
+ K=1 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.18 [-0.44,+0.14] p=0.236
34
+ K=1 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.18 [-0.44,+0.14] p=0.236
35
+ K=1 st=0 diff_s1 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.09 [-0.12,+0.25] p=0.342
36
+ K=1 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.09 [-0.21,+0.05] p=0.216
37
+ K=1 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.21 [-0.48,-0.01] p=0.153
38
+ K=1 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.21 [-0.48,-0.01] p=0.153
39
+ K=1 st=0 diff_s2 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.04 [-0.09,+0.14] p=0.422
40
+ K=1 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.13 [-0.29,-0.04] p=0.216
41
+ K=1 st=1 diff_s0 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.00 [-0.03,+0.00] p=0.484
42
+ K=1 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.00 [-0.03,+0.00] p=0.484
43
+ K=1 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 0-1 p=1 | regret +0.01 [+0.00,+0.02] p=4.26e-05
44
+ K=1 st=1 diff_s0 vs dls * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.00,+0.02] p=0.0393
45
+ K=1 st=1 diff_s1 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.00 [-0.02,+0.00] p=0.357
46
+ K=1 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.00 [-0.02,+0.00] p=0.357
47
+ K=1 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [+0.00,+0.01] p=0.00551
48
+ K=1 st=1 diff_s1 vs dls * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.00,+0.00] p=0.414
49
+ K=1 st=1 diff_s2 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.00 [-0.03,-0.00] p=0.328
50
+ K=1 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.00 [-0.03,-0.00] p=0.328
51
+ K=1 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [+0.00,+0.01] p=0.004
52
+ K=1 st=1 diff_s2 vs dls * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.00,+0.01] p=0.357
53
+ K=4 st=0 diff_s0 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.53 [+0.41,+0.84] p=2.3e-08
54
+ K=4 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.53 [+0.38,+0.84] p=5.35e-07
55
+ K=4 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.04 [-0.13,+0.16] p=0.702
56
+ K=4 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.11 [-0.26,-0.03] p=0.263
57
+ K=4 st=0 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.55 [+0.42,+0.87] p=3.1e-08
58
+ K=4 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.55 [+0.36,+0.87] p=6e-07
59
+ K=4 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [-0.15,+0.11] p=0.906
60
+ K=4 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.13 [-0.31,+0.01] p=0.276
61
+ K=4 st=0 diff_s2 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.59 [+0.41,+0.80] p=6.89e-09
62
+ K=4 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.57 [+0.39,+0.79] p=2.1e-07
63
+ K=4 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.02 [-0.12,+0.13] p=1
64
+ K=4 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.12 [-0.19,-0.06] p=0.198
65
+ K=4 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=5.48e-05
66
+ K=4 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.02] p=0.000118
67
+ K=4 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.0105
68
+ K=4 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.178
69
+ K=4 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [+0.00,+0.03] p=5.83e-05
70
+ K=4 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [+0.00,+0.03] p=0.000125
71
+ K=4 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.0158
72
+ K=4 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.0625
73
+ K=4 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [+0.00,+0.03] p=3.4e-05
74
+ K=4 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [+0.00,+0.02] p=1.77e-05
75
+ K=4 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [+0.00,+0.01] p=0.000747
76
+ K=4 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.12
77
+ K=8 st=0 diff_s0 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.68 [+0.46,+0.84] p=2.21e-10
78
+ K=8 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.68 [+0.45,+0.84] p=2.32e-10
79
+ K=8 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.02 [-0.09,+0.11] p=0.648
80
+ K=8 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.18 [-0.28,-0.11] p=0.385
81
+ K=8 st=0 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.76 [+0.56,+0.96] p=1.74e-10
82
+ K=8 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.76 [+0.56,+0.96] p=1.74e-10
83
+ K=8 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.03 [-0.06,+0.19] p=0.151
84
+ K=8 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.12 [-0.23,-0.03] p=0.675
85
+ K=8 st=0 diff_s2 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.68 [+0.57,+0.80] p=1.58e-10
86
+ K=8 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.68 [+0.57,+0.80] p=1.74e-10
87
+ K=8 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.08,+0.07] p=1
88
+ K=8 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.21 [-0.28,-0.03] p=0.16
89
+ K=8 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=9.45e-07
90
+ K=8 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=1.36e-07
91
+ K=8 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.00806
92
+ K=8 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.0636
93
+ K=8 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=2.46e-06
94
+ K=8 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=3.12e-07
95
+ K=8 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.00721
96
+ K=8 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.0515
97
+ K=8 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=2.46e-07
98
+ K=8 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=2.23e-08
99
+ K=8 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.0477
100
+ K=8 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.0812
101
+
102
+ === lvl1.1: N=60 NN-dist ratio test/bank 1.9x (mean NN 0.268); oracle finite 60 ===
103
+ -- no steering: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
104
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
105
+ diff_s0 | 1.000 1.000 1.29 | 1.000 1.000 1.00 | 1.000 1.000 0.92
106
+ diff_s1 | 1.000 1.000 1.28 | 1.000 1.000 1.03 | 1.000 1.000 0.92
107
+ diff_s2 | 1.000 1.000 1.14 | 1.000 1.000 0.94 | 1.000 1.000 0.84
108
+ x_s42 | 1.000 1.000 1.20 | 1.000 1.000 0.96 | 1.000 1.000 0.88
109
+ x_s43 | 1.000 0.983 1.15 | 1.000 1.000 0.85 | 1.000 1.000 0.74
110
+ x_s44 | 1.000 1.000 1.17 | 1.000 1.000 0.93 | 1.000 1.000 0.88
111
+ nn | 1.000 1.000 1.41 | 1.000 1.000 0.68 | 1.000 1.000 0.53
112
+ idw | 1.000 1.000 1.41 | 1.000 1.000 0.75 | 1.000 1.000 0.55
113
+ nns | 1.000 1.000 1.27 | 1.000 1.000 1.06 | 1.000 1.000 0.95
114
+ dls | 1.000 1.000 1.17 | 1.000 1.000 1.05 | 1.000 1.000 0.95
115
+
116
+ -- 50-step steering on every candidate: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
117
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
118
+ diff_s0 | 1.000 0.917 0.03 | 1.000 0.967 0.00 | 1.000 0.967 0.00
119
+ diff_s1 | 1.000 0.933 0.02 | 1.000 0.983 0.00 | 1.000 1.000 0.00
120
+ diff_s2 | 1.000 0.917 0.02 | 1.000 0.967 0.00 | 1.000 0.967 0.00
121
+ x_s42 | 1.000 0.900 0.03 | 1.000 0.917 0.00 | 1.000 0.933 0.00
122
+ x_s43 | 1.000 0.867 0.04 | 1.000 0.933 0.01 | 1.000 0.933 0.00
123
+ x_s44 | 1.000 0.917 0.02 | 1.000 0.950 0.00 | 1.000 0.950 0.00
124
+ nn | 1.000 0.967 0.01 | 1.000 1.000 -0.01 | 1.000 1.000 -0.01
125
+ idw | 1.000 0.967 0.01 | 1.000 1.000 -0.00 | 1.000 1.000 -0.01
126
+ nns | 1.000 0.950 0.01 | 1.000 0.950 0.01 | 1.000 0.950 0.00
127
+ dls | 1.000 0.950 0.01 | 1.000 0.950 0.00 | 1.000 0.950 0.00
128
+
129
+ -- paired tests, diffusion (model - retrieval): [cov McNemar model-only/retr-only p | certified-cov | paired regret diff median [95% CI], Wilcoxon p]; * = best retrieval arm by median regret
130
+ K=1 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret +0.04 [-0.33,+0.69] p=0.617
131
+ K=1 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.04 [-0.33,+0.69] p=0.617
132
+ K=1 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.17 [-0.03,+0.29] p=0.046
133
+ K=1 st=0 diff_s0 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.04 [-0.13,+0.22] p=0.245
134
+ K=1 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.76,+0.47] p=0.814
135
+ K=1 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.76,+0.47] p=0.814
136
+ K=1 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.05 [-0.08,+0.21] p=0.373
137
+ K=1 st=0 diff_s1 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.05 [-0.22,+0.15] p=0.802
138
+ K=1 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.07 [-0.55,+0.48] p=0.883
139
+ K=1 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.07 [-0.55,+0.48] p=0.883
140
+ K=1 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.22,+0.17] p=0.48
141
+ K=1 st=0 diff_s2 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.04 [-0.19,+0.17] p=0.912
142
+ K=1 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-3 p=0.25 | regret +0.00 [-0.00,+0.05] p=0.0126
143
+ K=1 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-3 p=0.25 | regret +0.00 [-0.00,+0.05] p=0.0126
144
+ K=1 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.00 [-0.00,+0.03] p=0.00373
145
+ K=1 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.00 [-0.00,+0.02] p=0.00705
146
+ K=1 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-2 p=0.5 | regret -0.00 [-0.01,+0.01] p=0.401
147
+ K=1 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret -0.00 [-0.01,+0.01] p=0.401
148
+ K=1 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-1 p=1 | regret -0.00 [-0.01,+0.00] p=0.982
149
+ K=1 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 0-1 p=1 | regret -0.00 [-0.00,+0.00] p=0.571
150
+ K=1 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 1-4 p=0.375 | regret +0.00 [-0.00,+0.02] p=0.052
151
+ K=1 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 1-4 p=0.375 | regret +0.00 [-0.00,+0.02] p=0.052
152
+ K=1 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 1-3 p=0.625 | regret +0.00 [-0.00,+0.03] p=0.00737
153
+ K=1 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 1-3 p=0.625 | regret +0.00 [-0.00,+0.01] p=0.105
154
+ K=4 st=0 diff_s0 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.31 [+0.00,+0.67] p=0.0084
155
+ K=4 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.20 [-0.06,+0.58] p=0.052
156
+ K=4 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.05 [-0.14,+0.05] p=0.906
157
+ K=4 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.14 [-0.24,-0.03] p=0.178
158
+ K=4 st=0 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.28 [-0.03,+0.65] p=0.0252
159
+ K=4 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.17 [-0.11,+0.42] p=0.113
160
+ K=4 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.08 [-0.17,-0.01] p=0.306
161
+ K=4 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.17 [-0.29,-0.08] p=0.0668
162
+ K=4 st=0 diff_s2 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.34 [+0.04,+0.54] p=0.0192
163
+ K=4 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.24 [-0.03,+0.53] p=0.0947
164
+ K=4 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.05 [-0.23,+0.02] p=0.393
165
+ K=4 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.13 [-0.22,-0.04] p=0.0595
166
+ K=4 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.02 [+0.00,+0.05] p=2.16e-05
167
+ K=4 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.04] p=6.02e-05
168
+ K=4 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 1-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.424
169
+ K=4 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 1-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.686
170
+ K=4 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.01 [+0.00,+0.03] p=0.000141
171
+ K=4 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret +0.01 [+0.00,+0.03] p=0.000268
172
+ K=4 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 2-0 p=0.5 | regret -0.00 [-0.01,+0.00] p=0.478
173
+ K=4 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 2-0 p=0.5 | regret -0.00 [-0.00,+0.00] p=0.695
174
+ K=4 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.04] p=1.11e-05
175
+ K=4 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.04] p=8.46e-05
176
+ K=4 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 1-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.422
177
+ K=4 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 1-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.466
178
+ K=8 st=0 diff_s0 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.28 [+0.02,+0.68] p=0.00419
179
+ K=8 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.26 [-0.07,+0.56] p=0.0111
180
+ K=8 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.06 [-0.15,+0.07] p=0.825
181
+ K=8 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.18 [-0.29,-0.03] p=0.296
182
+ K=8 st=0 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.21 [+0.02,+0.69] p=0.00492
183
+ K=8 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.19 [+0.01,+0.62] p=0.0113
184
+ K=8 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.02 [-0.20,+0.12] p=0.883
185
+ K=8 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.18 [-0.29,-0.05] p=0.339
186
+ K=8 st=0 diff_s2 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.31 [+0.01,+0.58] p=0.00753
187
+ K=8 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.23 [+0.01,+0.56] p=0.0154
188
+ K=8 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.06 [-0.17,+0.01] p=0.393
189
+ K=8 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.15 [-0.30,-0.08] p=0.18
190
+ K=8 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.05] p=2.09e-05
191
+ K=8 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.04] p=6.61e-05
192
+ K=8 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 1-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.397
193
+ K=8 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 1-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.437
194
+ K=8 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=3.16e-06
195
+ K=8 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=4.84e-06
196
+ K=8 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 3-0 p=0.25 | regret -0.00 [-0.00,+0.00] p=0.551
197
+ K=8 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 3-0 p=0.25 | regret -0.00 [-0.00,+0.00] p=0.478
198
+ K=8 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.05] p=6.55e-07
199
+ K=8 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.04] p=4.43e-07
200
+ K=8 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 1-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.592
201
+ K=8 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 1-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.299
202
+
203
+ === lvl1.25: N=60 NN-dist ratio test/bank 3.9x (mean NN 0.565); oracle finite 60 ===
