matCercola18 commited on
Commit
cad6165
·
verified ·
1 Parent(s): 164a408

X28-M: S4 with repair (July sector projection)

Browse files
Files changed (35) hide show
  1. results/icml_r3/X28/PREREG.md +25 -0
  2. results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m2b_s0.npz +3 -0
  3. results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m2b_s1.npz +3 -0
  4. results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m2b_s2.npz +3 -0
  5. results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m3b_s0.npz +3 -0
  6. results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m3b_s1.npz +3 -0
  7. results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m3b_s2.npz +3 -0
  8. results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m2b_s0.npz +3 -0
  9. results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m2b_s1.npz +3 -0
  10. results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m2b_s2.npz +3 -0
  11. results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m3b_s0.npz +3 -0
  12. results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m3b_s1.npz +3 -0
  13. results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m3b_s2.npz +3 -0
  14. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2a_s0.npz +3 -0
  15. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2a_s1.npz +3 -0
  16. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2a_s2.npz +3 -0
  17. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2b_s0.npz +3 -0
  18. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2b_s1.npz +3 -0
  19. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2b_s2.npz +3 -0
  20. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3a_s0.npz +3 -0
  21. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3a_s1.npz +3 -0
  22. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3a_s2.npz +3 -0
  23. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3b_s0.npz +3 -0
  24. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3b_s1.npz +3 -0
  25. results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3b_s2.npz +3 -0
  26. results/icml_r3/X28/mimo/gen_rep_fresh_m2.npz +3 -0
  27. results/icml_r3/X28/mimo/gen_rep_fresh_m2.txt +7 -0
  28. results/icml_r3/X28/mimo/gen_rep_fresh_m3.npz +3 -0
  29. results/icml_r3/X28/mimo/gen_rep_fresh_m3.txt +7 -0
  30. results/icml_r3/X28/mimo/gen_rep_paired_m2.txt +17 -0
  31. results/icml_r3/X28/mimo/gen_rep_paired_m3.txt +17 -0
  32. results/icml_r3/X28/mimo/rep_run.txt +99 -0
  33. results/icml_r3/X28/mimo/rep_selfcheck.txt +6 -0
  34. results/icml_r3/X28/mimo/rep_tails.txt +25 -0
  35. results/icml_r3/X28/mimo/rep_val_choice.txt +21 -0
results/icml_r3/X28/PREREG.md CHANGED
@@ -224,3 +224,28 @@ Compute: inside GPL_move's job 17025 on node 03 (Matteo 2026-11-13: share that s
224
 
225
  Written before any MIMO label run, dataset or model exists, after a round-trip check of the code on the study's validation labels. The gauge of the previous addendum (K row i times r_i(1), ratios divided by r_i(1)) is not a symmetry of the so3 class: the knots sit at fixed positions, and a row scaling c is only matched by psi_i(w / c), whose knots move to c x_k. On the 64 validation labels the round trip changed J by up to 72% (m = 2) and 225% (m = 3); without it the round trip is exact (3e-16). The gauge starts c = 1, 2, 5 of the optimizer stay valid, because they start from a linear psi, for which the scaling is exact.
226
  Target of arm (b), replacing the previous one: K (m^2) and log r_i(1), log r_i(3), log r_i(5) per channel (3m), each ratio projected to [0.1, 10] after decoding. Nothing else changes.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
224
 
225
  Written before any MIMO label run, dataset or model exists, after a round-trip check of the code on the study's validation labels. The gauge of the previous addendum (K row i times r_i(1), ratios divided by r_i(1)) is not a symmetry of the so3 class: the knots sit at fixed positions, and a row scaling c is only matched by psi_i(w / c), whose knots move to c x_k. On the 64 validation labels the round trip changed J by up to 72% (m = 2) and 225% (m = 3); without it the round trip is exact (3e-16). The gauge starts c = 1, 2, 5 of the optimizer stay valid, because they start from a linear psi, for which the scaling is exact.
226
  Target of arm (b), replacing the previous one: K (m^2) and log r_i(1), log r_i(3), log r_i(5) per channel (3m), each ratio projected to [0.1, 10] after decoding. Nothing else changes.
227
+
228
+ ## Addendum 2026-11-14: MIMO spline repair (the July sector projection)
229
+
230
+ Written after the S4-M result (`gen_paired_m*.txt`, `gen_fresh_m*.txt`) and before any repair code, validation sample or repair number exists. The S4-M verdict stays the registered primary result of X28; this test is reported next to it as "S4-M with repair".
231
+
232
+ **Why** (Matteo, 2026-11-14: "Are you sure that you tested at the best the July idea?", then "Yes, make it work"). In July (`nonlinear/data_generation/optimize_mimo_targets.py`) the plants were stable and every spline was projected into the sector [0, 0.7 beta] of its plant, so every sample was certified by construction. The unstable X27 family has no sector known in advance, so S4-M certified each sample after sampling, and that is where arm (b) lost: 57% to 76% of its samples certified against 93% to 99.5% for arm (a). Two gaps against July: no projection step for (b), and a circle certificate that is conservative for a static time-invariant nonlinearity, while arm (a) gets the exact Lyapunov test.
233
+
234
+ **Repair R (arm (b) only; arm (a) is linear and unchanged).** For a decoded sample (K, r_i(1), r_i(3), r_i(5)), shrink the log ratios of each channel toward a center g_i: log r_i(x; lam) = log g_i + lam (log r_i(x) - log g_i) for x = 1, 3, 5. Rebuild the so3 ratios (r(2) = (r(1) + 3 r(3)) / 4, r(4) = (3 r(3) + 5 r(5)) / 8, odd), keep K, and keep the largest lam in {1, 0.75, 0.5, 0.25, 0} that certifies. The sectors are nested and shrink to the point g_i. lam = 0 is the linear controller u = -diag(g) K y, which gets the Lyapunov test. One repaired controller per sample, with no cost-based choice among the lam values, so the budget b still counts generated samples. A sample with no certified lam stays uncertified.
