File size: 24,209 Bytes
374b6ec
 
 
1778de1
374b6ec
 
 
10c5142
374b6ec
 
 
 
 
 
 
 
1778de1
374b6ec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1778de1
374b6ec
 
 
 
 
1778de1
374b6ec
 
 
 
 
10c5142
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
---
title: C+phi next to C-only with the same baselines and KPIs (X29). SISO, the two generators tie at 16 samples on the common J* and both have lower median regret than kNN, the regression twin and the flat references. MIMO spline setting (200 fresh plants per m), the repaired C+phi generator has the lowest median regret of the generators and baselines at 16 samples at both m, and C-only certifies more at 1 sample. Actuator setting, at the per-plant optimum C+phi equals C-only (median gain 0.00% at m = 3, 4, 5)
date: 2026-11-17
status: complete. SISO (with CPU), MIMO spline setting (with CPU) and actuator step 3A done; actuator step 3B (C+phi generators on X27's banks) not run, by Matteo's decision of 2026-11-17 (the deck draws C+phi at the C-only values in that setting, labeled)
claim: X29 (Matteo 2026-11-16, via dc_presentation and in the move-HPC session): "I need to have the final baseline comparison also for the spline methods". C+phi (the model generates the linear part and the spline phi, X28) gets the same baselines and KPIs as C-only (phi fixed: SISO sat_1, MIMO identity) in every setting of the deck v10 slides.
maps_to: deck v10 slides [35] [36] [37] [38] [73] (SISO); the three new MIMO spline-setting slides and [73]; [50] [51] [52] [84] [73] (actuator); summaries [54] [72] [77]. No paper edit until Matteo agrees the claim.
scripts: nonlinear/experiments/x29_siso.py (SISO kNN, twins, optimizer from the best sample, timing, report), nonlinear/mimo_unstable/x29_t2.py (MIMO spline setting on T2: kNN, twins, 48-start solver, Adam references, timing, report), nonlinear/mimo_unstable/x29_act.py (actuator 3A: composed-sector certificate, simulated X27 cost, optimizer, new directions, report, X27 rows and re-score)
artifacts: results/icml_r3/X29/{PREREG.md, PREREG_siso.md, PREREG_t2.md}; results/icml_r3/X29/siso/{report.txt, eqcheck.txt, eval/, refine/, timing/, timing_batched.json}; results/icml_r3/X29/t2/{report.txt, numbers.json, cpu_report.txt, timing_m2.json, timing_m3.json, knn_m*_{a,b}.npz, twin_*.npz, solver_m*.npz, ard_m2.npz}; results/icml_r3/X29/act/{selfcheck_*.json, report_3a_heldout_m*.txt, heldout_m*/, x27rows_*.npy, x27rescore_*.txt, x27rescore_*.json}; models/x29_twin_clS{,_s1,_s2}/, models/x29t/; HF matCercola18/DiffusionControl commits 8f92ad1d (SISO), d7d0258 (MIMO spline setting), 09f35d5 (actuator 3A), 374b6ec (SISO CPU, X27 re-score in the X29 convention, this write-up); git branch icml-experiments 225e6cf
---