204
+ -- no steering: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
205
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
206
+ diff_s0 | 1.000 0.983 1.21 | 1.000 1.000 0.83 | 1.000 1.000 0.70
207
+ diff_s1 | 1.000 0.983 1.11 | 1.000 1.000 0.81 | 1.000 1.000 0.67
208
+ diff_s2 | 1.000 1.000 1.26 | 1.000 1.000 0.81 | 1.000 1.000 0.72
209
+ x_s42 | 1.000 0.983 1.28 | 1.000 1.000 0.85 | 1.000 1.000 0.69
210
+ x_s43 | 1.000 1.000 1.37 | 1.000 1.000 0.88 | 1.000 1.000 0.65
211
+ x_s44 | 1.000 0.983 1.19 | 1.000 1.000 0.79 | 1.000 1.000 0.71
212
+ nn | 1.000 1.000 1.80 | 1.000 1.000 1.22 | 1.000 1.000 1.06
213
+ idw | 1.000 1.000 1.80 | 1.000 1.000 1.27 | 1.000 1.000 1.00
214
+ nns | 1.000 1.000 1.11 | 1.000 1.000 0.97 | 1.000 1.000 0.95
215
+ dls | 1.000 1.000 1.06 | 1.000 1.000 1.00 | 1.000 1.000 0.97
216
+
217
+ -- 50-step steering on every candidate: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
218
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
219
+ diff_s0 | 1.000 0.817 0.01 | 1.000 0.833 0.00 | 1.000 0.883 -0.00
220
+ diff_s1 | 1.000 0.783 0.01 | 1.000 0.917 0.00 | 1.000 0.950 -0.00
221
+ diff_s2 | 1.000 0.767 0.01 | 1.000 0.883 -0.00 | 1.000 0.967 -0.00
222
+ x_s42 | 1.000 0.750 0.01 | 1.000 0.817 -0.00 | 1.000 0.833 -0.01
223
+ x_s43 | 1.000 0.800 0.03 | 1.000 0.833 0.00 | 1.000 0.917 0.00
224
+ x_s44 | 1.000 0.817 0.02 | 1.000 0.867 0.00 | 1.000 0.933 0.00
225
+ nn | 1.000 0.950 -0.00 | 1.000 1.000 -0.01 | 1.000 1.000 -0.01
226
+ idw | 1.000 0.950 -0.00 | 1.000 1.000 -0.00 | 1.000 1.000 -0.01
227
+ nns | 1.000 0.850 0.01 | 1.000 0.883 0.00 | 1.000 0.883 0.00
228
+ dls | 1.000 0.833 0.01 | 1.000 0.867 0.00 | 1.000 0.883 0.00
229
+
230
+ -- paired tests, diffusion (model - retrieval): [cov McNemar model-only/retr-only p | certified-cov | paired regret diff median [95% CI], Wilcoxon p]; * = best retrieval arm by median regret
231
+ K=1 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.86 [-1.73,+0.08] p=0.178
232
+ K=1 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.86 [-1.73,+0.08] p=0.178
233
+ K=1 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-1 p=1 | regret +0.12 [-0.14,+0.36] p=0.107
234
+ K=1 st=0 diff_s0 vs dls * cov 0-0 p=1 | cert 0-1 p=1 | regret -0.03 [-0.25,+0.16] p=0.837
235
+ K=1 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.65 [-1.80,+0.29] p=0.248
236
+ K=1 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.65 [-1.80,+0.29] p=0.248
237
+ K=1 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-1 p=1 | regret -0.02 [-0.09,+0.29] p=0.242
238
+ K=1 st=0 diff_s1 vs dls * cov 0-0 p=1 | cert 0-1 p=1 | regret -0.04 [-0.30,+0.23] p=0.848
239
+ K=1 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.55 [-1.94,+0.30] p=0.279
240
+ K=1 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.55 [-1.94,+0.30] p=0.279
241
+ K=1 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.10 [-0.14,+0.38] p=0.0933
242
+ K=1 st=0 diff_s2 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.10 [-0.14,+0.31] p=0.494
243
+ K=1 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-8 p=0.00781 | regret +0.01 [+0.00,+0.08] p=0.0069
244
+ K=1 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-8 p=0.00781 | regret +0.01 [+0.00,+0.08] p=0.0069
245
+ K=1 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 1-3 p=0.625 | regret +0.00 [-0.00,+0.01] p=0.381
246
+ K=1 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 1-2 p=1 | regret -0.00 [-0.02,+0.00] p=0.248
247
+ K=1 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-10 p=0.00195 | regret +0.01 [-0.01,+0.05] p=0.085
248
+ K=1 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-10 p=0.00195 | regret +0.01 [-0.01,+0.05] p=0.085
249
+ K=1 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 2-6 p=0.289 | regret +0.00 [-0.00,+0.00] p=0.527
250
+ K=1 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 2-5 p=0.453 | regret -0.00 [-0.02,-0.00] p=0.155
251
+ K=1 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-11 p=0.000977 | regret +0.01 [-0.00,+0.08] p=0.0113
252
+ K=1 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-11 p=0.000977 | regret +0.01 [-0.00,+0.08] p=0.0113
253
+ K=1 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 0-5 p=0.0625 | regret +0.00 [-0.00,+0.02] p=0.0836
254
+ K=1 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 1-5 p=0.219 | regret +0.00 [-0.00,+0.02] p=0.489
255
+ K=4 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.24 [-0.81,+0.24] p=0.466
256
+ K=4 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.27 [-0.98,+0.13] p=0.23
257
+ K=4 st=0 diff_s0 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.09 [-0.16,+0.06] p=0.269
258
+ K=4 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.18 [-0.27,-0.03] p=0.0469
259
+ K=4 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.27 [-0.71,+0.35] p=0.729
260
+ K=4 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.36 [-1.14,+0.38] p=0.354
261
+ K=4 st=0 diff_s1 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.09 [-0.19,-0.05] p=0.0346
262
+ K=4 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.18 [-0.27,-0.05] p=0.0174
263
+ K=4 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.30 [-0.68,+0.12] p=0.556
264
+ K=4 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.37 [-0.96,+0.07] p=0.282
265
+ K=4 st=0 diff_s2 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.13 [-0.19,-0.02] p=0.0234
266
+ K=4 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.18 [-0.26,-0.09] p=0.0252
267
+ K=4 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-10 p=0.00195 | regret +0.00 [+0.00,+0.07] p=0.000446
268
+ K=4 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-10 p=0.00195 | regret +0.00 [-0.00,+0.07] p=0.00355
269
+ K=4 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 1-4 p=0.375 | regret -0.00 [-0.01,+0.00] p=0.124
270
+ K=4 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 1-3 p=0.625 | regret -0.00 [-0.02,-0.00] p=0.0566
271
+ K=4 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-5 p=0.0625 | regret +0.01 [+0.00,+0.03] p=0.00148
272
+ K=4 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-5 p=0.0625 | regret +0.00 [-0.00,+0.03] p=0.0077
273
+ K=4 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 3-1 p=0.625 | regret -0.00 [-0.02,-0.00] p=0.0174
274
+ K=4 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 4-1 p=0.375 | regret -0.00 [-0.02,-0.00] p=0.052
275
+ K=4 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-7 p=0.0156 | regret +0.01 [+0.00,+0.05] p=0.000571
276
+ K=4 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-7 p=0.0156 | regret +0.01 [+0.00,+0.05] p=0.00152
277
+ K=4 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 2-2 p=1 | regret -0.00 [-0.01,-0.00] p=0.0309
278
+ K=4 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 3-2 p=1 | regret -0.01 [-0.02,-0.00] p=0.0511
279
+ K=8 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.28 [-0.66,+0.24] p=0.763
280
+ K=8 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.19 [-0.64,+0.27] p=0.752
281
+ K=8 st=0 diff_s0 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.17 [-0.24,-0.10] p=0.0026
282
+ K=8 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.24 [-0.30,-0.13] p=0.0104
283
+ K=8 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.21 [-0.53,+0.22] p=0.848
284
+ K=8 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.14 [-0.56,+0.22] p=0.854
285
+ K=8 st=0 diff_s1 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.17 [-0.21,-0.07] p=0.00117
286
+ K=8 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.22 [-0.30,-0.15] p=0.0262
287
+ K=8 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.21 [-0.54,+0.16] p=0.814
288
+ K=8 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.19 [-0.46,+0.16] p=0.735
289
+ K=8 st=0 diff_s2 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.12 [-0.19,-0.03] p=0.00111
290
+ K=8 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.20 [-0.26,-0.09] p=0.0657
291
+ K=8 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-7 p=0.0156 | regret +0.01 [+0.00,+0.06] p=0.000159
292
+ K=8 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-7 p=0.0156 | regret +0.01 [-0.00,+0.06] p=0.000356
293
+ K=8 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 1-1 p=1 | regret -0.00 [-0.01,-0.00] p=0.0059
294
+ K=8 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 2-2 p=1 | regret -0.00 [-0.01,-0.00] p=0.245
295
+ K=8 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-3 p=0.25 | regret +0.00 [+0.00,+0.03] p=0.000516
296
+ K=8 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-3 p=0.25 | regret +0.00 [+0.00,+0.03] p=0.0014
297
+ K=8 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 5-1 p=0.219 | regret -0.00 [-0.04,-0.00] p=0.00515
298
+ K=8 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 4-0 p=0.125 | regret -0.00 [-0.02,-0.00] p=0.143
299
+ K=8 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.03] p=7.48e-05
300
+ K=8 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.03] p=0.000118
301
+ K=8 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 5-0 p=0.0625 | regret -0.01 [-0.03,-0.00] p=0.000708
302
+ K=8 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 5-0 p=0.0625 | regret -0.00 [-0.01,-0.00] p=0.19
303
+
304
+ === lvl1.5: N=60 NN-dist ratio test/bank 7.5x (mean NN 1.075); oracle finite 60 ===
305
+ -- no steering: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
306
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
307
+ diff_s0 | 1.000 0.967 1.70 | 1.000 0.983 1.07 | 1.000 0.983 0.83
308
+ diff_s1 | 1.000 0.967 1.58 | 1.000 0.983 1.02 | 1.000 0.983 0.92
309
+ diff_s2 | 1.000 0.983 1.48 | 1.000 0.983 0.98 | 1.000 0.983 0.86
310
+ x_s42 | 1.000 0.950 1.70 | 1.000 0.983 1.04 | 1.000 0.983 0.97
311
+ x_s43 | 1.000 0.900 2.13 | 1.000 0.967 1.02 | 1.000 0.983 0.96
312
+ x_s44 | 1.000 0.967 2.04 | 1.000 0.983 1.05 | 1.000 0.983 0.99
313
+ nn | 1.000 1.000 3.77 | 1.000 1.000 1.53 | 1.000 1.000 1.08
314
+ idw | 1.000 1.000 3.77 | 1.000 1.000 2.38 | 1.000 1.000 1.16
315
+ nns | 1.000 1.000 1.19 | 1.000 1.000 1.07 | 1.000 1.000 0.98
316
+ dls | 1.000 1.000 1.37 | 1.000 1.000 1.11 | 1.000 1.000 1.05
317
+
318
+ -- 50-step steering on every candidate: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
319
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
320
+ diff_s0 | 1.000 0.817 0.05 | 1.000 0.900 0.01 | 1.000 0.917 0.00
321
+ diff_s1 | 1.000 0.783 0.03 | 1.000 0.850 0.00 | 1.000 0.917 0.00
322
+ diff_s2 | 1.000 0.850 0.02 | 1.000 0.883 0.00 | 1.000 0.933 -0.00
323
+ x_s42 | 1.000 0.783 0.05 | 1.000 0.883 0.00 | 1.000 0.900 0.00
324
+ x_s43 | 1.000 0.767 0.07 | 1.000 0.883 0.01 | 1.000 0.917 0.00
325
+ x_s44 | 1.000 0.800 0.09 | 1.000 0.917 0.01 | 1.000 0.917 0.00
326
+ nn | 1.000 0.950 0.02 | 1.000 1.000 -0.01 | 1.000 1.000 -0.02
327
+ idw | 1.000 0.950 0.02 | 1.000 1.000 -0.01 | 1.000 1.000 -0.01
328
+ nns | 1.000 0.833 0.04 | 1.000 0.850 0.03 | 1.000 0.850 0.03
329
+ dls | 1.000 0.833 0.05 | 1.000 0.833 0.02 | 1.000 0.833 0.01
330
+
331
+ -- paired tests, diffusion (model - retrieval): [cov McNemar model-only/retr-only p | certified-cov | paired regret diff median [95% CI], Wilcoxon p]; * = best retrieval arm by median regret
332
+ K=1 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-2 p=0.5 | regret -1.50 [-3.13,+0.45] p=0.617
333
+ K=1 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret -1.50 [-3.13,+0.45] p=0.617
334
+ K=1 st=0 diff_s0 vs nns * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.57 [+0.26,+1.44] p=3.65e-06
335
+ K=1 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.45 [+0.09,+0.93] p=0.000602
336
+ K=1 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-2 p=0.5 | regret -2.30 [-3.30,-0.05] p=0.317
337
+ K=1 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret -2.30 [-3.30,-0.05] p=0.317
338
+ K=1 st=0 diff_s1 vs nns * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.50 [+0.19,+1.18] p=0.000195
339
+ K=1 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.22 [-0.06,+0.51] p=0.0668
340
+ K=1 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -1.89 [-3.17,-0.26] p=0.248
341
+ K=1 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -1.89 [-3.17,-0.26] p=0.248
342
+ K=1 st=0 diff_s2 vs nns * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.14 [-0.01,+0.97] p=0.00515
343
+ K=1 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-1 p=1 | regret +0.05 [-0.13,+0.46] p=0.0962
344
+ K=1 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-8 p=0.00781 | regret +0.01 [-0.01,+0.12] p=0.129
345
+ K=1 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-8 p=0.00781 | regret +0.01 [-0.01,+0.12] p=0.129
346
+ K=1 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 1-2 p=1 | regret +0.00 [-0.00,+0.08] p=0.0109
347
+ K=1 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 1-2 p=1 | regret +0.01 [-0.00,+0.11] p=0.0142
348
+ K=1 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-10 p=0.00195 | regret +0.01 [-0.03,+0.07] p=0.317
349
+ K=1 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-10 p=0.00195 | regret +0.01 [-0.03,+0.07] p=0.317
350
+ K=1 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-3 p=0.25 | regret +0.00 [-0.01,+0.04] p=0.171
351
+ K=1 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 0-3 p=0.25 | regret -0.00 [-0.02,+0.01] p=0.988
352
+ K=1 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 1-7 p=0.0703 | regret +0.01 [-0.03,+0.03] p=0.414
353
+ K=1 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 1-7 p=0.0703 | regret +0.01 [-0.03,+0.03] p=0.414
354
+ K=1 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 2-1 p=1 | regret -0.00 [-0.01,+0.01] p=0.44
355
+ K=1 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 2-1 p=1 | regret -0.00 [-0.04,+0.00] p=0.935
356
+ K=4 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.19 [-0.91,+0.36] p=0.977
357
+ K=4 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.13 [-1.40,+0.18] p=0.752
358
+ K=4 st=0 diff_s0 vs nns * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.14 [-0.02,+0.52] p=0.0151
359
+ K=4 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-1 p=1 | regret +0.02 [-0.27,+0.24] p=0.317
360
+ K=4 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.33 [-1.03,+0.18] p=0.527
361
+ K=4 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.19 [-1.49,+0.08] p=0.414
362
+ K=4 st=0 diff_s1 vs nns * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.16 [-0.08,+0.42] p=0.0452
363
+ K=4 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-1 p=1 | regret +0.06 [-0.28,+0.20] p=0.686
364
+ K=4 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.29 [-0.91,+0.19] p=0.576
365
+ K=4 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.30 [-1.47,+0.19] p=0.457
366
+ K=4 st=0 diff_s2 vs nns * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.03 [-0.19,+0.16] p=0.44
367