235
+
236
+ **Certificate of arm (b) (primary).** A sample is certified iff the circle LMI of section 2.1 or the multivariable Popov LMI holds. Popov: V = x'Px + 2 sum_i q_i int_0^{w_i} phi'_i(s) ds, with w = K C x, phi'_i(w) = psi_i(w) - alpha_i w in [0, beta_i - alpha_i] and Q = diag(q) >= 0. It is the LMI of section 2.1 with M12 = C'K' diag(beta - alpha) T - P B + A_K' C'K' Q and M22 = -2T - Q K C B - B'C'K' Q (Q = 0 gives the circle LMI). It holds for static time-invariant nonlinearities, which these knots are. Same solver, eps and numerical re-check, with q clipped at 0 before the re-check.
237
+
238
+ Self-checks before the validation run, on arm (b) samples of the validation plants:
239
+ 1. On 400 random circle-certified samples, the Popov LMI alone must be feasible on at least 95%, since Q = 0 is a feasible point. Below that, the Popov code counts as faulty.
240
+ 2. The first 200 Popov-certified, circle-rejected samples are simulated for 60 s from 64 random unit directions at radii 1 and 5 (seed (7878, plant id)). Any rollout with final |x| >= |x0| or a non-finite state counts as a failure. The fraction with final |x| > 1e-2 |x0| is reported, next to the same numbers for 200 circle-certified samples.
241
+ 3. If check 1 or check 2 fails, the primary drops the Popov LMI and uses the circle LMI with repair.
242
+
243
+ **Center g (chosen on validation).** Two candidates: the geometric mean of r_i(1), r_i(3), r_i(5), or the small-signal ratio r_i(1) (the analogue of the SISO t-repair).
244
+ - Validation set: the 64 bank plants [0, 64) per m, never used to train the generators. 16 samples per plant from all 12 generators (seed = model seed, CPU sampler). J*_val = min over the class-study labels of those plants (`study_m*/cert_*.npz`).
245
+ - Rule: pooled over the 3 seeds and both m, at b = 16, take the center with the lower median regret (uncertified = 1e3) among those whose certified rate is at most 2 points below arm (a)'s on the same plants. If neither qualifies, take the one with the higher certified rate. Ties go to the geometric mean.
246
+
247
+ **Test.** Run once on the 120 held-out plants per m, with the stored S4-M samples (`gen/cand_*.npz`) and arm (a) results (`gen/eval_*.npz`) unchanged. The S4-M rule applies: b = 16 primary, b = 1 and 4 reported, J* unchanged, Holm over the 3 seed pairs, verdict iff on >= 2 of 3 pairs. The fresh-direction diagnostic is repeated for the repaired arm.
248
+ Secondaries, reported with the primary: (i) the either-or certificate without repair (lam = 1 only); (ii) repair with the circle LMI alone. Also reported: the distribution of lam and the certified rate per sample.
249
+ Outputs: `mimo/gen_rep/`, `mimo/gen_rep_paired_m{2,3}.txt`, `mimo/gen_rep_fresh_m{2,3}.txt`, `mimo/rep_selfcheck.txt`, `mimo/rep_val_choice.txt`.
250
+
251
+ Compute: the CPU node `move-hpc-01-cpu` (back up), in one process, because the /home NFS is slow to import from.
results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m2b_s0.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:03de6d0ab3b2969d56925cb8acb58353f3a6451ad009a22ad34f4b45a3e0841f
3
+ size 243352
results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m2b_s1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6de5000bf131e69c821f158d574c239ecf726d195b426bd4760129de2b9c2c3
3
+ size 243506
results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m2b_s2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b234d4e3d18dd1ca83082bdb2181e96ae465cfa83fcfbf9610538c3d302fb9b
3
+ size 243060
results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m3b_s0.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fa0305ec410103d475ac8f19bd6fada07102829e4e73b8f625b65b22307fa285
3
+ size 401624
results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m3b_s1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:16530e30cbbc8e5b36957b7215e7e1a6d051ba8b8c0e9579d37c530885fb38c0
3
+ size 399985
results/icml_r3/X28/mimo/gen_rep/eval_circle_x28m3b_s2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b5708545cc65dc7b03bfa269475ed558300b912bb953ef120b3b2f8afd6c1d3
3
+ size 399911
results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m2b_s0.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:03de6d0ab3b2969d56925cb8acb58353f3a6451ad009a22ad34f4b45a3e0841f
3
+ size 243352
results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m2b_s1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6de5000bf131e69c821f158d574c239ecf726d195b426bd4760129de2b9c2c3
3
+ size 243506
results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m2b_s2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b234d4e3d18dd1ca83082bdb2181e96ae465cfa83fcfbf9610538c3d302fb9b
3
+ size 243060
results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m3b_s0.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fa0305ec410103d475ac8f19bd6fada07102829e4e73b8f625b65b22307fa285