# C+phi next to C-only with the same baselines and KPIs (X29)

## TL;DR
- **Question:** The deck compares C-only (phi fixed; the model generates the linear part) with kNN, regression twins, per-plant optimizers and classical references. Matteo asked for the same comparison for C+phi (the model also generates the spline phi, X28) in every setting, on one regret reference per setting.
- **SISO (199 test plants):** at 16 samples the C+phi generator certifies 95.6% of plants with median regret 6.59%, and C-only 98.0% with 6.74% (3-seed means). Both have lower regret than their kNN (11.3%, 10.2%), their regression twins (31.8%, 20.1%), the tuned LQG or static gain (34.05%) and the best of 16 HIFOO starts (50.75%). At 1 sample C+phi certifies 50.6% of plants against 64.0%. The joint optimizer reaches the C+phi generator's regret with 2 starts; the F-only optimizer needs 8 to reach C-only's. On one CPU thread the C+phi generator costs 29.1 s per plant at 16 samples against 11.5 s (same sampling time, slower scoring), and the variant's own optimizer at the start count that reaches the generator's regret costs 106 times the C+phi generator's CPU and 138 times the C-only generator's (October: 140 times).
- **MIMO spline setting (200 fresh plants per m):** at 16 samples the repaired C+phi generator has median regret 6.56% at m = 3 and 3.59% at m = 2, against 10.22% and 6.42% for C-only, with certified shares within 0.5 points. Its kNN and twin also have lower regret than C-only's. The 48-start solver has 11.60% and 12.33%. At 1 sample C-only certifies more plants (95.7% against 79.7% at m = 3). Each generator costs about 3 s per plant on one CPU thread and about 0.05 s per plant batched on a GPU; the per-plant Adam optimizer costs about 4,200 to 5,100 s and 34 to 41 s.
- **Actuator setting (X27's 120 test plants per m):** at each plant's optimum C+phi costs the same as C-only (median gain 0.00% at m = 3, 4 and 5), and the best certified spline is linear on most plants. So a C+phi generator can at best tie here. Matteo decided not to run step 3B (C+phi generators on X27's banks, 2 to 3 GPU-days); the deck draws C+phi at the C-only values in this setting, labeled as not trained separately. X27's C-only pipeline, re-run on its stored candidates, reproduces every published number; on the common J*_A the regret of its filtered arms at 16 samples rises by 0.5 to 1.2 points in X27's convention (median over certified plants).
- **So what:** C+phi changes the comparison only in the MIMO spline setting, where it has lower regret than C-only for every method family at 16 samples. On SISO it ties, and in the actuator setting it has nothing to gain at the optimum. With 1 sample, C-only certifies more plants in SISO and in the MIMO spline setting.

## 1. Problem / Goal
- **C-only:** SISO u = -sat_1(clip(F y, +-5)) with a generated dynamic filter F (X21); MIMO a generated gain K with phi = identity, in the X28-M spline setting (u = -K y) and in X27's actuator setting (u = -Phi(K y), Phi unknown in a sector).
- **C+phi:** the same model also generates phi. SISO: an 11-knot spline with |phi| <= 1, phi(0) = 0 (X28 S3). MIMO: per channel an odd piecewise-linear spline with knots at |x| = 1, 3, 5 and knot ratios in [0.1, 10] (the so3 class, X28-M).
- **Goal:** every deck KPI (certified share and median regret vs budget b, CPU vs regret, compute ratio, generator minus kNN at b = 1) for both variants, on one reference J* per setting, with the same baselines. Nothing published is replaced; C-only numbers are re-scored on the common J*.
- **No new verdict rule** in SISO and the MIMO spline setting: the registered X28 verdicts stand. The actuator setting had no C+phi yet; its first step (3A) measures whether C+phi can gain at all.

## 2. Method
Pre-registration: `results/icml_r3/X29/PREREG.md` (2026-11-16 18:30 UTC, before any X29 code ran) and its addenda in `PREREG.md`, `PREREG_siso.md` and `PREREG_t2.md`, each filed before the number it affects.