+ K=4 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-1 p=1 | regret -0.13 [-0.36,+0.04] p=0.263
368
+ K=4 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-6 p=0.0312 | regret +0.01 [-0.00,+0.09] p=0.0106
369
+ K=4 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-6 p=0.0312 | regret +0.01 [-0.00,+0.10] p=0.0196
370
+ K=4 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 4-1 p=0.375 | regret +0.00 [-0.01,+0.01] p=0.431
371
+ K=4 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 5-1 p=0.219 | regret -0.00 [-0.03,+0.00] p=0.959
372
+ K=4 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-9 p=0.00391 | regret +0.01 [-0.00,+0.05] p=0.0353
373
+ K=4 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-9 p=0.00391 | regret +0.01 [-0.00,+0.05] p=0.0529
374
+ K=4 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 2-2 p=1 | regret +0.00 [-0.02,+0.01] p=0.527
375
+ K=4 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 2-1 p=1 | regret -0.01 [-0.04,+0.00] p=0.982
376
+ K=4 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-7 p=0.0156 | regret +0.01 [+0.00,+0.06] p=0.0077
377
+ K=4 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-7 p=0.0156 | regret +0.01 [+0.00,+0.06] p=0.0167
378
+ K=4 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 3-1 p=0.625 | regret -0.00 [-0.02,+0.00] p=0.69
379
+ K=4 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 4-1 p=0.375 | regret -0.01 [-0.05,-0.00] p=0.178
380
+ K=8 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.09 [-0.45,+0.15] p=0.941
381
+ K=8 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.09 [-0.51,+0.15] p=0.837
382
+ K=8 st=0 diff_s0 vs nns * cov 0-0 p=1 | cert 0-1 p=1 | regret -0.03 [-0.18,+0.10] p=0.475
383
+ K=8 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-1 p=1 | regret -0.15 [-0.35,+0.11] p=0.686
384
+ K=8 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.10 [-0.42,+0.30] p=0.906
385
+ K=8 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.10 [-0.52,+0.30] p=0.774
386
+ K=8 st=0 diff_s1 vs nns * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.08 [-0.11,+0.37] p=0.076
387
+ K=8 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.27,+0.23] p=0.422
388
+ K=8 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret -0.05 [-0.28,+0.21] p=0.877
389
+ K=8 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret -0.05 [-0.50,+0.21] p=0.718
390
+ K=8 st=0 diff_s2 vs nns * cov 0-0 p=1 | cert 0-1 p=1 | regret -0.03 [-0.14,+0.09] p=0.691
391
+ K=8 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-1 p=1 | regret -0.17 [-0.31,+0.03] p=0.643
392
+ K=8 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-5 p=0.0625 | regret +0.01 [-0.00,+0.05] p=0.00936
393
+ K=8 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-5 p=0.0625 | regret +0.01 [-0.00,+0.05] p=0.0128
394
+ K=8 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 5-1 p=0.219 | regret -0.00 [-0.02,+0.00] p=0.982
395
+ K=8 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 6-1 p=0.125 | regret -0.00 [-0.02,+0.00] p=0.752
396
+ K=8 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-5 p=0.0625 | regret +0.00 [-0.00,+0.02] p=0.0293
397
+ K=8 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-5 p=0.0625 | regret +0.00 [-0.00,+0.02] p=0.052
398
+ K=8 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 5-1 p=0.219 | regret -0.00 [-0.02,+0.00] p=0.68
399
+ K=8 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 6-1 p=0.125 | regret -0.01 [-0.04,+0.00] p=0.994
400
+ K=8 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-4 p=0.125 | regret +0.00 [-0.00,+0.06] p=0.0154
401
+ K=8 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-4 p=0.125 | regret +0.00 [-0.00,+0.06] p=0.0221
402
+ K=8 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 6-1 p=0.125 | regret -0.00 [-0.02,+0.00] p=0.774
403
+ K=8 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 7-1 p=0.0703 | regret -0.02 [-0.04,-0.00] p=0.289
404
+
405
+ === lvl2.0: N=60 NN-dist ratio test/bank 23.5x (mean NN 3.376); oracle finite 60 ===
406
+ -- no steering: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
407
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
408
+ diff_s0 | 1.000 0.867 3.21 | 1.000 0.933 0.88 | 1.000 0.933 0.73
409
+ diff_s1 | 1.000 0.817 1.92 | 1.000 0.950 0.73 | 1.000 0.950 0.66
410
+ diff_s2 | 1.000 0.800 2.10 | 1.000 0.917 0.81 | 1.000 0.950 0.65
411
+ x_s42 | 0.967 0.767 4.09 | 1.000 0.900 0.85 | 1.000 0.933 0.62
412
+ x_s43 | 0.983 0.817 4.11 | 1.000 0.867 1.49 | 1.000 0.917 0.89
413
+ x_s44 | 0.983 0.817 4.48 | 1.000 0.917 1.13 | 1.000 0.933 0.78
414
+ nn | 0.967 0.950 8.97 | 0.967 0.950 5.67 | 0.967 0.950 4.00
415
+ idw | 0.967 0.950 8.97 | 0.950 0.933 5.92 | 0.950 0.933 4.31
416
+ nns | 1.000 1.000 1.01 | 1.000 1.000 0.87 | 1.000 1.000 0.78
417
+ dls | 1.000 1.000 1.02 | 1.000 1.000 0.85 | 1.000 1.000 0.81
418
+
419
+ -- 50-step steering on every candidate: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
420
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
421
+ diff_s0 | 1.000 0.767 0.02 | 1.000 0.850 -0.00 | 1.000 0.883 -0.00
422
+ diff_s1 | 1.000 0.700 0.00 | 1.000 0.900 -0.00 | 1.000 0.917 -0.01
423
+ diff_s2 | 1.000 0.650 0.02 | 1.000 0.817 -0.00 | 1.000 0.867 -0.00
424
+ x_s42 | 1.000 0.700 0.03 | 1.000 0.817 -0.00 | 1.000 0.900 -0.01
425
+ x_s43 | 1.000 0.783 0.31 | 1.000 0.850 0.00 | 1.000 0.917 -0.00
426
+ x_s44 | 1.000 0.733 0.26 | 1.000 0.817 0.00 | 1.000 0.883 -0.00
427
+ nn | 0.967 0.900 0.15 | 0.983 0.950 0.02 | 0.983 0.950 0.00
428
+ idw | 0.967 0.900 0.15 | 0.983 0.950 0.03 | 0.983 0.950 0.00
429
+ nns | 1.000 0.750 0.02 | 1.000 0.750 0.01 | 1.000 0.750 0.00
430
+ dls | 1.000 0.700 0.01 | 1.000 0.700 0.00 | 1.000 0.733 0.00
431
+
432
+ -- paired tests, diffusion (model - retrieval): [cov McNemar model-only/retr-only p | certified-cov | paired regret diff median [95% CI], Wilcoxon p]; * = best retrieval arm by median regret
433
+ K=1 st=0 diff_s0 vs nn cov 2-0 p=0.5 | cert 0-5 p=0.0625 | regret -2.61 [-8.11,+0.41] p=0.135
434
+ K=1 st=0 diff_s0 vs idw cov 2-0 p=0.5 | cert 0-5 p=0.0625 | regret -2.61 [-8.11,+0.41] p=0.135
435
+ K=1 st=0 diff_s0 vs nns * cov 0-0 p=1 | cert 0-8 p=0.00781 | regret +2.56 [+0.98,+3.75] p=9.44e-08
436
+ K=1 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-8 p=0.00781 | regret +1.68 [+0.94,+3.37] p=6.23e-07
437
+ K=1 st=0 diff_s1 vs nn cov 2-0 p=0.5 | cert 0-8 p=0.00781 | regret -4.63 [-9.49,+0.51] p=0.0485
438
+ K=1 st=0 diff_s1 vs idw cov 2-0 p=0.5 | cert 0-8 p=0.00781 | regret -4.63 [-9.49,+0.51] p=0.0485
439
+ K=1 st=0 diff_s1 vs nns * cov 0-0 p=1 | cert 0-11 p=0.000977 | regret +0.87 [+0.30,+2.79] p=0.000498
440
+ K=1 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-11 p=0.000977 | regret +0.90 [+0.30,+2.94] p=0.000949
441
+ K=1 st=0 diff_s2 vs nn cov 2-0 p=0.5 | cert 0-9 p=0.00391 | regret -3.68 [-9.38,+0.47] p=0.12
442
+ K=1 st=0 diff_s2 vs idw cov 2-0 p=0.5 | cert 0-9 p=0.00391 | regret -3.68 [-9.38,+0.47] p=0.12
443
+ K=1 st=0 diff_s2 vs nns * cov 0-0 p=1 | cert 0-12 p=0.000488 | regret +1.33 [+0.26,+2.69] p=4.68e-05
444
+ K=1 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-12 p=0.000488 | regret +1.28 [-0.02,+2.69] p=0.000268
445
+ K=1 st=1 diff_s0 vs nn cov 2-0 p=0.5 | cert 3-11 p=0.0574 | regret -0.01 [-0.16,+0.07] p=0.757
446
+ K=1 st=1 diff_s0 vs idw cov 2-0 p=0.5 | cert 3-11 p=0.0574 | regret -0.01 [-0.16,+0.07] p=0.757
447
+ K=1 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 8-7 p=1 | regret +0.01 [+0.00,+0.19] p=0.00539
448
+ K=1 st=1 diff_s0 vs dls * cov 0-0 p=1 | cert 11-7 p=0.481 | regret +0.01 [-0.00,+0.22] p=0.0109
449
+ K=1 st=1 diff_s1 vs nn cov 2-0 p=0.5 | cert 3-15 p=0.00754 | regret -0.03 [-0.54,+0.00] p=0.233
450
+ K=1 st=1 diff_s1 vs idw cov 2-0 p=0.5 | cert 3-15 p=0.00754 | regret -0.03 [-0.54,+0.00] p=0.233
451
+ K=1 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 5-8 p=0.581 | regret -0.00 [-0.03,+0.01] p=0.659
452
+ K=1 st=1 diff_s1 vs dls * cov 0-0 p=1 | cert 8-8 p=1 | regret -0.01 [-0.03,+0.01] p=0.669
453
+ K=1 st=1 diff_s2 vs nn cov 2-0 p=0.5 | cert 3-18 p=0.00149 | regret -0.01 [-0.36,+0.02] p=0.527
454
+ K=1 st=1 diff_s2 vs idw cov 2-0 p=0.5 | cert 3-18 p=0.00149 | regret -0.01 [-0.36,+0.02] p=0.527
455
+ K=1 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 3-9 p=0.146 | regret +0.01 [-0.00,+0.14] p=0.0204
456
+ K=1 st=1 diff_s2 vs dls * cov 0-0 p=1 | cert 6-9 p=0.607 | regret +0.00 [-0.01,+0.13] p=0.0605
457
+ K=4 st=0 diff_s0 vs nn cov 2-0 p=0.5 | cert 0-1 p=1 | regret -1.85 [-7.30,-0.06] p=0.00753
458
+ K=4 st=0 diff_s0 vs idw cov 3-0 p=0.25 | cert 1-1 p=1 | regret -3.53 [-8.30,+0.01] p=0.00409
459
+ K=4 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-4 p=0.125 | regret +0.06 [-0.13,+0.46] p=0.155
460
+ K=4 st=0 diff_s0 vs dls * cov 0-0 p=1 | cert 0-4 p=0.125 | regret -0.09 [-0.31,+0.46] p=0.484
461
+ K=4 st=0 diff_s1 vs nn cov 2-0 p=0.5 | cert 0-0 p=1 | regret -3.33 [-7.54,-0.35] p=0.0039
462
+ K=4 st=0 diff_s1 vs idw cov 3-0 p=0.25 | cert 1-0 p=1 | regret -3.49 [-8.54,-0.35] p=0.00194
463
+ K=4 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-3 p=0.25 | regret -0.00 [-0.18,+0.22] p=0.536
464
+ K=4 st=0 diff_s1 vs dls * cov 0-0 p=1 | cert 0-3 p=0.25 | regret +0.06 [-0.41,+0.23] p=0.669
465
+ K=4 st=0 diff_s2 vs nn cov 2-0 p=0.5 | cert 0-2 p=0.5 | regret -2.14 [-7.53,-0.21] p=0.00355
466
+ K=4 st=0 diff_s2 vs idw cov 3-0 p=0.25 | cert 1-2 p=1 | regret -3.55 [-8.53,-0.37] p=0.00117
467
+ K=4 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-5 p=0.0625 | regret -0.11 [-0.22,+0.45] p=0.303
468
+ K=4 st=0 diff_s2 vs dls * cov 0-0 p=1 | cert 0-5 p=0.0625 | regret -0.17 [-0.38,+0.37] p=0.627
469
+ K=4 st=1 diff_s0 vs nn cov 1-0 p=1 | cert 0-6 p=0.0312 | regret -0.02 [-0.22,+0.00] p=0.0798
470
+ K=4 st=1 diff_s0 vs idw cov 1-0 p=1 | cert 0-6 p=0.0312 | regret -0.02 [-0.35,+0.00] p=0.0657
471
+ K=4 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 10-4 p=0.18 | regret -0.00 [-0.02,+0.00] p=0.517
472
+ K=4 st=1 diff_s0 vs dls * cov 0-0 p=1 | cert 13-4 p=0.049 | regret -0.01 [-0.03,+0.00] p=0.32
473
+ K=4 st=1 diff_s1 vs nn cov 1-0 p=1 | cert 0-3 p=0.25 | regret -0.04 [-0.29,-0.00] p=0.046
474
+ K=4 st=1 diff_s1 vs idw cov 1-0 p=1 | cert 0-3 p=0.25 | regret -0.01 [-0.49,+0.00] p=0.0836
475
+ K=4 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 13-4 p=0.049 | regret -0.02 [-0.04,-0.00] p=0.0529
476
+ K=4 st=1 diff_s1 vs dls * cov 0-0 p=1 | cert 16-4 p=0.0118 | regret -0.01 [-0.06,-0.00] p=0.0386
477
+ K=4 st=1 diff_s2 vs nn cov 1-0 p=1 | cert 0-8 p=0.00781 | regret -0.05 [-0.25,+0.00] p=0.0556
478
+ K=4 st=1 diff_s2 vs idw cov 1-0 p=1 | cert 0-8 p=0.00781 | regret -0.04 [-0.38,+0.00] p=0.0372
479
+ K=4 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 8-4 p=0.388 | regret -0.00 [-0.03,+0.00] p=0.41
480
+ K=4 st=1 diff_s2 vs dls * cov 0-0 p=1 | cert 11-4 p=0.118 | regret -0.01 [-0.03,+0.00] p=0.263
481
+ K=8 st=0 diff_s0 vs nn cov 2-0 p=0.5 | cert 0-1 p=1 | regret -1.33 [-4.89,-0.04] p=0.00449
482
+ K=8 st=0 diff_s0 vs idw cov 3-0 p=0.25 | cert 1-1 p=1 | regret -1.33 [-5.16,-0.04] p=0.00316
483
+ K=8 st=0 diff_s0 vs nns * cov 0-0 p=1 | cert 0-4 p=0.125 | regret -0.08 [-0.20,+0.29] p=0.606
484
+ K=8 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-4 p=0.125 | regret -0.11 [-0.41,+0.08] p=0.912
485
+ K=8 st=0 diff_s1 vs nn cov 2-0 p=0.5 | cert 0-0 p=1 | regret -1.67 [-4.70,-0.50] p=0.00287
486
+ K=8 st=0 diff_s1 vs idw cov 3-0 p=0.25 | cert 1-0 p=1 | regret -1.67 [-4.94,-0.41] p=0.00171
487
+ K=8 st=0 diff_s1 vs nns * cov 0-0 p=1 | cert 0-3 p=0.25 | regret -0.10 [-0.31,+0.13] p=0.819
488
+ K=8 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-3 p=0.25 | regret -0.26 [-0.59,+0.15] p=0.696
489
+ K=8 st=0 diff_s2 vs nn cov 2-0 p=0.5 | cert 0-0 p=1 | regret -1.45 [-5.03,-0.05] p=0.00787
490
+ K=8 st=0 diff_s2 vs idw cov 3-0 p=0.25 | cert 1-0 p=1 | regret -1.45 [-5.60,-0.05] p=0.0028
491
+ K=8 st=0 diff_s2 vs nns * cov 0-0 p=1 | cert 0-3 p=0.25 | regret -0.12 [-0.30,+0.07] p=0.947
492
+ K=8 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-3 p=0.25 | regret -0.21 [-0.40,+0.03] p=0.74
493
+ K=8 st=1 diff_s0 vs nn cov 1-0 p=1 | cert 0-4 p=0.125 | regret -0.00 [-0.05,+0.00] p=0.0983
494
+ K=8 st=1 diff_s0 vs idw * cov 1-0 p=1 | cert 0-4 p=0.125 | regret -0.00 [-0.08,+0.00] p=0.122
495
+ K=8 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 12-4 p=0.0768 | regret -0.00 [-0.04,-0.00] p=0.0221
496
+ K=8 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 13-4 p=0.049 | regret -0.01 [-0.08,-0.00] p=0.04
497
+ K=8 st=1 diff_s1 vs nn cov 1-0 p=1 | cert 0-2 p=0.5 | regret -0.02 [-0.16,+0.00] p=0.052
498
+ K=8 st=1 diff_s1 vs idw * cov 1-0 p=1 | cert 0-2 p=0.5 | regret -0.01 [-0.14,+0.00] p=0.0494
499
+ K=8 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 14-4 p=0.0309 | regret -0.02 [-0.11,-0.00] p=0.0154
500
+ K=8 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 15-4 p=0.0192 | regret -0.02 [-0.08,-0.01] p=0.0106
501
+ K=8 st=1 diff_s2 vs nn cov 1-0 p=1 | cert 0-5 p=0.0625 | regret -0.01 [-0.10,+0.00] p=0.12
502
+ K=8 st=1 diff_s2 vs idw * cov 1-0 p=1 | cert 0-5 p=0.0625 | regret -0.01 [-0.12,+0.00] p=0.085
503
+ K=8 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 10-3 p=0.0923 | regret -0.01 [-0.03,+0.00] p=0.0773
504
+ K=8 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 11-3 p=0.0574 | regret -0.01 [-0.03,-0.00] p=0.0918
505
+
506
+ === hole_in: N=60 NN-dist ratio test/bank 2.5x (mean NN 0.368); oracle finite 60 ===
507
+ -- no steering: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
508
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
509
+ diff_s0 | 1.000 1.000 1.62 | 1.000 1.000 0.91 | 1.000 1.000 0.68
510
+ diff_s1 | 1.000 1.000 1.49 | 1.000 1.000 0.82 | 1.000 1.000 0.68
511
+ diff_s2 | 1.000 1.000 1.58 | 1.000 1.000 0.95 | 1.000 1.000 0.74
512
+ x_s43 | 1.000 1.000 1.55 | 1.000 1.000 0.61 | 1.000 1.000 0.42
513
+ x_s44 | 1.000 1.000 1.30 | 1.000 1.000 0.64 | 1.000 1.000 0.41
514
+ nn | 1.000 1.000 1.95 | 1.000 1.000 0.56 | 1.000 1.000 0.45
515
+ idw | 1.000 1.000 1.95 | 1.000 1.000 0.64 | 1.000 1.000 0.41
516
+ nns | 1.000 1.000 1.46 | 1.000 1.000 1.30 | 1.000 1.000 1.17
517
+ dls | 1.000 1.000 1.39 | 1.000 1.000 1.23 | 1.000 1.000 1.16
518
+
519
+ -- 50-step steering on every candidate: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
520
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
521
+ diff_s0 | 1.000 1.000 0.01 | 1.000 1.000 -0.00 | 1.000 1.000 -0.00
522
+ diff_s1 | 1.000 1.000 0.00 | 1.000 1.000 -0.00 | 1.000 1.000 -0.00
523
+ diff_s2 | 1.000 1.000 0.01 | 1.000 1.000 -0.00 | 1.000 1.000 -0.00
524
+ x_s43 | 1.000 0.983 0.01 | 1.000 1.000 -0.00 | 1.000 1.000 -0.00
525
+ x_s44 | 1.000 0.983 0.00 | 1.000 1.000 -0.00 | 1.000 1.000 -0.00
526
+ nn | 1.000 1.000 0.04 | 1.000 1.000 0.00 | 1.000 1.000 -0.00
527