3
+ size 401624
results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m3b_s1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:16530e30cbbc8e5b36957b7215e7e1a6d051ba8b8c0e9579d37c530885fb38c0
3
+ size 399985
results/icml_r3/X28/mimo/gen_rep/eval_primary_x28m3b_s2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b5708545cc65dc7b03bfa269475ed558300b912bb953ef120b3b2f8afd6c1d3
3
+ size 399911
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2a_s0.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:46f7198e2188ab1f5f59b3c527496d8198ee24e106c2b6317d73e4d4eaa056f4
3
+ size 31460
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2a_s1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5d53f822f1b5f52c4a1b0f252d75fd6992a2c47f4135916d0cf1346467ca073e
3
+ size 31483
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2a_s2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:68ed7a7a1bd848a5fd1236dd0f90795959f93a4d8a43064b6e1ae6900f699fac
3
+ size 31473
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2b_s0.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:223fa32f8a98b830c200382216672b7a5b7e8037ffa26e14232801cea4a959a9
3
+ size 78694
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2b_s1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d7a98b367e3e2a53443afcd50fdb2a8699697d6e9314bcfb1902057537a56bd8
3
+ size 78625
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m2b_s2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f83dc47217ae0efe185ea08c7c9642edd4d1980e27d2cb129433e5a0292ce0e4
3
+ size 78579
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3a_s0.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6c0c79ddd44b222ca68b0dccbb2de56f277d8e5a94f91759a113b0a0f16fd91
3
+ size 70802
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3a_s1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4821020cd720628d518a10106134597e9239955a87ab3442e86696f07d0ab310
3
+ size 70831
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3a_s2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d72697e93723b6f4882b81c2f4f8c6f97ae66391e6f6f8e1f54995ca647a4abe
3
+ size 70848
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3b_s0.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:20493c0789d1db7a8f8a5810a609c3ec13ba2e54e2e4e497545e7eec0845fe21
3
+ size 141153
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3b_s1.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ba22a46811be774678c4ad6866f76110b48bc1209761ec1b7e9a2046d97a99b2
3
+ size 141113
results/icml_r3/X28/mimo/gen_rep/val_cand_x28m3b_s2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e1131b295c961e4d026c2fedde520fb7735dc8f903ddb1c6f1baf8eb46618c2
3
+ size 141202
results/icml_r3/X28/mimo/gen_rep_fresh_m2.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:39d15c2942030cfda414d3231b13ad58fb08e31870837f112fb9af72e054ec76
3
+ size 707964
results/icml_r3/X28/mimo/gen_rep_fresh_m2.txt ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ X28-M S4 with repair, fresh directions, m = 2: best certified b=16 sample per held-out plant on 64 unit directions (seed (7575, id)), vs x28m2a_s0; (b) repaired (primary)
2
+ x28m2a_s0: median over plants of per-plant median (J_a0 - J) / J +0.00% [+0.00, +0.00], > 0 on 0 of 120; non-convergent 0 of 7680
3
+ x28m2a_s1: median over plants of per-plant median (J_a0 - J) / J -0.07% [-0.18, +0.01], > 0 on 50 of 120; non-convergent 35 of 7680
4
+ x28m2a_s2: median over plants of per-plant median (J_a0 - J) / J -0.07% [-0.16, +0.14], > 0 on 57 of 120; non-convergent 0 of 7680
5
+ x28m2b_s0: median over plants of per-plant median (J_a0 - J) / J +2.03% [+0.40, +3.08], > 0 on 77 of 120; non-convergent 123 of 7680
6
+ x28m2b_s1: median over plants of per-plant median (J_a0 - J) / J +1.69% [+0.51, +2.54], > 0 on 78 of 120; non-convergent 42 of 7680
7
+ x28m2b_s2: median over plants of per-plant median (J_a0 - J) / J +1.29% [+0.49, +2.41], > 0 on 75 of 120; non-convergent 66 of 7680
results/icml_r3/X28/mimo/gen_rep_fresh_m3.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05797c3766d83cb15e660f167feea8ebf8735b474d6a657cead2dd7f60ce48ae
3
+ size 704217
results/icml_r3/X28/mimo/gen_rep_fresh_m3.txt ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ X28-M S4 with repair, fresh directions, m = 3: best certified b=16 sample per held-out plant on 64 unit directions (seed (7575, id)), vs x28m3a_s0; (b) repaired (primary)
2
+ x28m3a_s0: median over plants of per-plant median (J_a0 - J) / J +0.00% [+0.00, +0.00], > 0 on 0 of 120; non-convergent 54 of 7680
3
+ x28m3a_s1: median over plants of per-plant median (J_a0 - J) / J +0.26% [-0.16, +0.61], > 0 on 66 of 120; non-convergent 20 of 7680
4
+ x28m3a_s2: median over plants of per-plant median (J_a0 - J) / J +0.33% [-0.23, +0.63], > 0 on 65 of 119; non-convergent 64 of 7680