- **Regret** at budget b: (J_b - J*) / J*, with J_b the cheapest certified of the first b candidates. A plant with no certified candidate counts 1e3. A cell where fewer than half the plants certify has no finite median ("n/a"). Median with a bootstrap 95% CI over plants (10,000 resamples). Seeded methods: 3-seed mean and range, plus seed 0.
- **J\* per setting:** SISO min(J_Fonly, X28 test joint labels) as X28 S4; MIMO spline setting min of the T2 id and so3 labels (gauges c = 1, 2, 5); actuator min(J_ref, J_id, J_so3) from step 3A.
- **SISO arms** (199 X21 test plants, clip 1.5, t-repair on every arm):
  - generators: clL (C-only) and clS (C+phi), seeds 0 to 2, stored X28 evaluations; 64 samples for seed 0;
  - kNN: X22's rule on the X21 F-only bank (C-only) and on the c = 1 joint rows of the 1,971 training plants (C+phi), ARD metric of X21;
  - regression twins: X23's twins of clL (stored X21 outputs, see deviations) and new twins of clS (X23 recipe, 60,000 steps, best validation checkpoint), seeds 0 to 2;
  - per-plant optimizers: F-only, best of the first k X21 labels (C-only's own); joint, best of the first k joint rows (C+phi's own); k = 1, 2, 4, 8, 16;
  - optimizer from the generator's best b = 16 sample: 1 F-only start (X24 refine) for C-only, 1 joint start (S2 warm arm) for C+phi, seed 0;
  - flat references: tuned saturated LQG or static gain (J_lin), best of 16 HIFOO starts.
- **MIMO spline setting** (T2: 200 fresh plants per m, m = 2, 3, as the X28 addendum of 2026-11-15):
  - generators: C-only (a) and C+phi (b), seeds 0 to 2, stored; C+phi repaired (repair R: rejected samples' knots flattened toward r(1) until the circle LMI certifies) is primary, as drawn is secondary;
  - kNN: label (a) or (b) of the 16 nearest S2-M bank plants, ARD metric (m = 3 X10's, m = 2 fitted on the m = 2 bank before any kNN number);
  - regression twins: `mh_train` arm "reg" on the S3-M datasets (a) and (b), seeds 0 to 2;
  - flat references: the 48-start solver's best gain (X27 oracle, Lyapunov test, X28 cost) and the per-plant Adam optimizer of each variant (C-only: id start; C+phi: J*_T2 itself, regret 0 by construction).
- **Actuator setting, step 3A** (X27's 120 test plants per m, m = 3, 4, 5):
  - controller u = -Phi(phi(K y)), with Phi the unknown actuator (each channel in the sector [0.4, 2.2], X27) and phi in the so3 class; C-only is phi = identity;
  - certificate: circle LMI on the composed sector [0.4 alpha_i, 2.2 beta_i], alpha_i and beta_i the min and max knot ratio of channel i. It is sound because each channel's composition Phi_i(psi_i(.)) is memoryless and lies in that sector; at identity knots it is X27's test;
  - cost: X27's quadratic cost (mean over the 2m unit state directions of the integral of |x|^2 + 0.1 |u|^2, Phi = identity) by simulation, RK4 at dt 0.01, T = 20 s;
  - label optimizer: X28-M's procedure from X27's stored best gain K_ref; arm id (knots frozen) and arm so3 at gauges c = 1, 2, 5; K lr and knot lr 0.02; 400 steps; the cheapest certified snapshot;
  - reported: d = (J_id - J_so3) / J_so3; (J_ref - J_id) / J_id (extra optimization alone); the composed-sector ratio of the best so3 controller; d on 64 new random unit state directions per plant;
  - X27's stored C-only candidates re-run through X27's pipeline per plant and re-scored on J*_A.
- **CPU:** one timing job per setting on a node with nothing else of this project running. "One" = one plant on one CPU thread, median over 10 timing plants (SISO: 6 such tasks at a time on 8 cores; MIMO: one at a time).
  - MIMO (node move-hpc-08-gpu): "batched" = all 200 plants in one call, with sampling, rollouts and Adam on the node's NVIDIA L40S GPU and certificates on 6 CPU workers, wall time / 200.
  - SISO (node move-hpc-06-gpu, also an L40S): "batched" = generator and twin sampling (199 plants x 16) and kNN retrieval in one call, wall time / 199, without scoring (PREREG section 2).
  - Per-point CPU follows X24's formulas; the joint optimizer at k starts is charged k HIFOO starts + the F-only optimizer at k + the joint optimizer at k.
- **Self-checks** (two failed and were handled by addenda before the numbers they affect; see deviations):
  - SISO: C-only kNN reproduces X21's stored kNN (K16 +0.16 points per candidate, +0.50 plants; K64 +0.07, +0.00); batched optimizers equal the single-plant functions (relative J difference 0 at 25 and 100 steps);
  - MIMO spline setting: 5 bank plants per m retrieve themselves first; the solver's cost equals the stored label snapshot (maximum relative difference 0);
  - actuator: the composed-sector certificate at identity knots agrees with `family_sofm.verify` on 200 gains per m (0 disagreements at m = 3, 4, 5); the simulated cost equals tr(P)/(2m) on the 120 stored gains per m (median relative difference 1.3e-5 to 2.2e-5, maximum 2.6e-4 to 1.4e-3).