+ idw | 1.000 1.000 0.04 | 1.000 1.000 0.01 | 1.000 1.000 -0.00
528
+ nns | 1.000 1.000 0.00 | 1.000 1.000 -0.00 | 1.000 1.000 -0.00
529
+ dls | 1.000 1.000 0.00 | 1.000 1.000 -0.00 | 1.000 1.000 -0.00
530
+
531
+ -- paired tests, diffusion (model - retrieval): [cov McNemar model-only/retr-only p | certified-cov | paired regret diff median [95% CI], Wilcoxon p]; * = best retrieval arm by median regret
532
+ K=1 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.14 [-0.76,+0.27] p=0.178
533
+ K=1 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.14 [-0.76,+0.27] p=0.178
534
+ K=1 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.09 [-0.16,+0.25] p=0.627
535
+ K=1 st=0 diff_s0 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.12 [-0.23,+0.33] p=0.317
536
+ K=1 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.21 [-0.59,+0.33] p=0.292
537
+ K=1 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.21 [-0.59,+0.33] p=0.292
538
+ K=1 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.38,+0.30] p=0.494
539
+ K=1 st=0 diff_s1 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.05 [-0.25,+0.57] p=0.183
540
+ K=1 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.09 [-0.73,+0.36] p=0.541
541
+ K=1 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.09 [-0.73,+0.36] p=0.541
542
+ K=1 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.15 [+0.06,+0.22] p=0.169
543
+ K=1 st=0 diff_s2 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.06 [-0.09,+0.31] p=0.149
544
+ K=1 st=1 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.02 [-0.06,+0.00] p=0.00419
545
+ K=1 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.02 [-0.06,+0.00] p=0.00419
546
+ K=1 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [+0.00,+0.03] p=0.004
547
+ K=1 st=1 diff_s0 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.05] p=0.0316
548
+ K=1 st=1 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.03 [-0.05,-0.00] p=0.00897
549
+ K=1 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.03 [-0.05,-0.00] p=0.00897
550
+ K=1 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.076
551
+ K=1 st=1 diff_s1 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.109
552
+ K=1 st=1 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.01 [-0.04,+0.00] p=0.052
553
+ K=1 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.01 [-0.04,+0.00] p=0.052
554
+ K=1 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.02 [+0.00,+0.04] p=3.19e-05
555
+ K=1 st=1 diff_s2 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [-0.00,+0.02] p=0.00381
556
+ K=4 st=0 diff_s0 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.25 [-0.07,+0.43] p=0.576
557
+ K=4 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.16 [-0.19,+0.41] p=0.707
558
+ K=4 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.37 [-0.58,-0.19] p=1.07e-05
559
+ K=4 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.25 [-0.62,-0.19] p=0.0012
560
+ K=4 st=0 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.22 [+0.05,+0.48] p=0.195
561
+ K=4 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.18 [-0.04,+0.41] p=0.842
562
+ K=4 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.35 [-0.63,-0.09] p=0.000154
563
+ K=4 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.36 [-0.65,-0.11] p=0.0102
564
+ K=4 st=0 diff_s2 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.22 [+0.04,+0.39] p=0.248
565
+ K=4 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.14 [-0.14,+0.35] p=0.994
566
+ K=4 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.26 [-0.43,-0.07] p=0.000831
567
+ K=4 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.31 [-0.47,-0.12] p=0.0123
568
+ K=4 st=1 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.01,+0.00] p=0.0181
569
+ K=4 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.02,-0.00] p=0.000949
570
+ K=4 st=1 diff_s0 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.78
571
+ K=4 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.494
572
+ K=4 st=1 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.219
573
+ K=4 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.02,-0.00] p=0.0238
574
+ K=4 st=1 diff_s1 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.844
575
+ K=4 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.556
576
+ K=4 st=1 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.01,+0.00] p=0.04
577
+ K=4 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.01 [-0.02,-0.00] p=0.00504
578
+ K=4 st=1 diff_s2 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.953
579
+ K=4 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.587
580
+ K=8 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret +0.15 [+0.01,+0.35] p=0.0257
581
+ K=8 st=0 diff_s0 vs idw * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.18 [+0.02,+0.35] p=0.0366
582
+ K=8 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.37 [-0.56,-0.20] p=1.8e-07
583
+ K=8 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.51 [-0.71,-0.34] p=0.00054
584
+ K=8 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret +0.20 [+0.04,+0.29] p=0.0422
585
+ K=8 st=0 diff_s1 vs idw * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.16 [+0.04,+0.29] p=0.0407
586
+ K=8 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.41 [-0.59,-0.25] p=6.28e-08
587
+ K=8 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.55 [-0.70,-0.24] p=0.000347
588
+ K=8 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret +0.14 [+0.05,+0.30] p=0.0257
589
+ K=8 st=0 diff_s2 vs idw * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.14 [+0.05,+0.30] p=0.0267
590
+ K=8 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret -0.34 [-0.57,-0.23] p=9.1e-07
591
+ K=8 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.41 [-0.62,-0.14] p=0.000586
592
+ K=8 st=1 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.328
593
+ K=8 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.185
594
+ K=8 st=1 diff_s0 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.0904
595
+ K=8 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.405
596
+ K=8 st=1 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.785
597
+ K=8 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.664
598
+ K=8 st=1 diff_s1 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.723
599
+ K=8 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.591
600
+ K=8 st=1 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.402
601
+ K=8 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.31
602
+ K=8 st=1 diff_s2 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.0393
603
+ K=8 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret -0.00 [-0.00,+0.00] p=0.994
604
+
605
+ === hole_out: N=60 NN-dist ratio test/bank 1.0x (mean NN 0.149); oracle finite 60 ===
606
+ -- no steering: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
607
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
608
+ diff_s0 | 1.000 1.000 1.59 | 1.000 1.000 1.29 | 1.000 1.000 1.22
609
+ diff_s1 | 1.000 1.000 1.58 | 1.000 1.000 1.24 | 1.000 1.000 1.16
610
+ diff_s2 | 1.000 1.000 1.47 | 1.000 1.000 1.18 | 1.000 1.000 1.12
611
+ x_s43 | 1.000 0.983 1.43 | 1.000 1.000 0.95 | 1.000 1.000 0.81
612
+ x_s44 | 1.000 1.000 1.31 | 1.000 1.000 0.97 | 1.000 1.000 0.94
613
+ nn | 1.000 1.000 1.40 | 1.000 1.000 0.76 | 1.000 1.000 0.36
614
+ idw | 1.000 1.000 1.40 | 1.000 1.000 0.78 | 1.000 1.000 0.36
615
+ nns | 1.000 1.000 1.35 | 1.000 1.000 1.15 | 1.000 1.000 1.04
616
+ dls | 1.000 1.000 1.38 | 1.000 1.000 1.10 | 1.000 1.000 1.01
617
+
618
+ -- 50-step steering on every candidate: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
619
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
620
+ diff_s0 | 1.000 0.950 0.01 | 1.000 0.967 0.00 | 1.000 0.967 0.00
621
+ diff_s1 | 1.000 0.950 0.02 | 1.000 0.950 0.00 | 1.000 0.983 0.00
622
+ diff_s2 | 1.000 0.950 0.01 | 1.000 0.967 0.00 | 1.000 0.967 0.00
623
+ x_s43 | 1.000 0.933 0.03 | 1.000 0.933 0.00 | 1.000 0.933 -0.00
624
+ x_s44 | 1.000 0.933 0.01 | 1.000 0.933 0.00 | 1.000 0.933 -0.00
625
+ nn | 1.000 0.967 0.01 | 1.000 0.983 -0.00 | 1.000 1.000 -0.01
626
+ idw | 1.000 0.967 0.01 | 1.000 0.983 -0.00 | 1.000 1.000 -0.01
627
+ nns | 1.000 0.950 0.01 | 1.000 0.950 0.00 | 1.000 0.950 0.00
628
+ dls | 1.000 0.950 0.01 | 1.000 0.950 0.00 | 1.000 0.950 0.00
629
+
630
+ -- paired tests, diffusion (model - retrieval): [cov McNemar model-only/retr-only p | certified-cov | paired regret diff median [95% CI], Wilcoxon p]; * = best retrieval arm by median regret
631
+ K=1 st=0 diff_s0 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret +0.17 [-0.04,+0.49] p=0.151
632
+ K=1 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.17 [-0.04,+0.49] p=0.151
633
+ K=1 st=0 diff_s0 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.20 [+0.10,+0.40] p=0.000129
634
+ K=1 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.14 [+0.01,+0.39] p=0.0161
635
+ K=1 st=0 diff_s1 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret +0.20 [-0.14,+0.32] p=0.339
636
+ K=1 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.20 [-0.14,+0.32] p=0.339
637
+ K=1 st=0 diff_s1 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.21 [+0.04,+0.47] p=0.00136
638
+ K=1 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.15 [+0.01,+0.42] p=0.0167
639
+ K=1 st=0 diff_s2 vs nn cov 0-0 p=1 | cert 0-0 p=1 | regret +0.12 [-0.05,+0.38] p=0.401
640
+ K=1 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.12 [-0.05,+0.38] p=0.401
641
+ K=1 st=0 diff_s2 vs nns * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.22 [+0.06,+0.33] p=0.000292
642
+ K=1 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.02 [-0.05,+0.28] p=0.0299
643
+ K=1 st=1 diff_s0 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.01,+0.02] p=0.361
644
+ K=1 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.01,+0.02] p=0.361
645
+ K=1 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 1-1 p=1 | regret +0.00 [+0.00,+0.03] p=0.000946
646
+ K=1 st=1 diff_s0 vs dls * cov 0-0 p=1 | cert 1-1 p=1 | regret +0.01 [+0.00,+0.03] p=0.0014
647
+ K=1 st=1 diff_s1 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.00,+0.01] p=0.435
648
+ K=1 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.00,+0.01] p=0.435
649
+ K=1 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=0.000102
650
+ K=1 st=1 diff_s1 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.02] p=0.000526
651
+ K=1 st=1 diff_s2 vs nn cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.01,+0.01] p=0.531
652
+ K=1 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [-0.01,+0.01] p=0.531
653
+ K=1 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.01 [+0.00,+0.03] p=0.000337
654
+ K=1 st=1 diff_s2 vs dls * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [+0.00,+0.02] p=0.00189
655
+ K=4 st=0 diff_s0 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.45 [+0.31,+0.80] p=3e-07
656
+ K=4 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.43 [+0.28,+0.64] p=1.27e-05
657
+ K=4 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.17 [+0.08,+0.22] p=0.00155
658
+ K=4 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.03 [-0.09,+0.15] p=0.122
659
+ K=4 st=0 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.45 [+0.34,+0.75] p=6.02e-08
660
+ K=4 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.39 [+0.29,+0.71] p=5.97e-06
661
+ K=4 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.17 [+0.08,+0.26] p=0.00176
662
+ K=4 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.05 [-0.03,+0.26] p=0.0359
663
+ K=4 st=0 diff_s2 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.45 [+0.31,+0.69] p=2.77e-07
664
+ K=4 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.40 [+0.24,+0.66] p=1.36e-05
665
+ K=4 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.09 [+0.02,+0.26] p=0.0084
666
+ K=4 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.02 [-0.06,+0.16] p=0.105
667
+ K=4 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.01 [+0.00,+0.03] p=0.0003
668
+ K=4 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [+0.00,+0.03] p=0.00178
669
+ K=4 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 1-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.00574
670
+ K=4 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 1-0 p=1 | regret +0.00 [-0.00,+0.02] p=0.0113
671
+ K=4 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.03] p=0.00105
672
+ K=4 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [-0.00,+0.02] p=0.00214
673
+ K=4 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.0386
674
+ K=4 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.0216
675
+ K=4 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.01 [+0.00,+0.04] p=0.000213
676
+ K=4 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret +0.00 [+0.00,+0.02] p=0.00316
677
+ K=4 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 1-0 p=1 | regret +0.00 [+0.00,+0.01] p=0.0547
678
+ K=4 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 1-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.00459
679
+ K=8 st=0 diff_s0 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.68 [+0.48,+0.92] p=5.19e-10
680
+ K=8 st=0 diff_s0 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.68 [+0.48,+0.92] p=6.26e-10
681
+ K=8 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.16 [+0.05,+0.19] p=0.000239
682
+ K=8 st=0 diff_s0 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.02 [-0.04,+0.16] p=0.0798
683
+ K=8 st=0 diff_s1 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.69 [+0.48,+0.92] p=4.3e-10
684
+ K=8 st=0 diff_s1 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.69 [+0.48,+0.92] p=6.26e-10
685
+ K=8 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.13 [+0.08,+0.27] p=0.000399
686
+ K=8 st=0 diff_s1 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.07 [-0.05,+0.18] p=0.0657
687
+ K=8 st=0 diff_s2 vs nn * cov 0-0 p=1 | cert 0-0 p=1 | regret +0.66 [+0.43,+0.97] p=4.3e-10
688
+ K=8 st=0 diff_s2 vs idw cov 0-0 p=1 | cert 0-0 p=1 | regret +0.66 [+0.43,+0.97] p=5.19e-10
689
+ K=8 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-0 p=1 | regret +0.08 [+0.02,+0.18] p=0.00631
690
+ K=8 st=0 diff_s2 vs dls cov 0-0 p=1 | cert 0-0 p=1 | regret +0.04 [-0.08,+0.12] p=0.185
691
+ K=8 st=1 diff_s0 vs nn * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.02 [+0.00,+0.03] p=1.08e-05
692
+ K=8 st=1 diff_s0 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.02 [+0.00,+0.03] p=8.16e-06
693
+ K=8 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 1-0 p=1 | regret +0.00 [-0.00,+0.01] p=0.0106
694
+ K=8 st=1 diff_s0 vs dls cov 0-0 p=1 | cert 1-0 p=1 | regret +0.00 [-0.00,+0.02] p=0.00645
695
+ K=8 st=1 diff_s1 vs nn * cov 0-0 p=1 | cert 0-1 p=1 | regret +0.01 [+0.00,+0.03] p=0.000169
696
+ K=8 st=1 diff_s1 vs idw cov 0-0 p=1 | cert 0-1 p=1 | regret +0.01 [+0.00,+0.03] p=0.000169