5
+ x28m3b_s0: median over plants of per-plant median (J_a0 - J) / J +0.21% [-1.10, +2.12], > 0 on 65 of 119; non-convergent 0 of 7616
6
+ x28m3b_s1: median over plants of per-plant median (J_a0 - J) / J +0.66% [-0.24, +2.72], > 0 on 68 of 120; non-convergent 58 of 7680
7
+ x28m3b_s2: median over plants of per-plant median (J_a0 - J) / J +0.07% [-1.08, +1.67], > 0 on 62 of 120; non-convergent 0 of 7680
results/icml_r3/X28/mimo/gen_rep_paired_m2.txt ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ X28-M S4 with repair, m = 2: PRIMARY, repaired (b) (center r1, certificate circle) vs (a), 120 held-out plants, regret vs J*
2
+ a s0: certified per sample 99.5% | b=1: certified 100.0% regret median 12.25% | b=4: certified 100.0% regret median 10.01% | b=16: certified 100.0% regret median 8.22%
3
+ a s1: certified per sample 98.8% | b=1: certified 98.3% regret median 10.18% | b=4: certified 99.2% regret median 8.49% | b=16: certified 100.0% regret median 7.95%
4
+ a s2: certified per sample 99.0% | b=1: certified 99.2% regret median 12.09% | b=4: certified 100.0% regret median 8.72% | b=16: certified 100.0% regret median 7.99%
5
+ b s0: certified per sample 88.2% | b=1: certified 86.7% regret median 11.65% | b=4: certified 95.0% regret median 6.25% | b=16: certified 100.0% regret median 4.76%
6
+ b s1: certified per sample 87.5% | b=1: certified 86.7% regret median 13.77% | b=4: certified 96.7% regret median 7.08% | b=16: certified 100.0% regret median 4.79%
7
+ b s2: certified per sample 85.0% | b=1: certified 79.2% regret median 14.90% | b=4: certified 94.2% regret median 7.02% | b=16: certified 100.0% regret median 5.27%
8
+ b= 1 seed 0: (a) - (b) regret median -2.01 pts [-3.81, +0.90] p_holm 0.019 | certified (b) - (a) -13.3 pts McNemar p 3.1e-05 | (b) cheaper on 44.2%, dearer on 55.8%
9
+ b= 1 seed 1: (a) - (b) regret median -3.54 pts [-10.03, +0.07] p_holm 0.00027 | certified (b) - (a) -11.7 pts McNemar p 0.0013 | (b) cheaper on 41.7%, dearer on 58.3%
10
+ b= 1 seed 2: (a) - (b) regret median -2.74 pts [-9.21, +0.94] p_holm 0.00027 | certified (b) - (a) -20.0 pts McNemar p 8e-07 | (b) cheaper on 45.8%, dearer on 54.2%
11
+ b= 4 seed 0: (a) - (b) regret median +1.48 pts [-0.26, +2.46] p_holm 0.93 | certified (b) - (a) -5.0 pts McNemar p 0.031 | (b) cheaper on 57.5%, dearer on 42.5%
12
+ b= 4 seed 1: (a) - (b) regret median +0.84 pts [-1.35, +2.28] p_holm 0.93 | certified (b) - (a) -2.5 pts McNemar p 0.38 | (b) cheaper on 55.8%, dearer on 44.2%
13
+ b= 4 seed 2: (a) - (b) regret median +0.32 pts [-1.43, +1.91] p_holm 0.62 | certified (b) - (a) -5.8 pts McNemar p 0.016 | (b) cheaper on 52.5%, dearer on 47.5%
14
+ b=16 seed 0: (a) - (b) regret median +2.52 pts [+1.56, +3.47] p_holm 0.11 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 62.5%, dearer on 37.5%
15
+ b=16 seed 1: (a) - (b) regret median +1.94 pts [+0.63, +2.66] p_holm 0.11 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 64.2%, dearer on 35.8%
16
+ b=16 seed 2: (a) - (b) regret median +1.96 pts [+0.91, +2.93] p_holm 0.11 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 64.2%, dearer on 35.8%
17
+ VERDICT m=2 b=16: better on 0, worse on 0 of 3 seed pairs -> no measurable difference
results/icml_r3/X28/mimo/gen_rep_paired_m3.txt ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ X28-M S4 with repair, m = 3: PRIMARY, repaired (b) (center r1, certificate circle) vs (a), 120 held-out plants, regret vs J*
2
+ a s0: certified per sample 96.0% | b=1: certified 95.0% regret median 17.17% | b=4: certified 100.0% regret median 13.51% | b=16: certified 100.0% regret median 10.66%
3
+ a s1: certified per sample 96.6% | b=1: certified 97.5% regret median 16.83% | b=4: certified 100.0% regret median 11.77% | b=16: certified 100.0% regret median 10.46%
4
+ a s2: certified per sample 93.4% | b=1: certified 94.2% regret median 16.33% | b=4: certified 97.5% regret median 12.10% | b=16: certified 100.0% regret median 10.56%
5
+ b s0: certified per sample 79.3% | b=1: certified 83.3% regret median 23.66% | b=4: certified 96.7% regret median 11.02% | b=16: certified 99.2% regret median 8.25%
6
+ b s1: certified per sample 79.3% | b=1: certified 80.0% regret median 23.94% | b=4: certified 95.0% regret median 11.72% | b=16: certified 100.0% regret median 8.51%
7
+ b s2: certified per sample 82.6% | b=1: certified 80.8% regret median 23.89% | b=4: certified 99.2% regret median 13.16% | b=16: certified 100.0% regret median 9.04%
8
+ b= 1 seed 0: (a) - (b) regret median -3.10 pts [-8.25, +0.31] p_holm 0.006 | certified (b) - (a) -11.7 pts McNemar p 0.0043 | (b) cheaper on 41.7%, dearer on 56.7%
9
+ b= 1 seed 1: (a) - (b) regret median -4.18 pts [-9.36, +0.04] p_holm 0.00071 | certified (b) - (a) -17.5 pts McNemar p 1.9e-05 | (b) cheaper on 40.8%, dearer on 58.3%