## 3. Results

### 3.1 SISO, 199 test plants
Source: `results/icml_r3/X29/siso/report.txt`. Certified % / median regret %; generator and twin: 3-seed mean [range]; kNN: one run; b = 64: seed 0.

| method | b = 1 | b = 4 | b = 16 | b = 64 |
|---|---|---|---|---|
| C-only generator | 64.0 / 49.0 | 92.0 / 12.7 | 98.0 [97.5, 98.5] / 6.74 [6.62, 6.81] | 98.5 / 4.07 |
| C+phi generator | 50.6 / n/a on 2 of 3 seeds | 84.9 / 17.1 | 95.6 [95.0, 96.5] / 6.59 [6.30, 7.14] | 99.0 / 2.96 |
| C-only kNN | 57.3 / 81.3 | 86.9 / 19.4 | 96.5 / 10.19 | 98.5 / 6.49 |
| C+phi kNN | 46.2 / n/a | 77.9 / 22.7 | 93.5 / 11.32 | 98.0 / 6.26 |
| C-only twin | 68.5 / 48.1 | 77.2 / 30.3 | 85.6 [85.4, 85.9] / 20.1 [19.9, 20.3] | |
| C+phi twin | 42.2 / n/a | 61.8 / 66.1 | 71.0 [68.8, 72.9] / 31.8 [30.0, 32.9] | |
| LQG or static gain | | | 100 / 34.05 | |
| HIFOO, best of 16 | | | 100 / 50.75 | |

Per-plant optimizers, median regret % at k = 1, 2, 4, 8, 16 starts (certified 100%):
- F-only (C-only's own): 35.91, 13.55, 8.11, 4.19, 3.12.
- Joint (C+phi's own): 9.13, 4.39, 0.92, 0.00, 0.00. The joint labels define J*, so 0 at 8 and 16 starts is by construction.
- Generator at 16 plus 1 optimizer start (seed 0): C-only 98.5% / 2.60%, C+phi 95.0% / 0.01%. Only plants with a certified b = 16 sample get a start (196 C-only, 189 C+phi); the rest count as uncertified.

Compute ratio inputs ([38], seed 0): the C-only generator's 6.79% is first reached by the F-only optimizer at k = 8 (4.19%; k = 4 gives 8.11%). The C+phi generator's 7.14% is first reached by the joint optimizer at k = 2 (4.39%; k = 1 gives 9.13%). The paired verdict at b = 16 stays X28 S4's: no measurable difference on 3 of 3 seed pairs (gain -0.32 to +0.19 points).

[73], certified at b = 1, generator minus its own variant's kNN (points, plant-paired, 95% CI), seeds 0 / 1 / 2:
- C-only: +8.0 [-0.5, +16.6] / +8.0 [-0.5, +16.6] / +4.0 [-4.0, +12.1].
- C+phi: +3.0 [-6.0, +12.1] / +6.5 [-3.0, +16.1] / +3.5 [-5.5, +12.6].
- All six intervals include 0.