697
+ K=8 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 2-0 p=0.5 | regret +0.00 [-0.00,+0.01] p=0.0701
698
+ K=8 st=1 diff_s1 vs dls cov 0-0 p=1 | cert 2-0 p=0.5 | regret +0.00 [-0.00,+0.02] p=0.00419
699
+ K=8 st=1 diff_s2 vs nn * cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.03] p=8.15e-06
700
+ K=8 st=1 diff_s2 vs idw cov 0-0 p=1 | cert 0-2 p=0.5 | regret +0.01 [+0.00,+0.02] p=7.61e-06
701
+ K=8 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 1-0 p=1 | regret +0.00 [-0.00,+0.00] p=0.115
702
+ K=8 st=1 diff_s2 vs dls cov 0-0 p=1 | cert 1-0 p=1 | regret +0.00 [-0.00,+0.02] p=0.00617
703
+
704
+ === pilotfar: N=60 NN-dist ratio test/bank 27.2x (mean NN 3.897); oracle finite 60 ===
705
+ -- no steering: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
706
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
707
+ diff_s0 | 0.967 0.000 1.20 | 1.000 0.033 -5.36 | 1.000 0.033 -6.49
708
+ diff_s1 | 0.950 0.000 1.37 | 1.000 0.000 -4.15 | 1.000 0.000 -6.74
709
+ diff_s2 | 0.933 0.000 0.80 | 1.000 0.000 -3.90 | 1.000 0.017 -6.68
710
+ x_s42 | 0.950 0.017 1.03 | 1.000 0.067 -6.11 | 1.000 0.117 -7.49
711
+ x_s43 | 0.967 0.017 4.81 | 1.000 0.050 -3.51 | 1.000 0.083 -5.40
712
+ x_s44 | 0.950 0.033 3.56 | 1.000 0.083 -4.13 | 1.000 0.117 -5.92
713
+ nn | 0.717 0.517 69.64 | 0.717 0.517 69.64 | 0.717 0.517 69.64
714
+ idw | 0.717 0.517 69.64 | 0.383 0.250 1000.00 | 0.383 0.233 1000.00
715
+ nns | 1.000 1.000 2.25 | 1.000 1.000 0.25 | 1.000 1.000 0.04
716
+ dls | 1.000 1.000 1.05 | 1.000 1.000 0.18 | 1.000 1.000 -1.73
717
+
718
+ -- 50-step steering on every candidate: cov / certified-stable cov / median regret % (inf = unstable counts as failure)
719
+ arm K=1: cov cert regret | K=4: cov cert regret | K=8: cov cert regret
720
+ diff_s0 | 1.000 0.000 -3.21 | 1.000 0.033 -7.11 | 1.000 0.033 -8.32
721
+ diff_s1 | 0.983 0.000 -3.35 | 1.000 0.000 -7.20 | 1.000 0.000 -8.70
722
+ diff_s2 | 0.983 0.000 -3.61 | 1.000 0.000 -6.68 | 1.000 0.017 -8.44
723
+ x_s42 | 0.983 0.017 -3.16 | 1.000 0.067 -7.46 | 1.000 0.117 -8.68
724
+ x_s43 | 0.983 0.017 -0.83 | 1.000 0.050 -6.11 | 1.000 0.083 -7.52
725
+ x_s44 | 0.983 0.033 -1.83 | 1.000 0.083 -6.67 | 1.000 0.117 -7.78
726
+ nn | 0.867 0.550 32.16 | 0.867 0.550 32.16 | 0.867 0.550 32.16
727
+ idw | 0.867 0.550 32.16 | 0.617 0.233 107.65 | 0.617 0.283 99.44
728
+ nns | 1.000 1.000 0.61 | 1.000 1.000 -1.34 | 1.000 1.000 -1.48
729
+ dls | 1.000 1.000 -0.64 | 1.000 1.000 -1.44 | 1.000 1.000 -3.25
730
+
731
+ -- paired tests, diffusion (model - retrieval): [cov McNemar model-only/retr-only p | certified-cov | paired regret diff median [95% CI], Wilcoxon p]; * = best retrieval arm by median regret
732
+ K=1 st=0 diff_s0 vs nn cov 15-0 p=6.1e-05 | cert 0-31 p=9.31e-10 | regret -58.08 [-92.63,-29.15] p=3.51e-11
733
+ K=1 st=0 diff_s0 vs idw cov 15-0 p=6.1e-05 | cert 0-31 p=9.31e-10 | regret -58.08 [-92.63,-29.15] p=3.51e-11
734
+ K=1 st=0 diff_s0 vs nns cov 0-2 p=0.5 | cert 0-60 p=1.73e-18 | regret -1.27 [-4.91,+2.54] p=0.842
735
+ K=1 st=0 diff_s0 vs dls * cov 0-2 p=0.5 | cert 0-60 p=1.73e-18 | regret -0.44 [-3.61,+3.92] p=0.365
736
+ K=1 st=0 diff_s1 vs nn cov 14-0 p=0.000122 | cert 0-31 p=9.31e-10 | regret -58.51 [-88.98,-24.34] p=1.08e-10
737
+ K=1 st=0 diff_s1 vs idw cov 14-0 p=0.000122 | cert 0-31 p=9.31e-10 | regret -58.51 [-88.98,-24.34] p=1.08e-10
738
+ K=1 st=0 diff_s1 vs nns cov 0-3 p=0.25 | cert 0-60 p=1.73e-18 | regret -0.38 [-2.89,+9.62] p=0.0247
739
+ K=1 st=0 diff_s1 vs dls * cov 0-3 p=0.25 | cert 0-60 p=1.73e-18 | regret +0.61 [-1.87,+10.49] p=0.00897
740
+ K=1 st=0 diff_s2 vs nn cov 15-2 p=0.00235 | cert 0-31 p=9.31e-10 | regret -46.96 [-99.93,-28.23] p=3.56e-09
741
+ K=1 st=0 diff_s2 vs idw cov 15-2 p=0.00235 | cert 0-31 p=9.31e-10 | regret -46.96 [-99.93,-28.23] p=3.56e-09
742
+ K=1 st=0 diff_s2 vs nns cov 0-4 p=0.125 | cert 0-60 p=1.73e-18 | regret -1.12 [-4.05,+3.92] p=0.263
743
+ K=1 st=0 diff_s2 vs dls * cov 0-4 p=0.125 | cert 0-60 p=1.73e-18 | regret -0.08 [-2.69,+4.63] p=0.0976
744
+ K=1 st=1 diff_s0 vs nn cov 8-0 p=0.00781 | cert 0-33 p=2.33e-10 | regret -32.32 [-52.19,-18.43] p=1.63e-11
745
+ K=1 st=1 diff_s0 vs idw cov 8-0 p=0.00781 | cert 0-33 p=2.33e-10 | regret -32.32 [-52.19,-18.43] p=1.63e-11
746
+ K=1 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -4.67 [-6.66,-2.56] p=0.0437
747
+ K=1 st=1 diff_s0 vs dls * cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -3.91 [-5.21,-1.34] p=0.0773
748
+ K=1 st=1 diff_s1 vs nn cov 8-1 p=0.0391 | cert 0-33 p=2.33e-10 | regret -34.71 [-45.72,-15.72] p=3.92e-10
749
+ K=1 st=1 diff_s1 vs idw cov 8-1 p=0.0391 | cert 0-33 p=2.33e-10 | regret -34.71 [-45.72,-15.72] p=3.92e-10
750
+ K=1 st=1 diff_s1 vs nns cov 0-1 p=1 | cert 0-60 p=1.73e-18 | regret -3.70 [-5.51,+1.78] p=0.883
751
+ K=1 st=1 diff_s1 vs dls * cov 0-1 p=1 | cert 0-60 p=1.73e-18 | regret -2.71 [-4.72,+2.44] p=0.611
752
+ K=1 st=1 diff_s2 vs nn cov 8-1 p=0.0391 | cert 0-33 p=2.33e-10 | regret -30.56 [-53.61,-16.60] p=1.57e-09
753
+ K=1 st=1 diff_s2 vs idw cov 8-1 p=0.0391 | cert 0-33 p=2.33e-10 | regret -30.56 [-53.61,-16.60] p=1.57e-09
754
+ K=1 st=1 diff_s2 vs nns cov 0-1 p=1 | cert 0-60 p=1.73e-18 | regret -4.43 [-6.23,-2.05] p=0.11
755
+ K=1 st=1 diff_s2 vs dls * cov 0-1 p=1 | cert 0-60 p=1.73e-18 | regret -3.21 [-4.98,-0.61] p=0.227
756
+ K=4 st=0 diff_s0 vs nn cov 17-0 p=1.53e-05 | cert 0-29 p=3.73e-09 | regret -75.78 [-111.25,-48.62] p=1.63e-11
757
+ K=4 st=0 diff_s0 vs idw cov 37-0 p=1.46e-11 | cert 0-13 p=0.000244 | regret -992.50 [-1001.55,-149.64] p=1.63e-11
758
+ K=4 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-58 p=6.94e-18 | regret -7.13 [-7.78,-5.43] p=0.000446
759
+ K=4 st=0 diff_s0 vs dls * cov 0-0 p=1 | cert 0-58 p=6.94e-18 | regret -6.43 [-7.10,-4.69] p=0.000876
760
+ K=4 st=0 diff_s1 vs nn cov 17-0 p=1.53e-05 | cert 0-31 p=9.31e-10 | regret -70.24 [-103.53,-48.90] p=1.63e-11
761
+ K=4 st=0 diff_s1 vs idw cov 37-0 p=1.46e-11 | cert 0-15 p=6.1e-05 | regret -992.34 [-1000.23,-150.59] p=1.63e-11
762
+ K=4 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -4.70 [-6.46,-3.24] p=0.000164
763
+ K=4 st=0 diff_s1 vs dls * cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -5.18 [-6.07,-3.05] p=0.000484
764
+ K=4 st=0 diff_s2 vs nn cov 17-0 p=1.53e-05 | cert 0-31 p=9.31e-10 | regret -71.89 [-114.87,-45.49] p=1.63e-11
765
+ K=4 st=0 diff_s2 vs idw cov 37-0 p=1.46e-11 | cert 0-15 p=6.1e-05 | regret -997.48 [-999.36,-147.56] p=1.63e-11
766
+ K=4 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -5.38 [-6.72,-3.39] p=4.84e-06
767
+ K=4 st=0 diff_s2 vs dls * cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -4.67 [-6.32,-2.59] p=1.71e-05
768
+ K=4 st=1 diff_s0 vs nn cov 8-0 p=0.00781 | cert 0-31 p=9.31e-10 | regret -39.27 [-52.67,-25.66] p=1.63e-11
769
+ K=4 st=1 diff_s0 vs idw cov 23-0 p=2.38e-07 | cert 0-12 p=0.000488 | regret -115.13 [-564.70,-68.45] p=1.63e-11
770
+ K=4 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 0-58 p=6.94e-18 | regret -7.11 [-7.58,-6.27] p=3.12e-07
771
+ K=4 st=1 diff_s0 vs dls * cov 0-0 p=1 | cert 0-58 p=6.94e-18 | regret -6.28 [-7.70,-5.03] p=2.29e-06
772
+ K=4 st=1 diff_s1 vs nn cov 8-0 p=0.00781 | cert 0-33 p=2.33e-10 | regret -36.37 [-51.65,-26.14] p=1.63e-11
773
+ K=4 st=1 diff_s1 vs idw cov 23-0 p=2.38e-07 | cert 0-14 p=0.000122 | regret -111.16 [-571.43,-69.65] p=1.63e-11
774
+ K=4 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -5.95 [-7.10,-5.00] p=3.24e-09
775
+ K=4 st=1 diff_s1 vs dls * cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -5.47 [-6.58,-4.63] p=4.9e-08
776
+ K=4 st=1 diff_s2 vs nn cov 8-0 p=0.00781 | cert 0-33 p=2.33e-10 | regret -37.30 [-56.56,-24.37] p=1.63e-11
777
+ K=4 st=1 diff_s2 vs idw cov 23-0 p=2.38e-07 | cert 0-14 p=0.000122 | regret -113.67 [-572.30,-67.88] p=1.63e-11
778
+ K=4 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -6.32 [-7.28,-5.21] p=9.35e-09
779
+ K=4 st=1 diff_s2 vs dls * cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -5.13 [-7.07,-4.54] p=1.21e-08
780
+ K=8 st=0 diff_s0 vs nn cov 17-0 p=1.53e-05 | cert 0-29 p=3.73e-09 | regret -75.78 [-115.18,-48.62] p=1.63e-11
781
+ K=8 st=0 diff_s0 vs idw cov 37-0 p=1.46e-11 | cert 0-12 p=0.000488 | regret -1001.19 [-1003.63,-149.64] p=1.63e-11
782
+ K=8 st=0 diff_s0 vs nns cov 0-0 p=1 | cert 0-58 p=6.94e-18 | regret -7.87 [-8.53,-6.59] p=7.7e-08
783
+ K=8 st=0 diff_s0 vs dls * cov 0-0 p=1 | cert 0-58 p=6.94e-18 | regret -5.28 [-6.10,-4.50] p=2.95e-06
784
+ K=8 st=0 diff_s1 vs nn cov 17-0 p=1.53e-05 | cert 0-31 p=9.31e-10 | regret -76.92 [-114.56,-48.90] p=1.63e-11
785
+ K=8 st=0 diff_s1 vs idw cov 37-0 p=1.46e-11 | cert 0-14 p=0.000122 | regret -999.54 [-1003.53,-150.96] p=1.63e-11
786
+ K=8 st=0 diff_s1 vs nns cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -7.19 [-7.89,-6.11] p=1.86e-08
787
+ K=8 st=0 diff_s1 vs dls * cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -5.14 [-5.94,-3.69] p=2.95e-06
788
+ K=8 st=0 diff_s2 vs nn cov 17-0 p=1.53e-05 | cert 0-30 p=1.86e-09 | regret -75.14 [-116.95,-47.11] p=1.63e-11
789
+ K=8 st=0 diff_s2 vs idw cov 37-0 p=1.46e-11 | cert 0-13 p=0.000244 | regret -1001.80 [-1003.53,-148.13] p=1.63e-11
790
+ K=8 st=0 diff_s2 vs nns cov 0-0 p=1 | cert 0-59 p=3.47e-18 | regret -7.13 [-8.37,-6.37] p=3.56e-10
791
+ K=8 st=0 diff_s2 vs dls * cov 0-0 p=1 | cert 0-59 p=3.47e-18 | regret -5.25 [-5.94,-3.91] p=1.38e-08
792
+ K=8 st=1 diff_s0 vs nn cov 8-0 p=0.00781 | cert 0-31 p=9.31e-10 | regret -39.55 [-57.32,-25.66] p=1.63e-11
793
+ K=8 st=1 diff_s0 vs idw cov 23-0 p=2.38e-07 | cert 0-15 p=6.1e-05 | regret -107.57 [-576.87,-68.61] p=1.63e-11
794
+ K=8 st=1 diff_s0 vs nns cov 0-0 p=1 | cert 0-58 p=6.94e-18 | regret -7.42 [-7.96,-7.02] p=5.44e-10
795
+ K=8 st=1 diff_s0 vs dls * cov 0-0 p=1 | cert 0-58 p=6.94e-18 | regret -5.01 [-5.26,-4.53] p=8.57e-09
796
+ K=8 st=1 diff_s1 vs nn cov 8-0 p=0.00781 | cert 0-33 p=2.33e-10 | regret -40.52 [-57.25,-26.39] p=1.63e-11
797
+ K=8 st=1 diff_s1 vs idw cov 23-0 p=2.38e-07 | cert 0-17 p=1.53e-05 | regret -106.24 [-576.80,-69.90] p=1.63e-11
798
+ K=8 st=1 diff_s1 vs nns cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -7.06 [-7.42,-6.61] p=4.1e-10
799
+ K=8 st=1 diff_s1 vs dls * cov 0-0 p=1 | cert 0-60 p=1.73e-18 | regret -4.88 [-5.15,-4.32] p=3.88e-09
800
+ K=8 st=1 diff_s2 vs nn cov 8-0 p=0.00781 | cert 0-32 p=4.66e-10 | regret -39.27 [-58.19,-25.03] p=1.63e-11
801
+ K=8 st=1 diff_s2 vs idw cov 23-0 p=2.38e-07 | cert 0-16 p=3.05e-05 | regret -106.29 [-578.28,-68.13] p=1.63e-11
802
+ K=8 st=1 diff_s2 vs nns cov 0-0 p=1 | cert 0-59 p=3.47e-18 | regret -7.15 [-7.37,-6.66] p=2.01e-10
803
+ K=8 st=1 diff_s2 vs dls * cov 0-0 p=1 | cert 0-59 p=3.47e-18 | regret -5.01 [-5.26,-4.22] p=2.21e-10
804
+
results/ood/siso_ood.txt ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === SISO lvl0: N=120, z-NN ratio test/bank 1.1x (median-based 0.242) ===
2
+ -- no steering: coverage (any of K stable) | per-candidate stable rate
3
+ diff_s0 K=1: 0.550 0.550 | K=4: 0.650 0.515 | K=8: 0.667 0.518
4
+ diff_s1 K=1: 0.483 0.483 | K=4: 0.667 0.525 | K=8: 0.692 0.512
5
+ diff_s2 K=1: 0.517 0.517 | K=4: 0.642 0.521 | K=8: 0.667 0.532
6
+ nn K=1: 0.375 0.375 | K=4: 0.608 0.394 | K=8: 0.642 0.401
7
+ idw K=1: 0.375 0.375 | K=4: 0.608 0.423 | K=8: 0.650 0.424
8
+ -- 50-step steering: coverage (any of K stable) | per-candidate stable rate
9
+ diff_s0 K=1: 0.550 0.550 | K=4: 0.617 0.531 | K=8: 0.642 0.531
10
+ diff_s1 K=1: 0.492 0.492 | K=4: 0.617 0.537 | K=8: 0.625 0.531
11
+ diff_s2 K=1: 0.542 0.542 | K=4: 0.650 0.544 | K=8: 0.650 0.545
12
+ nn K=1: 0.375 0.375 | K=4: 0.600 0.398 | K=8: 0.642 0.401
13
+ idw K=1: 0.375 0.375 | K=4: 0.600 0.429 | K=8: 0.642 0.425
14
+ -- paired McNemar, model vs retrieval (model-only / retrieval-only, p); * = best retrieval arm
15
+ K=1 st=0 diff_s0: vs nn* 27-6 p=0.000324 diff +0.175 [+0.083,+0.267]; vs idw 27-6 p=0.000324 diff +0.175 [+0.083,+0.267];
16
+ K=1 st=0 diff_s1: vs nn* 25-12 p=0.047 diff +0.108 [+0.008,+0.200]; vs idw 25-12 p=0.047 diff +0.108 [+0.008,+0.200];
17
+ K=1 st=0 diff_s2: vs nn* 26-9 p=0.00599 diff +0.142 [+0.050,+0.233]; vs idw 26-9 p=0.00599 diff +0.142 [+0.050,+0.233];
18
+ K=1 st=1 diff_s0: vs nn* 24-3 p=4.92e-05 diff +0.175 [+0.100,+0.250]; vs idw 24-3 p=4.92e-05 diff +0.175 [+0.100,+0.250];
19
+ K=1 st=1 diff_s1: vs nn* 24-10 p=0.0243 diff +0.117 [+0.025,+0.208]; vs idw 24-10 p=0.0243 diff +0.117 [+0.025,+0.208];
20
+ K=1 st=1 diff_s2: vs nn* 27-7 p=0.000821 diff +0.167 [+0.075,+0.258]; vs idw 27-7 p=0.000821 diff +0.167 [+0.075,+0.258];
21
+ K=4 st=0 diff_s0: vs nn* 7-2 p=0.18 diff +0.042 [-0.008,+0.092]; vs idw 8-3 p=0.227 diff +0.042 [-0.008,+0.100];
22
+ K=4 st=0 diff_s1: vs nn* 8-1 p=0.0391 diff +0.058 [+0.017,+0.108]; vs idw 8-1 p=0.0391 diff +0.058 [+0.008,+0.108];
23
+ K=4 st=0 diff_s2: vs nn* 5-1 p=0.219 diff +0.033 [-0.008,+0.075]; vs idw 5-1 p=0.219 diff +0.033 [+0.000,+0.075];
24
+ K=4 st=1 diff_s0: vs nn* 3-1 p=0.625 diff +0.017 [-0.017,+0.050]; vs idw 3-1 p=0.625 diff +0.017 [-0.017,+0.050];
25
+ K=4 st=1 diff_s1: vs nn* 4-2 p=0.688 diff +0.017 [-0.025,+0.058]; vs idw 3-1 p=0.625 diff +0.017 [-0.017,+0.050];
26
+ K=4 st=1 diff_s2: vs nn* 7-1 p=0.0703 diff +0.050 [+0.008,+0.092]; vs idw 6-0 p=0.0312 diff +0.050 [+0.017,+0.092];
27
+ K=8 st=0 diff_s0: vs nn 4-1 p=0.375 diff +0.025 [-0.008,+0.067]; vs idw* 4-2 p=0.688 diff +0.017 [-0.025,+0.058];
28
+ K=8 st=0 diff_s1: vs nn 6-0 p=0.0312 diff +0.050 [+0.017,+0.092]; vs idw* 5-0 p=0.0625 diff +0.042 [+0.008,+0.075];
29
+ K=8 st=0 diff_s2: vs nn 4-1 p=0.375 diff +0.025 [-0.008,+0.067]; vs idw* 4-2 p=0.688 diff +0.017 [-0.025,+0.058];
30
+ K=8 st=1 diff_s0: vs nn* 1-1 p=1 diff +0.000 [-0.025,+0.025]; vs idw 1-1 p=1 diff +0.000 [-0.025,+0.025];
31
+ K=8 st=1 diff_s1: vs nn* 1-3 p=0.625 diff -0.017 [-0.050,+0.017]; vs idw 1-3 p=0.625 diff -0.017 [-0.050,+0.017];
32
+ K=8 st=1 diff_s2: vs nn* 2-1 p=1 diff +0.008 [-0.017,+0.042]; vs idw 2-1 p=1 diff +0.008 [-0.017,+0.042];
33
+
34
+ === SISO lvl1.1: N=120, z-NN ratio test/bank 2.0x (median-based 0.499) ===
35
+ -- no steering: coverage (any of K stable) | per-candidate stable rate
36
+ diff_s0 K=1: 0.567 0.567 | K=4: 0.692 0.577 | K=8: 0.717 0.589
37
+ diff_s1 K=1: 0.550 0.550 | K=4: 0.675 0.550 | K=8: 0.692 0.572
38
+ diff_s2 K=1: 0.600 0.600 | K=4: 0.692 0.579 | K=8: 0.708 0.589
39
+ nn K=1: 0.542 0.542 | K=4: 0.675 0.521 | K=8: 0.683 0.502
40
+ idw K=1: 0.542 0.542 | K=4: 0.667 0.544 | K=8: 0.675 0.517
41
+ -- 50-step steering: coverage (any of K stable) | per-candidate stable rate
42
+ diff_s0 K=1: 0.592 0.592 | K=4: 0.683 0.596 | K=8: 0.700 0.601
43
+ diff_s1 K=1: 0.567 0.567 | K=4: 0.675 0.575 | K=8: 0.683 0.592
44
+ diff_s2 K=1: 0.600 0.600 | K=4: 0.708 0.590 | K=8: 0.717 0.594
45
+ nn K=1: 0.550 0.550 | K=4: 0.683 0.529 | K=8: 0.683 0.508
46
+ idw K=1: 0.550 0.550 | K=4: 0.675 0.550 | K=8: 0.675 0.523
47