10
+ b= 1 seed 2: (a) - (b) regret median -2.14 pts [-6.33, +0.15] p_holm 0.006 | certified (b) - (a) -13.3 pts McNemar p 0.0015 | (b) cheaper on 42.5%, dearer on 55.0%
11
+ b= 4 seed 0: (a) - (b) regret median +0.62 pts [-1.47, +2.24] p_holm 1 | certified (b) - (a) -3.3 pts McNemar p 0.12 | (b) cheaper on 52.5%, dearer on 47.5%
12
+ b= 4 seed 1: (a) - (b) regret median +0.45 pts [-0.97, +2.41] p_holm 1 | certified (b) - (a) -5.0 pts McNemar p 0.031 | (b) cheaper on 55.0%, dearer on 45.0%
13
+ b= 4 seed 2: (a) - (b) regret median +1.07 pts [-0.80, +2.78] p_holm 1 | certified (b) - (a) +1.7 pts McNemar p 0.62 | (b) cheaper on 53.3%, dearer on 46.7%
14
+ b=16 seed 0: (a) - (b) regret median +1.91 pts [+0.45, +2.71] p_holm 0.11 | certified (b) - (a) -0.8 pts McNemar p 1 | (b) cheaper on 61.7%, dearer on 38.3%
15
+ b=16 seed 1: (a) - (b) regret median +1.33 pts [+0.29, +3.40] p_holm 0.042 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 61.7%, dearer on 38.3%
16
+ b=16 seed 2: (a) - (b) regret median +1.80 pts [+0.30, +4.11] p_holm 0.042 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 61.7%, dearer on 38.3%
17
+ VERDICT m=3 b=16: better on 2, worse on 0 of 3 seed pairs -> spline better
results/icml_r3/X28/mimo/rep_run.txt ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 17:49:20 pool of 2 forked
2
+ 18:33:02 validation sampling
3
+ 18:57:38 x28m3a_s0: certified as sampled 96.1%
4
+ 18:59:10 x28m3a_s1: certified as sampled 98.4%
5
+ 19:00:13 x28m3a_s2: certified as sampled 96.6%
6
+ 19:01:09 x28m3b_s0: certified as sampled 60.7%
7
+ 19:02:09 x28m3b_s1: certified as sampled 64.1%
8
+ 19:03:37 x28m3b_s2: certified as sampled 60.4%
9
+ 19:04:50 x28m2a_s0: certified as sampled 99.1%
10
+ 19:06:13 x28m2a_s1: certified as sampled 99.2%
11
+ 19:07:12 x28m2a_s2: certified as sampled 99.4%
12
+ 19:08:49 x28m2b_s0: certified as sampled 80.2%
13
+ 19:09:48 x28m2b_s1: certified as sampled 73.7%
14
+ 19:11:10 x28m2b_s2: certified as sampled 77.7%
15
+ 19:11:12 self-checks
16
+ X28-M repair self-checks (PREREG addendum 2026-11-14), arm (b) samples on the 64 validation plants per m
17
+ check 1: Popov LMI alone feasible on 400 of 400 circle-certified samples (100.0%), need >= 95%: pass
18
+ Popov-certified among circle-rejected: 51 found in 1876 tried
19
+ check 2: Popov-only certified, 60 s from 64 directions x radii 1, 5: 24 of 6528 rollouts with final |x| >= |x0| or non-finite (FAIL); final |x| > 1e-2 |x0| on 180 of 6528
20
+ control, circle-certified: 0 of 25600 with final |x| >= |x0|; final |x| > 1e-2 |x0| on 12 of 25600
21
+ primary certificate: circle (Popov path stopped by the self-checks)
22
+ 19:14:20 validation repair, certificate circle
23
+ X28-M repair: choice of the center on the 64 validation plants per m (PREREG addendum 2026-11-14), certificate circle, b = 16
24
+ x28m3a_s0: certified 100.0%, median regret 11.48%
25
+ x28m3a_s1: certified 100.0%, median regret 12.49%
26
+ x28m3a_s2: certified 100.0%, median regret 11.63%
27
+ x28m3b_s0 center gmean: certified per sample 79.1%, b=16 certified 98.4%, median regret 8.09%
28
+ x28m3b_s0 center r1: certified per sample 82.8%, b=16 certified 98.4%, median regret 8.06%
29
+ x28m3b_s1 center gmean: certified per sample 81.2%, b=16 certified 100.0%, median regret 7.58%
30
+ x28m3b_s1 center r1: certified per sample 84.3%, b=16 certified 100.0%, median regret 7.58%
31
+ x28m3b_s2 center gmean: certified per sample 76.3%, b=16 certified 98.4%, median regret 7.40%
32
+ x28m3b_s2 center r1: certified per sample 80.4%, b=16 certified 98.4%, median regret 7.40%
33
+ x28m2a_s0: certified 100.0%, median regret 9.18%
34
+ x28m2a_s1: certified 100.0%, median regret 9.55%
35
+ x28m2a_s2: certified 100.0%, median regret 9.34%
36
+ x28m2b_s0 center gmean: certified per sample 90.1%, b=16 certified 100.0%, median regret 5.14%
37
+ x28m2b_s0 center r1: certified per sample 91.2%, b=16 certified 100.0%, median regret 5.14%
38
+ x28m2b_s1 center gmean: certified per sample 87.1%, b=16 certified 100.0%, median regret 5.40%
39
+ x28m2b_s1 center r1: certified per sample 89.3%, b=16 certified 100.0%, median regret 5.40%
40
+ x28m2b_s2 center gmean: certified per sample 89.3%, b=16 certified 100.0%, median regret 5.55%
41
+ x28m2b_s2 center r1: certified per sample 92.3%, b=16 certified 100.0%, median regret 5.32%
42
+ pooled: arm (a) certified 100.0%; gmean: certified 99.5%, median regret 6.52%; r1: certified 99.5%, median regret 6.46%
43
+ CHOICE: center r1 (eligible: ['gmean', 'r1'])
44
+ 19:17:25 held-out test: center r1, certificate circle
45
+ 19:17:59 x28m3b_s0: per sample certified as sampled 57.4% -> primary 79.3%, circle 79.3%; lam 1.0: 1103 0.75: 209 0.5: 104 0.25: 81 0.0: 26, none 397
46
+ 19:19:26 x28m3b_s1: per sample certified as sampled 58.6% -> primary 79.3%, circle 79.3%; lam 1.0: 1126 0.75: 213 0.5: 89 0.25: 54 0.0: 41, none 397