### 3.2 MIMO spline setting, 200 fresh plants per m
Source: `results/icml_r3/X29/t2/report.txt`. Certified % / median regret %, 3-seed means (kNN one run).

| method | m = 3, b = 1 | m = 3, b = 16 | m = 2, b = 1 | m = 2, b = 16 |
|---|---|---|---|---|
| C+phi generator, repaired | 79.7 / 17.73 | 99.8 / 6.56 | 90.5 / 10.17 | 99.7 / 3.59 |
| C+phi generator, as drawn | 61.5 / 22.85 | 95.3 / 6.60 | 78.0 / 10.68 | 97.2 / 3.65 |
| C-only generator | 95.7 / 16.32 | 99.5 / 10.22 | 98.5 / 8.96 | 99.8 / 6.42 |
| C+phi kNN, repaired | 90.5 / 61.40 | 100 / 15.88 | 88.0 / 34.85 | 100 / 8.89 |
| C+phi kNN, as drawn | 48.0 / n/a | 99.5 / 17.55 | 60.0 / 76.27 | 100 / 9.83 |
| C-only kNN | 80.0 / 70.87 | 100 / 23.77 | 89.5 / 42.67 | 100 / 13.56 |
| C+phi twin, repaired | 76.3 / 22.16 | 95.0 / 10.84 | 88.3 / 12.14 | 98.0 / 6.23 |
| C+phi twin, as drawn | 60.3 / 28.50 | 83.2 / 11.26 | 77.7 / 12.74 | 90.8 / 6.31 |
| C-only twin | 94.2 / 18.21 | 98.7 / 13.81 | 98.7 / 9.22 | 99.7 / 7.71 |
| 48-start solver | | 100 / 11.60 | | 100 / 12.33 |
| Adam, C-only (id start) | | 100 / 4.33 | | 100 / 5.07 |
| Adam, C+phi (id + so3 c = 1, 2, 5) | | 100 / 0 (defines J*) | | 100 / 0 (defines J*) |

- At b = 4 (same file): repaired C+phi generator 93.8 / 9.70 (m = 3) and 97.8 / 5.47 (m = 2); C-only generator 99.2 / 12.73 and 99.3 / 7.30.
- [73], certified at b = 1, generator minus its own kNN (points, 95% CI), seeds 0 / 1 / 2:
  - m = 3: C+phi repaired -10.0 [-17.0, -3.0] / -12.5 [-19.5, -5.5] / -10.0 [-16.5, -3.5]; as drawn +14.0 / +13.0 / +13.5 (CIs above 0); C-only +15.5 / +17.5 / +14.0 (CIs above 0).
  - m = 2: C+phi repaired +3.0 [-3.0, +9.0] / +2.0 [-4.0, +8.0] / +2.5 [-3.0, +8.0]; as drawn +18.0 / +15.5 / +20.5 (CIs above 0); C-only +9.0 on all seeds (CIs above 0).
  - With repair, the C+phi kNN certifies more plants at b = 1 than the C+phi generator at m = 3 (90.5% against 79.7%).

### 3.3 Actuator setting, step 3A (per-plant optimum), 120 test plants per m
Source: `results/icml_r3/X29/act/report_3a_heldout_m{3,4,5}.txt`. All id and so3 starts certified on 118 to 120 of 120 plants.

| | m = 3 | m = 4 | m = 5 |
|---|---|---|---|
| d = (J_id - J_so3) / J_so3, median [CI] | 0.00% [0.00, 0.00] | 0.00% [0.00, 0.00] | 0.00% [0.00, 0.00] |
| plants with d >= 2% | 16 of 120 | 9 of 120 | 13 of 120 |
| mean d | 0.68% | 0.34% | 0.44% |
| extra optimization alone, (J_ref - J_id) / J_id, median | 0.00% | 0.00% | 0.00% |
| composed-sector ratio of the best so3 controller, median (90th pct); 5.5 = linear | 5.50 (15.97) | 5.50 (11.37) | 5.50 (12.18) |
| d on 64 new directions per plant, median of per-plant medians | +0.00% | +0.00% | +0.00% |

- The best so3 controller comes from the id start on 69, 78 and 76 of 120 plants: the spline starts found nothing cheaper there.
- The written expectation (PREREG section 4) was a small d: the composed sector is 5.5 beta/alpha wide, so a nonlinear spline needs a wider robust sector than a linear gain, and the quadratic cost does not reward an amplitude-dependent gain.