+ -- paired McNemar, model vs retrieval (model-only / retrieval-only, p); * = best retrieval arm
48
+ K=1 st=0 diff_s0: vs nn* 15-12 p=0.701 diff +0.025 [-0.058,+0.108]; vs idw 15-12 p=0.701 diff +0.025 [-0.058,+0.108];
49
+ K=1 st=0 diff_s1: vs nn* 12-11 p=1 diff +0.008 [-0.067,+0.092]; vs idw 12-11 p=1 diff +0.008 [-0.067,+0.092];
50
+ K=1 st=0 diff_s2: vs nn* 16-9 p=0.23 diff +0.058 [-0.025,+0.142]; vs idw 16-9 p=0.23 diff +0.058 [-0.025,+0.142];
51
+ K=1 st=1 diff_s0: vs nn* 14-9 p=0.405 diff +0.042 [-0.042,+0.117]; vs idw 14-9 p=0.405 diff +0.042 [-0.042,+0.117];
52
+ K=1 st=1 diff_s1: vs nn* 11-9 p=0.824 diff +0.017 [-0.058,+0.092]; vs idw 11-9 p=0.824 diff +0.017 [-0.058,+0.092];
53
+ K=1 st=1 diff_s2: vs nn* 16-10 p=0.327 diff +0.050 [-0.033,+0.133]; vs idw 16-10 p=0.327 diff +0.050 [-0.033,+0.133];
54
+ K=4 st=0 diff_s0: vs nn* 2-0 p=0.5 diff +0.017 [+0.000,+0.042]; vs idw 3-0 p=0.25 diff +0.025 [+0.000,+0.058];
55
+ K=4 st=0 diff_s1: vs nn* 2-2 p=1 diff +0.000 [-0.033,+0.033]; vs idw 3-2 p=1 diff +0.008 [-0.025,+0.042];
56
+ K=4 st=0 diff_s2: vs nn* 3-1 p=0.625 diff +0.017 [-0.017,+0.050]; vs idw 4-1 p=0.375 diff +0.025 [-0.008,+0.067];
57
+ K=4 st=1 diff_s0: vs nn* 1-1 p=1 diff +0.000 [-0.025,+0.025]; vs idw 2-1 p=1 diff +0.008 [-0.017,+0.042];
58
+ K=4 st=1 diff_s1: vs nn* 0-1 p=1 diff -0.008 [-0.025,+0.000]; vs idw 1-1 p=1 diff +0.000 [-0.025,+0.025];
59
+ K=4 st=1 diff_s2: vs nn* 4-1 p=0.375 diff +0.025 [-0.008,+0.067]; vs idw 5-1 p=0.219 diff +0.033 [-0.008,+0.075];
60
+ K=8 st=0 diff_s0: vs nn* 4-0 p=0.125 diff +0.033 [+0.008,+0.067]; vs idw 5-0 p=0.0625 diff +0.042 [+0.008,+0.083];
61
+ K=8 st=0 diff_s1: vs nn* 3-2 p=1 diff +0.008 [-0.025,+0.042]; vs idw 3-1 p=0.625 diff +0.017 [-0.017,+0.050];
62
+ K=8 st=0 diff_s2: vs nn* 3-0 p=0.25 diff +0.025 [+0.000,+0.058]; vs idw 4-0 p=0.125 diff +0.033 [+0.008,+0.067];
63
+ K=8 st=1 diff_s0: vs nn* 3-1 p=0.625 diff +0.017 [-0.017,+0.050]; vs idw 3-0 p=0.25 diff +0.025 [+0.000,+0.058];
64
+ K=8 st=1 diff_s1: vs nn* 1-1 p=1 diff +0.000 [-0.025,+0.025]; vs idw 2-1 p=1 diff +0.008 [-0.017,+0.042];
65
+ K=8 st=1 diff_s2: vs nn* 4-0 p=0.125 diff +0.033 [+0.008,+0.067]; vs idw 5-0 p=0.0625 diff +0.042 [+0.008,+0.083];
66
+
67
+ === SISO lvl1.25: N=120, z-NN ratio test/bank 3.3x (median-based 0.592) ===
68
+ -- no steering: coverage (any of K stable) | per-candidate stable rate
69
+ diff_s0 K=1: 0.492 0.492 | K=4: 0.592 0.492 | K=8: 0.608 0.489
70
+ diff_s1 K=1: 0.508 0.508 | K=4: 0.583 0.481 | K=8: 0.583 0.482
71
+ diff_s2 K=1: 0.475 0.475 | K=4: 0.583 0.490 | K=8: 0.600 0.491
72
+ nn K=1: 0.458 0.458 | K=4: 0.575 0.429 | K=8: 0.575 0.431
73
+ idw K=1: 0.458 0.458 | K=4: 0.575 0.450 | K=8: 0.575 0.444
74
+ -- 50-step steering: coverage (any of K stable) | per-candidate stable rate
75
+ diff_s0 K=1: 0.517 0.517 | K=4: 0.583 0.510 | K=8: 0.583 0.508
76
+ diff_s1 K=1: 0.517 0.517 | K=4: 0.583 0.494 | K=8: 0.592 0.495
77
+ diff_s2 K=1: 0.492 0.492 | K=4: 0.583 0.502 | K=8: 0.600 0.505
78
+ nn K=1: 0.467 0.467 | K=4: 0.575 0.435 | K=8: 0.575 0.435
79
+ idw K=1: 0.467 0.467 | K=4: 0.575 0.454 | K=8: 0.575 0.447
80
+ -- paired McNemar, model vs retrieval (model-only / retrieval-only, p); * = best retrieval arm
81
+ K=1 st=0 diff_s0: vs nn* 12-8 p=0.503 diff +0.033 [-0.042,+0.108]; vs idw 12-8 p=0.503 diff +0.033 [-0.042,+0.108];
82
+ K=1 st=0 diff_s1: vs nn* 11-5 p=0.21 diff +0.050 [-0.017,+0.117]; vs idw 11-5 p=0.21 diff +0.050 [-0.017,+0.117];
83
+ K=1 st=0 diff_s2: vs nn* 12-10 p=0.832 diff +0.017 [-0.067,+0.092]; vs idw 12-10 p=0.832 diff +0.017 [-0.067,+0.092];
84
+ K=1 st=1 diff_s0: vs nn* 12-6 p=0.238 diff +0.050 [-0.017,+0.117]; vs idw 12-6 p=0.238 diff +0.050 [-0.017,+0.117];
85
+ K=1 st=1 diff_s1: vs nn* 11-5 p=0.21 diff +0.050 [-0.017,+0.117]; vs idw 11-5 p=0.21 diff +0.050 [-0.017,+0.117];
86
+ K=1 st=1 diff_s2: vs nn* 12-9 p=0.664 diff +0.025 [-0.050,+0.100]; vs idw 12-9 p=0.664 diff +0.025 [-0.050,+0.100];
87
+ K=4 st=0 diff_s0: vs nn* 2-0 p=0.5 diff +0.017 [+0.000,+0.042]; vs idw 2-0 p=0.5 diff +0.017 [+0.000,+0.042];
88
+ K=4 st=0 diff_s1: vs nn* 1-0 p=1 diff +0.008 [+0.000,+0.025]; vs idw 1-0 p=1 diff +0.008 [+0.000,+0.025];
89
+ K=4 st=0 diff_s2: vs nn* 1-0 p=1 diff +0.008 [+0.000,+0.025]; vs idw 1-0 p=1 diff +0.008 [+0.000,+0.025];
90
+ K=4 st=1 diff_s0: vs nn* 1-0 p=1 diff +0.008 [+0.000,+0.025]; vs idw 1-0 p=1 diff +0.008 [+0.000,+0.025];
91
+ K=4 st=1 diff_s1: vs nn* 1-0 p=1 diff +0.008 [+0.000,+0.025]; vs idw 1-0 p=1 diff +0.008 [+0.000,+0.025];
92
+ K=4 st=1 diff_s2: vs nn* 1-0 p=1 diff +0.008 [+0.000,+0.025]; vs idw 1-0 p=1 diff +0.008 [+0.000,+0.025];
93
+ K=8 st=0 diff_s0: vs nn* 4-0 p=0.125 diff +0.033 [+0.008,+0.067]; vs idw 4-0 p=0.125 diff +0.033 [+0.008,+0.067];
94
+ K=8 st=0 diff_s1: vs nn* 1-0 p=1 diff +0.008 [+0.000,+0.025]; vs idw 1-0 p=1 diff +0.008 [+0.000,+0.025];
95
+ K=8 st=0 diff_s2: vs nn* 3-0 p=0.25 diff +0.025 [+0.000,+0.058]; vs idw 3-0 p=0.25 diff +0.025 [+0.000,+0.058];
96
+ K=8 st=1 diff_s0: vs nn* 1-0 p=1 diff +0.008 [+0.000,+0.025]; vs idw 1-0 p=1 diff +0.008 [+0.000,+0.025];
97
+ K=8 st=1 diff_s1: vs nn* 2-0 p=0.5 diff +0.017 [+0.000,+0.042]; vs idw 2-0 p=0.5 diff +0.017 [+0.000,+0.042];
98
+ K=8 st=1 diff_s2: vs nn* 3-0 p=0.25 diff +0.025 [+0.000,+0.058]; vs idw 3-0 p=0.25 diff +0.025 [+0.000,+0.058];
99
+
100
+ === SISO lvl1.5: N=120, z-NN ratio test/bank 8.3x (median-based 1.519) ===
101
+ -- no steering: coverage (any of K stable) | per-candidate stable rate
102
+ diff_s0 K=1: 0.533 0.533 | K=4: 0.675 0.517 | K=8: 0.700 0.518
103
+ diff_s1 K=1: 0.508 0.508 | K=4: 0.642 0.481 | K=8: 0.683 0.508
104
+ diff_s2 K=1: 0.517 0.517 | K=4: 0.675 0.502 | K=8: 0.692 0.506
105
+ nn K=1: 0.583 0.583 | K=4: 0.700 0.550 | K=8: 0.700 0.549
106
+ idw K=1: 0.583 0.583 | K=4: 0.700 0.575 | K=8: 0.700 0.559
107
+ -- 50-step steering: coverage (any of K stable) | per-candidate stable rate
108
+ diff_s0 K=1: 0.542 0.542 | K=4: 0.675 0.535 | K=8: 0.692 0.537
109
+ diff_s1 K=1: 0.508 0.508 | K=4: 0.642 0.506 | K=8: 0.675 0.528
110
+ diff_s2 K=1: 0.525 0.525 | K=4: 0.683 0.515 | K=8: 0.708 0.527
111
+ nn K=1: 0.592 0.592 | K=4: 0.692 0.550 | K=8: 0.692 0.550
112
+ idw K=1: 0.592 0.592 | K=4: 0.692 0.573 | K=8: 0.692 0.559
113
+ -- paired McNemar, model vs retrieval (model-only / retrieval-only, p); * = best retrieval arm
114
+ K=1 st=0 diff_s0: vs nn* 10-16 p=0.327 diff -0.050 [-0.133,+0.033]; vs idw 10-16 p=0.327 diff -0.050 [-0.133,+0.033];
115
+ K=1 st=0 diff_s1: vs nn* 7-16 p=0.0931 diff -0.075 [-0.150,+0.000]; vs idw 7-16 p=0.0931 diff -0.075 [-0.150,+0.000];
116
+ K=1 st=0 diff_s2: vs nn* 8-16 p=0.152 diff -0.067 [-0.142,+0.008]; vs idw 8-16 p=0.152 diff -0.067 [-0.142,+0.008];
117
+ K=1 st=1 diff_s0: vs nn* 7-13 p=0.263 diff -0.050 [-0.125,+0.017]; vs idw 7-13 p=0.263 diff -0.050 [-0.125,+0.017];
118
+ K=1 st=1 diff_s1: vs nn* 5-15 p=0.0414 diff -0.083 [-0.158,-0.017]; vs idw 5-15 p=0.0414 diff -0.083 [-0.158,-0.017];
119
+ K=1 st=1 diff_s2: vs nn* 8-16 p=0.152 diff -0.067 [-0.142,+0.008]; vs idw 8-16 p=0.152 diff -0.067 [-0.142,+0.008];
120
+ K=4 st=0 diff_s0: vs nn* 3-6 p=0.508 diff -0.025 [-0.075,+0.025]; vs idw 3-6 p=0.508 diff -0.025 [-0.075,+0.025];
121
+ K=4 st=0 diff_s1: vs nn* 1-8 p=0.0391 diff -0.058 [-0.108,-0.017]; vs idw 1-8 p=0.0391 diff -0.058 [-0.108,-0.017];
122
+ K=4 st=0 diff_s2: vs nn* 2-5 p=0.453 diff -0.025 [-0.067,+0.017]; vs idw 2-5 p=0.453 diff -0.025 [-0.067,+0.017];
123
+ K=4 st=1 diff_s0: vs nn* 2-4 p=0.688 diff -0.017 [-0.058,+0.025]; vs idw 2-4 p=0.688 diff -0.017 [-0.058,+0.025];
124
+ K=4 st=1 diff_s1: vs nn* 1-7 p=0.0703 diff -0.050 [-0.100,-0.008]; vs idw 1-7 p=0.0703 diff -0.050 [-0.100,-0.008];
125
+ K=4 st=1 diff_s2: vs nn* 2-3 p=1 diff -0.008 [-0.042,+0.025]; vs idw 2-3 p=1 diff -0.008 [-0.042,+0.025];
126
+ K=8 st=0 diff_s0: vs nn* 3-3 p=1 diff +0.000 [-0.042,+0.042]; vs idw 3-3 p=1 diff +0.000 [-0.042,+0.042];
127
+ K=8 st=0 diff_s1: vs nn* 3-5 p=0.727 diff -0.017 [-0.067,+0.025]; vs idw 3-5 p=0.727 diff -0.017 [-0.067,+0.025];
128
+ K=8 st=0 diff_s2: vs nn* 2-3 p=1 diff -0.008 [-0.050,+0.025]; vs idw 2-3 p=1 diff -0.008 [-0.050,+0.025];
129
+ K=8 st=1 diff_s0: vs nn* 2-2 p=1 diff +0.000 [-0.033,+0.033]; vs idw 2-2 p=1 diff +0.000 [-0.033,+0.033];
130
+ K=8 st=1 diff_s1: vs nn* 3-5 p=0.727 diff -0.017 [-0.067,+0.025]; vs idw 3-5 p=0.727 diff -0.017 [-0.067,+0.025];
131
+ K=8 st=1 diff_s2: vs nn* 4-2 p=0.688 diff +0.017 [-0.025,+0.058]; vs idw 4-2 p=0.688 diff +0.017 [-0.025,+0.058];
132
+
133
+ === SISO lvl2.0: N=120, z-NN ratio test/bank 36.3x (median-based 6.683) ===
134
+ -- no steering: coverage (any of K stable) | per-candidate stable rate
135
+ diff_s0 K=1: 0.408 0.408 | K=4: 0.608 0.433 | K=8: 0.667 0.438
136
+ diff_s1 K=1: 0.433 0.433 | K=4: 0.592 0.435 | K=8: 0.642 0.453
137
+ diff_s2 K=1: 0.450 0.450 | K=4: 0.642 0.458 | K=8: 0.692 0.448
138
+ nn K=1: 0.792 0.792 | K=4: 0.850 0.744 | K=8: 0.850 0.717
139
+ idw K=1: 0.792 0.792 | K=4: 0.850 0.758 | K=8: 0.850 0.724
140
+ -- 50-step steering: coverage (any of K stable) | per-candidate stable rate
141
+ diff_s0 K=1: 0.442 0.442 | K=4: 0.608 0.446 | K=8: 0.658 0.456
142
+ diff_s1 K=1: 0.458 0.458 | K=4: 0.625 0.465 | K=8: 0.658 0.474
143
+ diff_s2 K=1: 0.458 0.458 | K=4: 0.650 0.471 | K=8: 0.708 0.463
144
+ nn K=1: 0.783 0.783 | K=4: 0.850 0.740 | K=8: 0.850 0.715
145
+ idw K=1: 0.783 0.783 | K=4: 0.850 0.756 | K=8: 0.850 0.723
146
+ -- paired McNemar, model vs retrieval (model-only / retrieval-only, p); * = best retrieval arm
147
+ K=1 st=0 diff_s0: vs nn* 3-49 p=1.04e-11 diff -0.383 [-0.483,-0.283]; vs idw 3-49 p=1.04e-11 diff -0.383 [-0.483,-0.283];
148
+ K=1 st=0 diff_s1: vs nn* 4-47 p=2.42e-10 diff -0.358 [-0.458,-0.258]; vs idw 4-47 p=2.42e-10 diff -0.358 [-0.458,-0.258];
149
+ K=1 st=0 diff_s2: vs nn* 4-45 p=8.23e-10 diff -0.342 [-0.442,-0.242]; vs idw 4-45 p=8.23e-10 diff -0.342 [-0.442,-0.242];
150
+ K=1 st=1 diff_s0: vs nn* 3-44 p=2.46e-10 diff -0.342 [-0.442,-0.250]; vs idw 3-44 p=2.46e-10 diff -0.342 [-0.442,-0.250];
151
+ K=1 st=1 diff_s1: vs nn* 5-44 p=7.6e-09 diff -0.325 [-0.425,-0.225]; vs idw 5-44 p=7.6e-09 diff -0.325 [-0.425,-0.225];
152
+ K=1 st=1 diff_s2: vs nn* 4-43 p=2.78e-09 diff -0.325 [-0.425,-0.225]; vs idw 4-43 p=2.78e-09 diff -0.325 [-0.425,-0.225];
153
+ K=4 st=0 diff_s0: vs nn* 1-30 p=2.98e-08 diff -0.242 [-0.325,-0.158]; vs idw 1-30 p=2.98e-08 diff -0.242 [-0.325,-0.158];
154
+ K=4 st=0 diff_s1: vs nn* 2-33 p=3.67e-08 diff -0.258 [-0.342,-0.175]; vs idw 2-33 p=3.67e-08 diff -0.258 [-0.342,-0.175];
155
+ K=4 st=0 diff_s2: vs nn* 1-26 p=4.17e-07 diff -0.208 [-0.283,-0.133]; vs idw 1-26 p=4.17e-07 diff -0.208 [-0.283,-0.133];
156
+ K=4 st=1 diff_s0: vs nn* 1-30 p=2.98e-08 diff -0.242 [-0.325,-0.158]; vs idw 1-30 p=2.98e-08 diff -0.242 [-0.325,-0.158];
157
+ K=4 st=1 diff_s1: vs nn* 3-30 p=1.4e-06 diff -0.225 [-0.308,-0.142]; vs idw 3-30 p=1.4e-06 diff -0.225 [-0.308,-0.142];
158
+ K=4 st=1 diff_s2: vs nn* 1-25 p=8.05e-07 diff -0.200 [-0.275,-0.125]; vs idw 1-25 p=8.05e-07 diff -0.200 [-0.275,-0.125];
159
+ K=8 st=0 diff_s0: vs nn* 1-23 p=2.98e-06 diff -0.183 [-0.258,-0.117]; vs idw 1-23 p=2.98e-06 diff -0.183 [-0.258,-0.117];
160
+ K=8 st=0 diff_s1: vs nn* 2-27 p=1.62e-06 diff -0.208 [-0.292,-0.125]; vs idw 2-27 p=1.62e-06 diff -0.208 [-0.292,-0.125];
161
+ K=8 st=0 diff_s2: vs nn* 1-20 p=2.1e-05 diff -0.158 [-0.233,-0.092]; vs idw 1-20 p=2.1e-05 diff -0.158 [-0.233,-0.092];
162
+ K=8 st=1 diff_s0: vs nn* 1-24 p=1.55e-06 diff -0.192 [-0.267,-0.117]; vs idw 1-24 p=1.55e-06 diff -0.192 [-0.267,-0.117];
163
+ K=8 st=1 diff_s1: vs nn* 3-26 p=1.52e-05 diff -0.192 [-0.275,-0.117]; vs idw 3-26 p=1.52e-05 diff -0.192 [-0.275,-0.117];
164
+ K=8 st=1 diff_s2: vs nn* 1-18 p=7.63e-05 diff -0.142 [-0.217,-0.075]; vs idw 1-18 p=7.63e-05 diff -0.142 [-0.217,-0.075];
165
+
results/ood/train_holes.sh ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+ cd /home/cercola/matte_server/DiffusionControl
3
+ export PEND_DATA=data/pendulum_dataset_hole20.pt OMP_NUM_THREADS=1 WANDB_MODE=disabled PYTHONPATH=. CUDA_VISIBLE_DEVICES=0
4
+ for s in 42 43 44; do sfx=""; [ $s != 42 ] && sfx="_s$s"
5
+ myenv/bin/python -u -m nonlinear.training.train_pendulum_ablation --tag ood_hole20$sfx --data_fraction 1.0 --seed $s --epochs 1000
6
+ done
results/pendulum_full/dp_bank_rank.txt ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === DP policy bank + rank study (torque-bounded pendulum) ===
2
+ 300 plants, policy grid 60x61 = 3660 values
3
+ family: l[0.5,1.5] m[0.3,1.2] zeta~logU(0.01, 0.15) rho_u~U(0.35, 0.65)
4
+
5
+ variance explained: k=1:42.43% k=2:66.55% k=4:78.23% k=8:85.51% k=16:91.72% k=32:95.43% k=64:97.74%
6
+
7
+ rank scored by ROLLOUT COST vs each plant's own DP (MSE is not the metric that matters):
8
+ k var% med regret p90 max fell over
9
+ (reference: 25/25 held-out DP policies stabilise at this grid)
10
+ 1 42.43% +9.52% +40.59% +71.57% 1/25
11
+ 2 66.55% +9.69% +22.06% +36.00% 0/25
12
+ 4 78.23% +4.33% +20.12% +32.76% 0/25
13
+ 8 85.51% +3.43% +17.22% +39.13% 0/25
14
+ 16 91.72% +2.07% +8.99% +15.06% 0/25
15
+ 32 95.43% +0.46% +1.54% +7.44% 0/25
16
+ 64 97.74% +0.16% +0.69% +0.82% 0/25
17
+
18
+ This picks the generation dimension only. It is NOT the test of whether a model is
19
+ worth training: a 256-solve interpolation table already reaches +0.1% median regret
20
+ against per-plant value iteration, so the median is settled. The open question is the
21
+ ~8-13% tail where that table loses 10-25%, and it is answered by the matched-budget
22
+ comparison (same K candidates, same rollout verifier, both scored vs per-plant VI),
23
+ not by anything measurable here.
results/pendulum_full/dp_baseline.txt ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === DP (value iteration) vs the hand-designed two-stage architecture ===
2
+ DP assumes no architecture: it returns the globally optimal u*(theta,omega) under the
3
+ torque bound. The architecture restricts u to -phi(c1*theta + c2*omega) plus an energy
4
+ swing-up stage. Same cost, same RK4 loop, same 12 s horizon, same hanging IC.
5
+
6
+ plant rho_u J_DP J_arch J_fixed arch/DP fixed/DP t_DP
7
+ 0 0.416 17.13 19.80 19.80 1.16x 1.16x 2.1s
8
+ 48 0.521 13.02 14.02 14.05 1.08x 1.08x 2.1s
9
+ 95 0.448 69.02 80.70 80.75 1.17x 1.17x 2.2s
10
+ 143 0.625 73.23 107.91 109.09 1.47x 1.49x 2.2s
11
+ 190 0.555 70.50 90.75 90.77 1.29x 1.29x 2.2s
12
+ 239 0.556 14.64 15.73 15.76 1.07x 1.08x 2.2s
13
+
14
+ median J_arch / J_DP = 1.16x (how much the hand-designed architecture gives up vs optimal)
15
+ median J_fixed / J_DP = 1.16x
16
+ median DP solve time = 2.2 s per plant
17
+
18
+ Reading: if J_arch/J_DP is close to 1 the architecture is near-optimal and there is
19
+ nothing for a richer generated controller to win. If it is well above 1, the earlier
20
+ '+0.3% headroom' was a property of the restricted controller class, not of the family,
21
+ and the amortization target should be DP itself.