47
+ 19:19:56 x28m3b_s2: per sample certified as sampled 60.5% -> primary 82.6%, circle 82.6%; lam 1.0: 1161 0.75: 188 0.5: 121 0.25: 68 0.0: 47, none 335
48
+ X28-M S4 with repair, m = 3: PRIMARY, repaired (b) (center r1, certificate circle) vs (a), 120 held-out plants, regret vs J*
49
+ a s0: certified per sample 96.0% | b=1: certified 95.0% regret median 17.17% | b=4: certified 100.0% regret median 13.51% | b=16: certified 100.0% regret median 10.66%
50
+ a s1: certified per sample 96.6% | b=1: certified 97.5% regret median 16.83% | b=4: certified 100.0% regret median 11.77% | b=16: certified 100.0% regret median 10.46%
51
+ a s2: certified per sample 93.4% | b=1: certified 94.2% regret median 16.33% | b=4: certified 97.5% regret median 12.10% | b=16: certified 100.0% regret median 10.56%
52
+ b s0: certified per sample 79.3% | b=1: certified 83.3% regret median 23.66% | b=4: certified 96.7% regret median 11.02% | b=16: certified 99.2% regret median 8.25%
53
+ b s1: certified per sample 79.3% | b=1: certified 80.0% regret median 23.94% | b=4: certified 95.0% regret median 11.72% | b=16: certified 100.0% regret median 8.51%
54
+ b s2: certified per sample 82.6% | b=1: certified 80.8% regret median 23.89% | b=4: certified 99.2% regret median 13.16% | b=16: certified 100.0% regret median 9.04%
55
+ b= 1 seed 0: (a) - (b) regret median -3.10 pts [-8.25, +0.31] p_holm 0.006 | certified (b) - (a) -11.7 pts McNemar p 0.0043 | (b) cheaper on 41.7%, dearer on 56.7%
56
+ b= 1 seed 1: (a) - (b) regret median -4.18 pts [-9.36, +0.04] p_holm 0.00071 | certified (b) - (a) -17.5 pts McNemar p 1.9e-05 | (b) cheaper on 40.8%, dearer on 58.3%
57
+ b= 1 seed 2: (a) - (b) regret median -2.14 pts [-6.33, +0.15] p_holm 0.006 | certified (b) - (a) -13.3 pts McNemar p 0.0015 | (b) cheaper on 42.5%, dearer on 55.0%
58
+ b= 4 seed 0: (a) - (b) regret median +0.62 pts [-1.47, +2.24] p_holm 1 | certified (b) - (a) -3.3 pts McNemar p 0.12 | (b) cheaper on 52.5%, dearer on 47.5%
59
+ b= 4 seed 1: (a) - (b) regret median +0.45 pts [-0.97, +2.41] p_holm 1 | certified (b) - (a) -5.0 pts McNemar p 0.031 | (b) cheaper on 55.0%, dearer on 45.0%
60
+ b= 4 seed 2: (a) - (b) regret median +1.07 pts [-0.80, +2.78] p_holm 1 | certified (b) - (a) +1.7 pts McNemar p 0.62 | (b) cheaper on 53.3%, dearer on 46.7%
61
+ b=16 seed 0: (a) - (b) regret median +1.91 pts [+0.45, +2.71] p_holm 0.11 | certified (b) - (a) -0.8 pts McNemar p 1 | (b) cheaper on 61.7%, dearer on 38.3%
62
+ b=16 seed 1: (a) - (b) regret median +1.33 pts [+0.29, +3.40] p_holm 0.042 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 61.7%, dearer on 38.3%
63
+ b=16 seed 2: (a) - (b) regret median +1.80 pts [+0.30, +4.11] p_holm 0.042 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 61.7%, dearer on 38.3%
64
+ VERDICT m=3 b=16: better on 2, worse on 0 of 3 seed pairs -> spline better
65
+ X28-M S4 with repair, fresh directions, m = 3: best certified b=16 sample per held-out plant on 64 unit directions (seed (7575, id)), vs x28m3a_s0; (b) repaired (primary)
66
+ x28m3a_s0: median over plants of per-plant median (J_a0 - J) / J +0.00% [+0.00, +0.00], > 0 on 0 of 120; non-convergent 54 of 7680
67
+ x28m3a_s1: median over plants of per-plant median (J_a0 - J) / J +0.26% [-0.16, +0.61], > 0 on 66 of 120; non-convergent 20 of 7680
68
+ x28m3a_s2: median over plants of per-plant median (J_a0 - J) / J +0.33% [-0.23, +0.63], > 0 on 65 of 119; non-convergent 64 of 7680
69
+ x28m3b_s0: median over plants of per-plant median (J_a0 - J) / J +0.21% [-1.10, +2.12], > 0 on 65 of 119; non-convergent 0 of 7616
70
+ x28m3b_s1: median over plants of per-plant median (J_a0 - J) / J +0.66% [-0.24, +2.72], > 0 on 68 of 120; non-convergent 58 of 7680
71
+ x28m3b_s2: median over plants of per-plant median (J_a0 - J) / J +0.07% [-1.08, +1.67], > 0 on 62 of 120; non-convergent 0 of 7680
72
+ 19:22:42 x28m2b_s0: per sample certified as sampled 75.9% -> primary 88.2%, circle 88.2%; lam 1.0: 1458 0.75: 111 0.5: 64 0.25: 30 0.0: 31, none 226
73
+ 19:22:59 x28m2b_s1: per sample certified as sampled 76.0% -> primary 87.5%, circle 87.5%; lam 1.0: 1460 0.75: 95 0.5: 53 0.25: 38 0.0: 34, none 240
74
+ 19:23:13 x28m2b_s2: per sample certified as sampled 73.4% -> primary 85.0%, circle 85.0%; lam 1.0: 1409 0.75: 99 0.5: 56 0.25: 34 0.0: 34, none 288
75
+ X28-M S4 with repair, m = 2: PRIMARY, repaired (b) (center r1, certificate circle) vs (a), 120 held-out plants, regret vs J*
76
+ a s0: certified per sample 99.5% | b=1: certified 100.0% regret median 12.25% | b=4: certified 100.0% regret median 10.01% | b=16: certified 100.0% regret median 8.22%
77
+ a s1: certified per sample 98.8% | b=1: certified 98.3% regret median 10.18% | b=4: certified 99.2% regret median 8.49% | b=16: certified 100.0% regret median 7.95%
78
+ a s2: certified per sample 99.0% | b=1: certified 99.2% regret median 12.09% | b=4: certified 100.0% regret median 8.72% | b=16: certified 100.0% regret median 7.99%
79
+ b s0: certified per sample 88.2% | b=1: certified 86.7% regret median 11.65% | b=4: certified 95.0% regret median 6.25% | b=16: certified 100.0% regret median 4.76%
80
+ b s1: certified per sample 87.5% | b=1: certified 86.7% regret median 13.77% | b=4: certified 96.7% regret median 7.08% | b=16: certified 100.0% regret median 4.79%
81
+ b s2: certified per sample 85.0% | b=1: certified 79.2% regret median 14.90% | b=4: certified 94.2% regret median 7.02% | b=16: certified 100.0% regret median 5.27%
82
+ b= 1 seed 0: (a) - (b) regret median -2.01 pts [-3.81, +0.90] p_holm 0.019 | certified (b) - (a) -13.3 pts McNemar p 3.1e-05 | (b) cheaper on 44.2%, dearer on 55.8%
83
+ b= 1 seed 1: (a) - (b) regret median -3.54 pts [-10.03, +0.07] p_holm 0.00027 | certified (b) - (a) -11.7 pts McNemar p 0.0013 | (b) cheaper on 41.7%, dearer on 58.3%
84
+ b= 1 seed 2: (a) - (b) regret median -2.74 pts [-9.21, +0.94] p_holm 0.00027 | certified (b) - (a) -20.0 pts McNemar p 8e-07 | (b) cheaper on 45.8%, dearer on 54.2%
85
+ b= 4 seed 0: (a) - (b) regret median +1.48 pts [-0.26, +2.46] p_holm 0.93 | certified (b) - (a) -5.0 pts McNemar p 0.031 | (b) cheaper on 57.5%, dearer on 42.5%
86
+ b= 4 seed 1: (a) - (b) regret median +0.84 pts [-1.35, +2.28] p_holm 0.93 | certified (b) - (a) -2.5 pts McNemar p 0.38 | (b) cheaper on 55.8%, dearer on 44.2%
87
+ b= 4 seed 2: (a) - (b) regret median +0.32 pts [-1.43, +1.91] p_holm 0.62 | certified (b) - (a) -5.8 pts McNemar p 0.016 | (b) cheaper on 52.5%, dearer on 47.5%
88
+ b=16 seed 0: (a) - (b) regret median +2.52 pts [+1.56, +3.47] p_holm 0.11 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 62.5%, dearer on 37.5%
89
+ b=16 seed 1: (a) - (b) regret median +1.94 pts [+0.63, +2.66] p_holm 0.11 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 64.2%, dearer on 35.8%
90
+ b=16 seed 2: (a) - (b) regret median +1.96 pts [+0.91, +2.93] p_holm 0.11 | certified (b) - (a) +0.0 pts McNemar p 1 | (b) cheaper on 64.2%, dearer on 35.8%
91
+ VERDICT m=2 b=16: better on 0, worse on 0 of 3 seed pairs -> no measurable difference
92
+ X28-M S4 with repair, fresh directions, m = 2: best certified b=16 sample per held-out plant on 64 unit directions (seed (7575, id)), vs x28m2a_s0; (b) repaired (primary)
93
+ x28m2a_s0: median over plants of per-plant median (J_a0 - J) / J +0.00% [+0.00, +0.00], > 0 on 0 of 120; non-convergent 0 of 7680
94
+ x28m2a_s1: median over plants of per-plant median (J_a0 - J) / J -0.07% [-0.18, +0.01], > 0 on 50 of 120; non-convergent 35 of 7680
95
+ x28m2a_s2: median over plants of per-plant median (J_a0 - J) / J -0.07% [-0.16, +0.14], > 0 on 57 of 120; non-convergent 0 of 7680
96
+ x28m2b_s0: median over plants of per-plant median (J_a0 - J) / J +2.03% [+0.40, +3.08], > 0 on 77 of 120; non-convergent 123 of 7680
97
+ x28m2b_s1: median over plants of per-plant median (J_a0 - J) / J +1.69% [+0.51, +2.54], > 0 on 78 of 120; non-convergent 42 of 7680
98
+ x28m2b_s2: median over plants of per-plant median (J_a0 - J) / J +1.29% [+0.49, +2.41], > 0 on 75 of 120; non-convergent 66 of 7680
99
+ 19:24:58 repair test done
results/icml_r3/X28/mimo/rep_selfcheck.txt ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ X28-M repair self-checks (PREREG addendum 2026-11-14), arm (b) samples on the 64 validation plants per m
2
+ check 1: Popov LMI alone feasible on 400 of 400 circle-certified samples (100.0%), need >= 95%: pass
3
+ Popov-certified among circle-rejected: 51 found in 1876 tried
4
+ check 2: Popov-only certified, 60 s from 64 directions x radii 1, 5: 24 of 6528 rollouts with final |x| >= |x0| or non-finite (FAIL); final |x| > 1e-2 |x0| on 180 of 6528
5
+ control, circle-certified: 0 of 25600 with final |x| >= |x0|; final |x| > 1e-2 |x0| on 12 of 25600
6
+ primary certificate: circle (Popov path stopped by the self-checks)
results/icml_r3/X28/mimo/rep_tails.txt ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ X28-M S4 with repair, diagnostic: paired regret difference d = (a) - (b) at b = 16 (points), primary repaired (b) vs (a)
2
+ m=3 s0: median +1.91, mean +1.37; (b) better by > 5 on 35, worse by > 5 on 25, by > 10 on 13; Wilcoxon p before Holm 0.105
3
+ (b) worse by > 10 n = 13: samples certified as sampled per plant (of 16) median 2.0; best sample unrepaired on 30.8%; median regret (b) 37.4%, (a) 20.5%
4
+ (b) better by > 5 n = 35: samples certified as sampled per plant (of 16) median 10.0; best sample unrepaired on 88.6%; median regret (b) 5.1%, (a) 15.5%
5
+ all plants n = 120: samples certified as sampled per plant (of 16) median 10.0; best sample unrepaired on 78.3%; median regret (b) 8.3%, (a) 10.7%
6