22
+
23
+ NOTE the controller classes differ: DP gives u(theta,omega) on the full state; the
24
+ spline gives u(c1*theta + c2*omega), a 1-D projection. Part of any gap is that
25
+ projection, which a generated 2-D feedback map could remove.
results/pendulum_full/dp_interp.txt ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === is the pendulum plant->controller map just INTERPOLABLE? ===
2
+
3
+ --- A. same (zeta, rho_u), different (l, m): do the policies coincide? ---
4
+ l=1.00 m=0.75: reference, J=42.284
5
+ l=0.60 m=1.10: max|du_hat|=1.6000 rms=0.2239 J=30.267 J/J_ref=0.716
6
+ l=1.40 m=0.35: max|du_hat|=1.7000 rms=0.2459 J=33.782 J/J_ref=0.799
7
+ (small max|du_hat| => the family really is 2-D in (zeta, rho_u))
8
+
9
+ --- B. 5x5 table of DP solves over (log zeta, rho_u), then interpolate ---
10
+ table built: 25 DP solves in 57 s (2.3 s each), reused for every plant
11
+ plant zeta rho_u J_ownDP J_interp ratio
12
+ 55 0.0132 0.623 57.13 57.27 1.002x
13
+ 207 0.0115 0.503 57.30 57.49 1.003x
14
+ 190 0.0991 0.555 70.06 70.40 1.005x
15
+ 236 0.1094 0.623 22.66 27.03 1.193x
16
+ 42 0.0400 0.372 33.87 36.79 1.086x
17
+ 20 0.0196 0.387 101.09 109.08 1.079x
18
+ 139 0.0173 0.633 29.52 32.71 1.108x
19
+ 189 0.0215 0.572 40.48 40.41 0.998x
20
+
21
+ interpolation regret vs per-plant DP: median +4.2% p95 +16.3% max +19.3%
22
+ (for scale: the hand-designed two-stage architecture is +16% median, +47% worst)
23
+
24
+ Reading: if interpolation is within a few percent, a lookup table amortizes this family
25
+ and a generative model adds nothing HERE — the map is smooth. The Furuta certified
26
+ hand-off map is not smooth (a gain weighting fails between two that work), so the same
27
+ test there should behave very differently. That contrast, not the size of the plant
28
+ family, is the argument for a learned model.
results/pendulum_full/dp_interp4.txt ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === FAIR interpolation baseline: full 4-D table over (l, m, zeta, rho_u) ===
2
+ 4^4 = 256 DP solves in 620 s, reused for every plant
3
+ plant J_ownDP J_interp4 ratio
4
+ 55 57.13 56.95 0.997x
5
+ 207 57.30 57.15 0.997x
6
+ 190 70.06 75.64 1.080x
7
+ 236 22.66 27.02 1.192x
8
+ 42 33.87 37.54 1.108x
9
+ 20 101.09 100.40 0.993x
10
+ 139 29.52 29.61 1.003x
11
+ 189 40.48 40.46 0.999x
12
+
13
+ 4-D interpolation regret vs per-plant DP: median +0.1% p95 +16.3% max +19.2%
14
+ for scale: 2-D table (wrong parameterisation) +4.2% med / +19.3% max
15
+ hand-designed two-stage architecture +16% med / +47% max
16
+
17
+ This is the number a generated feedback map has to beat. Note the table cost grows as
18
+ npts^4 (256 solves here); a learned model needs a comparable number of training
19
+ solves but no grid, which is the real argument for amortisation over tabulation.
results/pendulum_full/dp_interp_kappa.txt ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === does interpolation fail specifically at LOW kappa (multi-swing)? ===
2
+ kappa-stratified sample, n=40, 4-D table (256 DP solves) as the baseline
3
+ kappa zeta rho_u J_ownDP J_interp ratio
4
+ 1.65 0.1420 0.468 69.42 66.64 0.960x
5
+ 1.96 0.0991 0.388 64.14 79.69 1.242x
6
+ 1.72 0.1060 0.364 109.63 115.87 1.057x (FELL OVER)
7
+ 1.92 0.1429 0.549 34.12 34.49 1.011x
8
+ 2.56 0.1060 0.542 23.65 24.68 1.043x
9
+ 2.52 0.1281 0.646 49.02 51.28 1.046x
10
+ 2.80 0.0931 0.521 16.13 17.05 1.057x
11
+ 2.14 0.1428 0.611 25.81 26.43 1.024x
12
+ 2.82 0.0688 0.388 51.33 52.02 1.013x
13
+ 4.11 0.0559 0.460 34.43 34.17 0.992x
14
+ 3.42 0.0647 0.443 26.22 26.32 1.004x
15
+ 3.19 0.0829 0.529 85.73 86.46 1.008x
16
+ 3.54 0.0652 0.462 31.02 31.06 1.001x
17
+ 3.68 0.0715 0.527 35.10 35.51 1.012x
18
+ 4.73 0.0389 0.368 20.77 24.88 1.198x
19
+ 5.60 0.0562 0.629 67.07 70.33 1.049x
20
+ 5.59 0.0331 0.371 47.45 47.61 1.003x
21
+ 5.71 0.0442 0.505 41.56 41.58 1.001x
22
+ 4.79 0.0381 0.365 54.08 54.58 1.009x
23
+ 7.16 0.0248 0.356 19.87 20.15 1.014x
24
+ 7.90 0.0301 0.476 45.58 45.57 1.000x
25
+ 7.72 0.0342 0.528 51.97 51.90 0.999x
26
+ 8.16 0.0337 0.550 37.03 37.90 1.024x
27
+ 7.14 0.0411 0.587 31.26 31.66 1.013x
28
+ 12.30 0.0264 0.649 83.20 82.87 0.996x
29
+ 11.82 0.0255 0.603 36.63 36.60 0.999x
30
+ 12.89 0.0235 0.606 41.91 42.25 1.008x
31
+ 9.53 0.0214 0.408 58.48 58.38 0.998x
32
+ 10.73 0.0185 0.397 25.14 25.48 1.014x
33
+ 17.52 0.0105 0.367 46.72 46.70 1.000x
34
+ 14.18 0.0222 0.629 104.14 103.26 0.992x
35
+ 18.52 0.0139 0.516 31.70 34.99 1.104x
36
+ 15.75 0.0200 0.630 20.79 20.67 0.994x
37
+ 13.58 0.0217 0.589 40.05 40.05 1.000x
38
+ 19.78 0.0164 0.647 75.17 75.86 1.009x
39
+ 26.83 0.0109 0.585 16.95 16.98 1.001x
40
+ 26.33 0.0102 0.537 106.99 105.82 0.989x
41
+ 29.19 0.0100 0.585 20.81 20.91 1.005x
42
+ 21.07 0.0118 0.495 32.34 32.36 1.001x
43
+
44
+ kappa in [ 1.3, 2.0): n= 4 median regret +3.4% max +24.2%
45
+ kappa in [ 2.0, 3.0): n= 5 median regret +4.3% max +5.7%
46
+ kappa in [ 3.0, 4.5): n= 5 median regret +0.4% max +1.2%
47
+ kappa in [ 4.5, 6.5): n= 5 median regret +0.9% max +19.8%
48
+ kappa in [ 6.5, 9.0): n= 5 median regret +1.3% max +2.4%
49
+ kappa in [ 9.0, 13.0): n= 5 median regret -0.1% max +1.4%
50
+ kappa in [ 13.0, 19.0): n= 5 median regret -0.0% max +10.4%
51
+ kappa in [ 19.0, 30.0): n= 5 median regret +0.1% max +0.9%
52
+
53
+ kappa < 5 : n=16 median +1.3% max +24.2%
54
+ kappa >= 9 : n=15 median -0.0% max +10.4%
55
+
56
+ If the low-kappa band really is where tabulation breaks down, THAT is the target for a
57
+ generated policy — and it is the same non-smooth-map failure as the Furuta gain holes.
results/pendulum_full/dp_target_eval.txt ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ === DP-target eval, matched budget (N=40, 40 with a valid VI ref) ===
2
+ ceiling: value iteration, 4.20 s/plant
3
+
4
+ --- K = 1 candidate(s) per plant, identical for every arm ---
5
+ generated (K=1) med +2.03% p90 +11.78% max +24.83% >10%: 15.4% fail 2.5% [33 ms]
6
+ grid table +verifier (K=1) med +8.25% p90 +24.57% max +49.17% >10%: 40.0% fail 25.0%
7
+ nearest neighbour (K=1) med +0.20% p90 +2.55% max +24.53% >10%: 2.6% fail 2.5%
8
+ paired vs grid table generated better on 22, worse on 7 -> p = 0.0081 *
9
+ paired vs nearest nb generated better on 7, worse on 32 -> p = 0.0001 *
10
+
11
+ --- K = 4 candidate(s) per plant, identical for every arm ---
12
+ generated (K=4) med +0.87% p90 +9.97% max +15.43% >10%: 10.3% fail 2.5% [64 ms]
13
+ grid table +verifier (K=4) med +2.67% p90 +13.67% max +29.92% >10%: 14.3% fail 12.5%
14
+ nearest neighbour (K=4) med -0.20% p90 +0.46% max +19.46% >10%: 2.5% fail 0.0%
15
+ paired vs grid table generated better on 20, worse on 14 -> p = 0.3915
16
+ paired vs nearest nb generated better on 9, worse on 30 -> p = 0.0011 *
17
+
18
+ --- K = 8 candidate(s) per plant, identical for every arm ---
19
+ generated (K=8) med +0.34% p90 +2.52% max +23.42% >10%: 5.1% fail 2.5% [94 ms]
20
+ grid table +verifier (K=8) med +0.56% p90 +8.33% max +56.93% >10%: 10.0% fail 0.0%
21
+ nearest neighbour (K=8) med -1.17% p90 +0.26% max +0.90% >10%: 0.0% fail 0.0%
22
+ paired vs grid table generated better on 22, worse on 17 -> p = 0.5224
23
+ paired vs nearest nb generated better on 7, worse on 32 -> p = 0.0001 *
24
+
results/pendulum_full/family_sweep.txt ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ --- 1. analytic swing-up feasibility (kappa = rho_u/(2 zeta) > 1.3614) ---
2
+ zeta~logU[0.01,0.25], rho~U[.35,1.2] feasible 96.4% kappa med 7.31 zeta med 0.0501 frac kappa<3 (multi-swing) 22.3%
3
+ zeta~logU[0.02,0.30], rho~U[.35,0.9] feasible 88.6% kappa med 3.89 zeta med 0.0778 frac kappa<3 (multi-swing) 40.3%
4
+ zeta~logU[0.01,0.20], rho~U[.40,1.2] feasible 99.1% kappa med 8.52 zeta med 0.0449 frac kappa<3 (multi-swing) 15.1%
5
+ LEGACY b~U[0.02,0.3], rho=0.4 feasible 96.3% kappa med 5.83 zeta med 0.0343 frac kappa<3 (multi-swing) 22.0%
6
+ LEGACY b~U[0.02,0.3], rho=0.6 feasible 99.1% kappa med 8.75 zeta med 0.0343 frac kappa<3 (multi-swing) 10.0%
7
+
8
+ --- 2. chosen family: zeta~logU(0.01, 0.15), rho_u~U(0.35, 0.65), N=240 ---
9
+ analytically feasible: 99.6%
10
+ kappa min 1.31 p5 1.78 med 6.46 p95 20.97 max 28.09
11
+ rho_u min 0.352 p5 0.365 med 0.489 p95 0.629 max 0.649
12
+ zeta min 0.0101 p5 0.0117 med 0.0368 p95 0.1329 max 0.1458
13
+
14
+ --- 3. certified saturation-free region (max-gamma SDP) ---
15
+ plants with a non-empty certified region: 100.0%
16
+ gamma (best over alpha_hat) min 0.125 p5 0.138 med 0.187 p95 0.254 max 0.278
17
+ gamma / rho_u min 0.329 p5 0.335 med 0.394 p95 0.435 max 0.440
18
+ gamma by alpha_hat (median): a=1.2:0.187 a=1.8:0.124 a=2.5:0.090 a=3.5:0.064 a=5.0:0.045 a=7.0:0.032 a=10.0:0.022
19
+
20
+ --- 4. end-to-end keep rate (certified AND full-cycle AND never saturating) ---
21
+ kept: 98.8% (237/240)
22
+ of the analytically-feasible ones: 99.2%
23
+ J = int(q)+int(u^2) (kept) min 12.31 p5 17.60 med 42.40 p95 149.83 max 350.17
24
+ t_settle (kept) min 2.320 p5 2.569 med 3.905 p95 6.704 max 10.385
25
+
26
+ selected by ABSOLUTE-WEIGHT COST J (the §4.7 convention):
27
+ alpha_hat: 1.2:24 1.8:144 2.5:54 3.5:14 5:1 7:0 10:0
28
+ ke_mult : 5:0 10:35 20:73 30:56 60:55 120:18 300:0 1000:0
29
+ catch box: 0.03:0 0.05:8 0.08:33 0.12:81 0.18:97 0.25:18
30
+ selected by SCALE-INVARIANT t_settle (contrast):
31
+ alpha_hat: 1.2:105 1.8:16 2.5:14 3.5:25 5:57 7:20 10:0
32
+ ke_mult : 5:0 10:0 20:0 30:0 60:0 120:0 300:9 1000:228
33
+ catch box: 0.03:194 0.05:11 0.08:0 0.12:5 0.18:17 0.25:10
34
+ the two rules pick the SAME config on 0.0% of kept plants
35
+
36
+ --- 5. NUISANCE BASELINE A: the BEST single fixed controller, no per-plant solve ---
37
+ tuned over the same grid -> alpha_hat=1.8, ke_mult=120, catch box (0.08, 0.08*wn)
38
+ full-cycle success without saturating: 98.8%
39
+ AND handing off inside the certified set: 98.8% (vs 98.8% for the per-plant pipeline)
40
+ COST regret vs per-plant % min +0.0 p5 +0.0 med +0.3 p95 +1.1 max +2.2
41
+ fraction of kept plants where the fixed config IS the per-plant argmin: 0.0%
42
+ best 5 fixed configs by certified success:
43
+ a=1.8 ke=120 box=0.08 success 98.8% certified 98.8%
44
+ a=1.2 ke=120 box=0.08 success 98.8% certified 98.8%
45
+ a=1.2 ke=120 box=0.12 success 98.8% certified 98.8%
46
+ a=2.5 ke=300 box=0.03 success 98.8% certified 98.8%
47
+ a=1.8 ke=300 box=0.03 success 98.8% certified 98.8%
48
+ -> a tie on success AND certification AND cost means nothing is being amortized.
49
+
50
+ --- 6. NUISANCE BASELINE B: static sector spline alone, launched from hanging ---
51
+ (no swing-up stage; predicted to succeed once rho_u > 2/pi = 0.637)
52
+ rho_u in [0.35,0.50): static-only success 0.0% (n=126)
53
+ rho_u in [0.50,0.65): static-only success 0.0% (n=114)
54
+
55
+ --- 7. spline grid width: how far does the loop actually travel in y? ---
56
+ max|y| during balance min 0.0001 p5 0.0014 med 0.0100 p95 0.0528 max 0.2296
57
+ certified radius r min 0.0352 p5 0.0450 med 0.1396 p95 0.4406 max 0.5412
58
+ grid spans [-0.60,0.60], spacing 0.120
59
+ a grid whose spacing exceeds the travelled |y| leaves most control points
60
+ unexercised on the certified set (they then matter only OUTSIDE it).