+ m=3 s1: median +1.33, mean +1.04; (b) better by > 5 on 36, worse by > 5 on 19, by > 10 on 12; Wilcoxon p before Holm 0.014
7
+ (b) worse by > 10 n = 12: samples certified as sampled per plant (of 16) median 3.0; best sample unrepaired on 75.0%; median regret (b) 36.6%, (a) 11.7%
8
+ (b) better by > 5 n = 36: samples certified as sampled per plant (of 16) median 12.0; best sample unrepaired on 80.6%; median regret (b) 6.2%, (a) 16.7%
9
+ all plants n = 120: samples certified as sampled per plant (of 16) median 10.0; best sample unrepaired on 86.7%; median regret (b) 8.5%, (a) 10.5%
10
+ m=3 s2: median +1.80, mean +1.63; (b) better by > 5 on 42, worse by > 5 on 22, by > 10 on 17; Wilcoxon p before Holm 0.014
11
+ (b) worse by > 10 n = 17: samples certified as sampled per plant (of 16) median 2.0; best sample unrepaired on 47.1%; median regret (b) 26.1%, (a) 10.2%
12
+ (b) better by > 5 n = 42: samples certified as sampled per plant (of 16) median 10.0; best sample unrepaired on 90.5%; median regret (b) 6.4%, (a) 17.6%
13
+ all plants n = 120: samples certified as sampled per plant (of 16) median 10.0; best sample unrepaired on 84.2%; median regret (b) 9.0%, (a) 10.6%
14
+ m=2 s0: median +2.52, mean -2.39; (b) better by > 5 on 37, worse by > 5 on 23, by > 10 on 17; Wilcoxon p before Holm 0.036
15
+ (b) worse by > 10 n = 17: samples certified as sampled per plant (of 16) median 3.0; best sample unrepaired on 52.9%; median regret (b) 38.9%, (a) 6.9%
16
+ (b) better by > 5 n = 37: samples certified as sampled per plant (of 16) median 16.0; best sample unrepaired on 94.6%; median regret (b) 3.6%, (a) 13.0%
17
+ all plants n = 120: samples certified as sampled per plant (of 16) median 15.0; best sample unrepaired on 90.0%; median regret (b) 4.8%, (a) 8.2%
18
+ m=2 s1: median +1.94, mean +0.36; (b) better by > 5 on 36, worse by > 5 on 24, by > 10 on 15; Wilcoxon p before Holm 0.050
19
+ (b) worse by > 10 n = 15: samples certified as sampled per plant (of 16) median 6.0; best sample unrepaired on 80.0%; median regret (b) 35.1%, (a) 9.6%
20
+ (b) better by > 5 n = 36: samples certified as sampled per plant (of 16) median 15.5; best sample unrepaired on 97.2%; median regret (b) 3.1%, (a) 12.9%
21
+ all plants n = 120: samples certified as sampled per plant (of 16) median 15.0; best sample unrepaired on 91.7%; median regret (b) 4.8%, (a) 7.9%
22
+ m=2 s2: median +1.96, mean -1.36; (b) better by > 5 on 37, worse by > 5 on 26, by > 10 on 19; Wilcoxon p before Holm 0.111
23
+ (b) worse by > 10 n = 19: samples certified as sampled per plant (of 16) median 4.0; best sample unrepaired on 57.9%; median regret (b) 41.4%, (a) 8.4%
24
+ (b) better by > 5 n = 37: samples certified as sampled per plant (of 16) median 16.0; best sample unrepaired on 100.0%; median regret (b) 4.2%, (a) 13.2%
25
+ all plants n = 120: samples certified as sampled per plant (of 16) median 15.0; best sample unrepaired on 90.0%; median regret (b) 5.3%, (a) 8.0%
results/icml_r3/X28/mimo/rep_val_choice.txt ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ X28-M repair: choice of the center on the 64 validation plants per m (PREREG addendum 2026-11-14), certificate circle, b = 16
2
+ x28m3a_s0: certified 100.0%, median regret 11.48%
3
+ x28m3a_s1: certified 100.0%, median regret 12.49%
4
+ x28m3a_s2: certified 100.0%, median regret 11.63%
5
+ x28m3b_s0 center gmean: certified per sample 79.1%, b=16 certified 98.4%, median regret 8.09%
6
+ x28m3b_s0 center r1: certified per sample 82.8%, b=16 certified 98.4%, median regret 8.06%
7
+ x28m3b_s1 center gmean: certified per sample 81.2%, b=16 certified 100.0%, median regret 7.58%
8
+ x28m3b_s1 center r1: certified per sample 84.3%, b=16 certified 100.0%, median regret 7.58%
9
+ x28m3b_s2 center gmean: certified per sample 76.3%, b=16 certified 98.4%, median regret 7.40%
10
+ x28m3b_s2 center r1: certified per sample 80.4%, b=16 certified 98.4%, median regret 7.40%
11
+ x28m2a_s0: certified 100.0%, median regret 9.18%
12
+ x28m2a_s1: certified 100.0%, median regret 9.55%
13
+ x28m2a_s2: certified 100.0%, median regret 9.34%
14
+ x28m2b_s0 center gmean: certified per sample 90.1%, b=16 certified 100.0%, median regret 5.14%
15
+ x28m2b_s0 center r1: certified per sample 91.2%, b=16 certified 100.0%, median regret 5.14%
16
+ x28m2b_s1 center gmean: certified per sample 87.1%, b=16 certified 100.0%, median regret 5.40%
17
+ x28m2b_s1 center r1: certified per sample 89.3%, b=16 certified 100.0%, median regret 5.40%
18
+ x28m2b_s2 center gmean: certified per sample 89.3%, b=16 certified 100.0%, median regret 5.55%
19
+ x28m2b_s2 center r1: certified per sample 92.3%, b=16 certified 100.0%, median regret 5.32%
20
+ pooled: arm (a) certified 100.0%; gmean: certified 99.5%, median regret 6.52%; r1: certified 99.5%, median regret 6.46%
21
+ CHOICE: center r1 (eligible: ['gmean', 'r1'])