results/refine_reference/refine_reference.png ADDED
results/timing/classical_mimo_cpu1.json ADDED
@@ -0,0 +1,253 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "machine": {
3
+ "host": "virgo-asl",
4
+ "python": "3.13.13",
5
+ "torch": "2.10.0+cu128",
6
+ "numpy": "2.2.6",
7
+ "scipy": "1.18.0",
8
+ "cvxpy": "1.9.2",
9
+ "diffusers": "0.38.0",
10
+ "threads": 1,
11
+ "tag": "cpu1",
12
+ "load1": 17.33642578125,
13
+ "date": "2026-09-30 07:20:14 PDT",
14
+ "cpu": "AMD Ryzen Threadripper 7960X 24-Cores",
15
+ "affinity": [
16
+ 21
17
+ ]
18
+ },
19
+ "family": "mimo",
20
+ "cells": {
21
+ "lmi/circle_beta_iters5/b1": {
22
+ "n": 200,
23
+ "median_ms": 68.87386087328196,
24
+ "p99_ms": 3816.0280430503167,
25
+ "mean_ms": 546.3023953093216,
26
+ "min_ms": 36.024462431669235,
27
+ "max_ms": 4755.798971280456,
28
+ "note": "certified_beta(bmax=8, iters=5): bisection of the decentralised circle-criterion LMI; controller = linear gain 0.3*(0.7*beta) (the oracle's own init) => synthesis+certificate, per plant",
29
+ "samples_ms": [
30
+ 1539.46105,
31
+ 84.89908,
32
+ 58.99341,
33
+ 158.97685,
34
+ 57.06773,
35
+ 46.85211,
36
+ 52.05384,
37
+ 2038.34223,
38
+ 317.71592,
39
+ 46.14167,
40
+ 47.39384,
41
+ 46.71565,
42
+ 86.24514,
43
+ 2685.7059,
44
+ 3813.63335,
45
+ 372.45745,
46
+ 55.60401,
47
+ 89.29427,
48
+ 53.32176,
49
+ 711.88774,
50
+ 3314.75986,
51
+ 55.28606,
52
+ 53.7078,
53
+ 46.19153,
54
+ 66.45013,
55
+ 1330.51316,
56
+ 56.31458,
57
+ 61.06931,
58
+ 2636.02433,
59
+ 90.99745,
60
+ 42.93542,
61
+ 44.30927,
62
+ 1233.37236,
63
+ 49.36498,
64
+ 852.99159,
65
+ 102.28822,
66
+ 2655.41108,
67
+ 54.04861,
68
+ 82.58464,
69
+ 61.15766,
70
+ 1476.64616,
71
+ 2740.09727,
72
+ 238.39509,
73
+ 49.03152,
74
+ 100.22615,
75
+ 52.01367,
76
+ 51.79844,
77
+ 2763.47313,
78
+ 273.39254,
79
+ 1506.32314,
80
+ 52.71514,
81
+ 84.49058,
82
+ 54.64163,
83
+ 195.18373,
84
+ 1010.33022,
85
+ 107.77339,
86
+ 162.07648,
87
+ 2714.82362,
88
+ 789.56922,
89
+ 70.2019,
90
+ 71.74916,
91
+ 50.32908,
92
+ 63.23676,
93
+ 54.94314,
94
+ 48.97369,
95
+ 63.71007,
96
+ 57.70152,
97
+ 46.84208,
98
+ 1641.44114,
99
+ 65.3701,
100
+ 1499.00568,
101
+ 1583.04308,
102
+ 155.7974,
103
+ 48.37973,
104
+ 57.47,
105
+ 2099.01559,
106
+ 44.31368,
107
+ 409.94296,
108
+ 61.88901,
109
+ 51.12013,
110
+ 46.062,
111
+ 70.53595,
112
+ 516.58963,
113
+ 45.07162,
114
+ 1932.52514,
115
+ 72.99564,
116
+ 980.11014,
117
+ 57.26145,
118
+ 63.32942,
119
+ 45.80935,
120
+ 70.64325,
121
+ 2655.0548,
122
+ 60.55124,
123
+ 1398.9475,
124
+ 54.07696,
125
+ 50.63864,
126
+ 1385.55271,
127
+ 4053.10305,
128
+ 62.20255,
129
+ 49.0091,
130
+ 47.3842,
131
+ 467.22038,
132
+ 50.84759,
133
+ 1313.331,
134
+ 172.51838,
135
+ 68.47264,
136
+ 55.15518,
137
+ 52.60684,
138
+ 41.62144,
139
+ 89.66991,
140
+ 47.79533,
141
+ 47.35295,
142
+ 58.36662,
143
+ 42.34562,
144
+ 3054.16087,
145
+ 1367.51997,
146
+ 57.07521,
147
+ 110.66511,
148
+ 60.31887,
149
+ 47.43358,
150
+ 56.15808,
151
+ 2782.96257,
152
+ 73.97002,
153
+ 572.89503,
154
+ 68.09734,
155
+ 168.34665,
156
+ 143.64995,
157
+ 91.76317,
158
+ 46.70815,
159
+ 2656.82335,
160
+ 44.81606,
161
+ 113.42168,
162
+ 59.43455,
163
+ 53.72118,
164
+ 533.67817,
165
+ 44.59412,
166
+ 44.90975,
167
+ 2548.15145,
168
+ 60.62354,
169
+ 2613.68835,
170
+ 41.80415,
171
+ 58.57468,
172
+ 3097.16299,
173
+ 47.00665,
174
+ 102.20869,
175
+ 55.85084,
176
+ 71.33744,
177
+ 47.38481,
178
+ 409.22454,
179
+ 1329.93732,
180
+ 1342.6898,
181
+ 65.32284,
182
+ 107.73836,
183
+ 59.92452,
184
+ 180.91818,
185
+ 2805.83807,
186
+ 66.97288,
187
+ 1850.27357,
188
+ 48.42656,
189
+ 62.97545,
190
+ 56.19333,
191
+ 52.17917,
192
+ 4755.79897,
193
+ 36.02446,
194
+ 40.4797,
195
+ 45.0965,
196
+ 43.91727,
197
+ 139.24127,
198
+ 42.6585,
199
+ 259.45502,
200
+ 588.53512,
201
+ 43.94901,
202
+ 77.34191,
203
+ 152.85974,
204
+ 38.38812,
205
+ 94.10778,
206
+ 56.41848,
207
+ 1435.67258,
208
+ 58.60595,
209
+ 182.33116,
210
+ 48.6597,
211
+ 50.30897,
212
+ 43.92796,
213
+ 570.07636,
214
+ 69.27508,
215
+ 47.14419,
216
+ 43.0231,
217
+ 2290.25402,
218
+ 55.12404,
219
+ 50.66497,
220
+ 62.87157,
221
+ 1245.61535,
222
+ 158.08528,
223
+ 69.43793,
224
+ 454.00061,
225
+ 69.28398,
226
+ 682.18592,
227
+ 1371.39834,
228
+ 49.64081,
229
+ 158.31147
230
+ ]
231
+ },
232
+ "lmi/circle_beta_iters5/batch120": {
233
+ "n": 2,
234
+ "median_ms": 540.0699076397965,
235
+ "p99_ms": 570.1727172235647,
236
+ "mean_ms": 540.0699076397965,
237
+ "min_ms": 509.35275500329834,
238
+ "max_ms": 570.7870602762947,
239
+ "note": "looped over batch=120 (cvxpy solves are per plant); ms per plant; n = batches",
240
+ "samples_ms": [
241
+ 570.78706,
242
+ 509.35276
243
+ ]
244
+ },
245
+ "lmi/quality_heldout120": {
246
+ "coverage": 1.0,
247
+ "regret_median_pct": 13.914860725402832,
248
+ "regret_p90_pct": 43.86730766296388,
249
+ "note": "linear decentralised gain at 0.3*cap under the circle-criterion-certified sector (sound certificate), vs stored eval_mimo_regret oracle"
250
+ }
251
+ },
252
+ "done": true
253
+ }
results/timing/classical_mimo_cpu1q.json ADDED
@@ -0,0 +1,253 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "machine": {
3
+ "host": "virgo-asl",
4
+ "python": "3.13.13",
5
+ "torch": "2.10.0+cu128",
6
+ "numpy": "2.2.6",
7
+ "scipy": "1.18.0",
8
+ "cvxpy": "1.9.2",
9
+ "diffusers": "0.38.0",
10
+ "threads": 1,
11
+ "tag": "cpu1q",
12
+ "load1": 8.916015625,
13
+ "date": "2026-09-30 09:11:48 PDT",
14
+ "cpu": "AMD Ryzen Threadripper 7960X 24-Cores",
15
+ "affinity": [
16
+ 20
17
+ ]
18
+ },
19
+ "family": "mimo",
20
+ "cells": {
21
+ "lmi/circle_beta_iters5/b1": {
22
+ "n": 200,
23
+ "median_ms": 66.19771895930171,
24
+ "p99_ms": 3756.5721647534497,
25
+ "mean_ms": 536.2993241753429,
26
+ "min_ms": 36.544194445014,
27
+ "max_ms": 4701.850468292832,
28
+ "note": "certified_beta(bmax=8, iters=5): bisection of the decentralised circle-criterion LMI; controller = linear gain 0.3*(0.7*beta) (the oracle's own init) => synthesis+certificate, per plant",
29
+ "samples_ms": [
30
+ 1513.84927,
31
+ 79.18051,
32
+ 52.53494,
33
+ 154.0505,
34
+ 50.48537,
35
+ 43.06555,
36
+ 47.73189,
37
+ 1990.77106,
38
+ 308.27518,
39
+ 39.86873,
40
+ 39.18953,
41
+ 39.90704,
42
+ 80.71017,
43
+ 2640.85822,
44
+ 3754.09782,
45
+ 362.3173,
46
+ 50.45921,
47
+ 85.18535,
48
+ 48.75933,
49
+ 703.31981,
50
+ 3278.91677,
51
+ 48.70152,
52
+ 48.8602,
53
+ 40.56209,
54
+ 60.90773,
55
+ 1314.93117,
56
+ 53.80343,
57
+ 56.41611,
58
+ 2593.9615,
59
+ 85.1521,
60
+ 37.23508,
61
+ 38.32452,
62
+ 1214.554,
63
+ 45.87818,
64
+ 841.45585,
65
+ 100.99499,
66
+ 2616.62442,
67
+ 48.40769,
68
+ 76.61527,
69
+ 54.58448,
70
+ 1444.48981,
71
+ 2700.64529,
72
+ 230.40055,
73
+ 43.38491,
74
+ 97.00003,
75
+ 47.4023,
76
+ 47.15563,
77
+ 2698.40246,
78
+ 266.67435,
79
+ 1473.26395,
80
+ 47.04946,
81
+ 73.90279,
82
+ 51.1841,
83
+ 189.5483,
84
+ 989.51642,
85
+ 102.48008,
86
+ 159.59952,
87
+ 2656.26725,
88
+ 773.19128,
89
+ 64.42876,
90
+ 67.6078,
91
+ 45.53384,
92
+ 56.61694,
93
+ 50.95768,
94
+ 44.0698,
95
+ 57.52367,
96
+ 53.23725,
97
+ 41.41171,
98
+ 1607.05642,
99
+ 61.50824,
100
+ 1490.06605,
101
+ 1558.93327,
102
+ 151.64866,
103
+ 44.4635,
104
+ 53.02526,
105
+ 2068.91261,
106
+ 40.47347,
107
+ 397.91506,
108
+ 58.21347,
109
+ 46.31337,
110
+ 40.89776,
111
+ 66.87334,
112
+ 510.13955,
113
+ 43.84304,
114
+ 1900.14085,
115
+ 68.75502,
116
+ 955.51208,
117
+ 53.99163,
118
+ 58.21375,
119
+ 42.01659,
120
+ 66.53192,
121
+ 2616.54124,
122
+ 60.78525,
123
+ 1365.12419,
124
+ 53.3804,
125
+ 49.62038,
126
+ 1364.8067,
127
+ 4001.53255,
128
+ 61.37351,
129
+ 47.13935,
130
+ 44.06032,
131
+ 458.43294,
132
+ 47.19989,
133
+ 1296.0189,
134
+ 176.05968,
135
+ 66.27581,
136
+ 55.15212,
137
+ 47.63964,
138
+ 37.87151,
139
+ 86.97425,
140
+ 45.56593,
141
+ 42.92863,
142
+ 55.09171,
143
+ 39.88255,
144
+ 3009.76388,
145
+ 1341.09018,
146
+ 55.08479,
147
+ 107.95607,
148
+ 57.45166,
149
+ 40.81644,
150
+ 54.03973,
151
+ 2751.96685,
152
+ 71.72596,
153
+ 561.86074,
154
+ 66.16003,
155
+ 164.70586,
156
+ 140.85416,
157
+ 87.79745,
158
+ 40.03635,
159
+ 2612.99797,
160
+ 42.57246,
161
+ 110.65331,
162
+ 57.25475,
163
+ 51.48026,
164
+ 506.16988,
165
+ 41.54455,
166
+ 41.59874,
167
+ 2522.02333,
168
+ 59.40518,
169
+ 2614.56268,
170
+ 38.08357,
171
+ 50.6474,
172
+ 3063.74876,
173
+ 48.43063,
174
+ 96.17522,
175
+ 57.26778,
176
+ 70.33154,
177
+ 46.36715,
178
+ 404.33629,
179
+ 1317.2864,
180
+ 1322.3159,
181
+ 63.02168,
182
+ 102.9791,
183
+ 57.52541,
184
+ 176.35879,
185
+ 2769.55748,
186
+ 64.22492,
187
+ 1817.0326,
188
+ 48.23126,
189
+ 60.74989,
190
+ 53.38415,
191
+ 50.49312,
192
+ 4701.85047,
193
+ 36.54419,
194
+ 40.09773,
195
+ 44.22369,
196
+ 42.31991,
197
+ 136.46588,
198
+ 40.87392,
199
+ 255.31355,
200
+ 584.29782,
201
+ 44.94602,
202
+ 75.77784,
203
+ 148.22984,
204
+ 38.3939,
205
+ 91.66273,
206
+ 56.50123,
207
+ 1425.14511,
208
+ 58.22189,
209
+ 181.3864,
210
+ 46.81972,
211
+ 48.84703,
212
+ 41.58017,
213
+ 558.48009,
214
+ 69.34392,
215
+ 44.08439,
216
+ 43.00825,
217
+ 2259.23838,
218
+ 53.32712,
219
+ 47.9518,
220
+ 59.84782,
221
+ 1227.37103,
222
+ 154.69733,
223
+ 67.36043,
224
+ 444.89144,
225
+ 66.23541,
226
+ 668.46963,
227
+ 1351.32746,
228
+ 47.17544,
229
+ 148.08138
230
+ ]
231
+ },
232
+ "lmi/circle_beta_iters5/batch120": {
233
+ "n": 2,
234
+ "median_ms": 479.01488344650716,
235
+ "p99_ms": 557.0329894960548,
236
+ "mean_ms": 479.01488344650716,
237
+ "min_ms": 399.40457115105045,
238
+ "max_ms": 558.6251957419639,
239
+ "note": "looped over batch=120 (cvxpy solves are per plant); ms per plant; n = batches",
240
+ "samples_ms": [
241
+ 558.6252,
242
+ 399.40457
243
+ ]
244
+ },
245
+ "lmi/quality_heldout120": {
246
+ "coverage": 1.0,
247
+ "regret_median_pct": 13.914860725402832,
248
+ "regret_p90_pct": 43.86730766296388,
249
+ "note": "linear decentralised gain at 0.3*cap under the circle-criterion-certified sector (sound certificate), vs stored eval_mimo_regret oracle"
250
+ }
251
+ },
252
+ "done": true
253
+ }
results/timing/classical_mimo_cpu8.json ADDED
@@ -0,0 +1,260 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "machine": {
3
+ "host": "virgo-asl",
4
+ "python": "3.13.13",
5
+ "torch": "2.10.0+cu128",
6
+ "numpy": "2.2.6",
7
+ "scipy": "1.18.0",
8
+ "cvxpy": "1.9.2",
9
+ "diffusers": "0.38.0",
10
+ "threads": 8,
11
+ "tag": "cpu8",
12
+ "load1": 23.9970703125,
13
+ "date": "2026-09-30 07:54:54 PDT",
14
+ "cpu": "AMD Ryzen Threadripper 7960X 24-Cores",
15
+ "affinity": [
16
+ 14,
17
+ 15,
18
+ 17,
19
+ 18,
20
+ 19,
21
+ 20,
22
+ 22,
23
+ 23
24
+ ]
25
+ },
26
+ "family": "mimo",
27
+ "cells": {
28
+ "lmi/circle_beta_iters5/b1": {
29
+ "n": 200,
30
+ "median_ms": 72.32201099395752,
31
+ "p99_ms": 3823.6898711882513,
32
+ "mean_ms": 549.2923274170607,
33
+ "min_ms": 36.78356297314167,
34
+ "max_ms": 4809.1598181054,
35
+ "note": "certified_beta(bmax=8, iters=5): bisection of the decentralised circle-criterion LMI; controller = linear gain 0.3*(0.7*beta) (the oracle's own init) => synthesis+certificate, per plant",
36
+ "samples_ms": [
37
+ 1527.77675,
38
+ 78.22582,
39
+ 51.39571,
40
+ 155.94155,
41
+ 51.53694,
42
+ 48.72184,
43
+ 53.77331,
44
+ 2011.14939,
45
+ 312.37963,
46
+ 41.20297,
47
+ 44.20665,
48
+ 46.50073,
49
+ 89.90167,
50
+ 2670.06131,
51
+ 3821.51445,
52
+ 378.58834,
53
+ 60.86336,
54
+ 99.62023,
55
+ 58.56525,
56
+ 727.18937,
57
+ 3344.48929,
58
+ 63.00118,
59
+ 60.56186,
60
+ 51.17881,
61
+ 73.37691,
62
+ 1338.93066,
63
+ 62.08986,
64
+ 66.98864,
65
+ 2647.59406,
66
+ 88.45621,
67
+ 37.99611,
68
+ 36.96601,
69
+ 1232.83921,
70
+ 57.47533,
71
+ 854.23513,
72
+ 110.16709,
73
+ 2649.42927,
74
+ 51.12484,
75
+ 72.32689,
76
+ 54.27222,
77
+ 1462.28117,
78
+ 2739.18881,
79
+ 251.10868,
80
+ 63.34201,
81
+ 115.7227,
82
+ 65.93608,
83
+ 65.77918,
84
+ 2737.63095,
85
+ 290.35823,
86
+ 1514.26963,
87
+ 67.85548,
88
+ 86.52914,
89
+ 62.82481,
90
+ 204.84217,
91
+ 1016.16613,
92
+ 117.729,
93
+ 176.0141,
94
+ 2726.83456,
95
+ 818.55577,
96
+ 84.46952,
97
+ 87.13758,
98
+ 65.15015,
99
+ 83.34302,
100
+ 62.24988,
101
+ 54.73903,
102
+ 67.388,
103
+ 62.95666,
104
+ 52.93454,
105
+ 1631.87829,
106
+ 69.88871,
107
+ 1507.96469,
108
+ 1608.04162,
109
+ 165.65432,
110
+ 57.12839,
111
+ 64.94526,
112
+ 2155.20359,
113
+ 60.23292,
114
+ 417.62554,
115
+ 72.31713,
116
+ 57.88626,
117
+ 50.91124,
118
+ 82.93688,
119
+ 535.6281,
120
+ 47.53878,
121
+ 1968.79024,
122
+ 95.56295,
123
+ 1013.76554,
124
+ 53.16884,
125
+ 54.09356,
126
+ 38.89047,
127
+ 63.42843,
128
+ 2660.12211,
129
+ 86.45663,
130
+ 1406.20906,
131
+ 75.55625,
132
+ 71.74694,
133
+ 1388.89967,
134
+ 4039.0562,
135
+ 71.4912,
136
+ 56.84959,
137
+ 55.11437,
138
+ 479.19634,
139
+ 62.62968,
140
+ 1373.05836,
141
+ 188.31666,
142
+ 104.92228,
143
+ 90.90466,
144
+ 58.57755,
145
+ 48.79924,
146
+ 97.79338,
147
+ 44.27852,
148
+ 41.38853,
149
+ 54.24409,
150
+ 37.54529,
151
+ 3093.92548,
152
+ 1354.13561,
153
+ 53.30585,
154
+ 106.70779,
155
+ 55.76267,
156
+ 37.69584,
157
+ 50.46429,
158
+ 2773.99922,
159
+ 70.60852,
160
+ 565.89937,
161
+ 63.99923,
162
+ 163.93369,
163
+ 143.35188,
164
+ 87.26837,
165
+ 37.56383,
166
+ 2628.61575,
167
+ 41.91269,
168
+ 108.21044,
169
+ 55.18098,
170
+ 50.32528,
171
+ 513.04181,
172
+ 39.77673,
173
+ 40.20145,
174
+ 2545.9515,
175
+ 57.029,
176
+ 2617.88251,
177
+ 37.43142,
178
+ 48.70009,
179
+ 3095.50921,
180
+ 46.63382,
181
+ 97.19735,
182
+ 55.36159,
183
+ 68.30094,
184
+ 44.91931,
185
+ 405.59209,
186
+ 1329.90252,
187
+ 1335.46317,
188
+ 60.15871,
189
+ 100.17693,
190
+ 56.598,
191
+ 174.14835,
192
+ 2820.7193,
193
+ 62.5164,
194
+ 1833.50239,
195
+ 46.7516,
196
+ 62.33442,
197
+ 53.0722,
198
+ 50.74952,
199
+ 4809.15982,
200
+ 37.87016,
201
+ 40.12395,
202
+ 43.84106,
203
+ 45.63039,
204
+ 141.18765,
205
+ 39.07686,
206
+ 256.61544,
207
+ 596.76008,
208
+ 47.17568,
209
+ 76.84058,
210
+ 150.91957,
211
+ 36.78356,
212
+ 89.76167,
213
+ 54.36926,
214
+ 1448.62717,
215
+ 59.61437,
216
+ 181.54364,
217
+ 44.4173,
218
+ 46.29237,
219
+ 38.50612,
220
+ 559.32421,
221
+ 66.76084,
222
+ 40.65791,
223
+ 39.96966,
224
+ 2289.57182,
225
+ 51.63352,
226
+ 44.79614,
227
+ 57.01578,
228
+ 1249.04927,
229
+ 152.37869,
230
+ 64.89577,
231
+ 453.71979,
232
+ 65.1298,
233
+ 679.60313,
234
+ 1370.48742,
235
+ 43.84292,
236
+ 149.43281
237
+ ]
238
+ },
239
+ "lmi/circle_beta_iters5/batch120": {
240
+ "n": 2,
241
+ "median_ms": 492.2846951871179,
242
+ "p99_ms": 569.9142948704928,
243
+ "mean_ms": 492.2846951871179,
244
+ "min_ms": 413.07081795918447,
245
+ "max_ms": 571.4985724150514,
246
+ "note": "looped over batch=120 (cvxpy solves are per plant); ms per plant; n = batches",
247
+ "samples_ms": [
248
+ 571.49857,
249
+ 413.07082
250
+ ]
251
+ },
252
+ "lmi/quality_heldout120": {
253
+ "coverage": 1.0,
254
+ "regret_median_pct": 13.914860725402832,
255
+ "regret_p90_pct": 43.86730766296388,
256
+ "note": "linear decentralised gain at 0.3*cap under the circle-criterion-certified sector (sound certificate), vs stored eval_mimo_regret oracle"
257
+ }
258
+ },
259
+ "done": true
260
+ }