X27's C-only pipeline re-scored on J*_A (`results/icml_r3/X29/act/x27rescore_*.txt`). Re-running it on the stored candidates reproduces every published certified share and regret. Certified shares are unchanged. X27 publishes the median over certified plants (seeds pooled); in that convention, median regret %, published (vs J_ref) and on J*_A:

| X27 arm (keep the cheapest certified) | m = 3 | m = 4 | m = 5 (16k bank) |
|---|---|---|---|
| diffusion, filter, b = 16 | 1.50 -> 2.18 | 3.04 -> 3.73 | 3.38 -> 3.92 |
| diffusion, filter + 10 solver iterations | 1.50 -> 2.21 | 3.24 -> 3.85 | 3.44 -> 4.00 |
| twin, filter, b = 16 | 2.50 -> 3.39 | 3.46 -> 4.00 | 3.57 -> 4.59 |
| kNN, filter + 10 solver iterations | 7.34 -> 8.58 | 9.03 -> 9.68 | 10.34 -> 10.97 |

Stopping at the first certified candidate instead, the diffusion arm with iterations has 4.01 -> 5.15, 5.71 -> 6.40 and 5.65 -> 6.75.

In the X29 convention (all plants, uncertified = 1e3, median per seed, 3-seed mean), on J*_A, cheapest certified at b = 16, certified % / median regret %:

| X27 arm | m = 3 | m = 4 | m = 5 |
|---|---|---|---|
| diffusion, filter + 10 solver iterations | 99.7 / 2.22 | 97.8 / 3.96 | 98.1 / 4.04 |
| diffusion, filter | 98.6 / 2.24 | 91.9 / 4.05 | 90.8 / 4.30 |
| twin, filter | 73.1 / 5.25 | 52.2 / 12.06 | 55.6 / 9.61 |
| kNN, filter + 10 solver iterations (one run) | 93.3 / 8.84 | 93.3 / 10.01 | 92.5 / 11.51 |

The twin's low certified share counts against it here, unlike in the published convention.

### 3.4 CPU
MIMO spline setting (`results/icml_r3/X29/t2/cpu_report.txt`; seconds per plant, one | batched; median regret on J*_T2):

| method | m = 3 | m = 2 |
|---|---|---|
| C+phi generator, b = 16 | 2.96 \| 0.055; 6.56% | 2.85 \| 0.054; 3.59% |
| C-only generator, b = 16 | 2.88 \| 0.050; 10.22% | 2.79 \| 0.049; 6.42% |
| C+phi kNN, b = 16 | 1.34 \| 0.054; 15.88% | 1.27 \| 0.053; 8.89% |
| C-only kNN, b = 16 | 1.25 \| 0.042; 23.77% | 1.18 \| 0.043; 13.56% |
| C+phi twin, b = 16 | 1.39 \| 0.050; 10.84% | 1.28 \| 0.051; 6.23% |
| C-only twin, b = 16 | 1.35 \| 0.045; 13.81% | 1.24 \| 0.044; 7.71% |
| 48-start solver, all starts | 6.01 \| not batched; 11.60% | 7.38 \| not batched; 12.33% |
| Adam, C+phi (4 starts, 400 steps + certify) | 5,131 \| 41.4; 0% | 5,102 \| 41.4; 0% |
| Adam, C-only (id start) | 4,177 \| 34.6; 4.33% | 4,170 \| 34.1; 5.07% |

At b = 1 a generator costs 1/16 of its b = 16 cost (X24's convention): 0.18 to 0.19 s one at a time.

SISO (`results/icml_r3/X29/siso/report.txt`, CPU section; seconds per plant on one CPU thread, median over the 10 timing plants; median regret on J*, seed 0):

| point | CPU | median regret |
|---|---|---|
| C-only generator, b = 1 / b = 16 | 0.7 s / 11.5 s | 54.90% / 6.79% |
| C+phi generator, b = 1 / b = 16 | 1.8 s / 29.1 s | n/a / 7.14% |
| F-only optimizer, 1 / 2 / 4 / 8 / 16 starts | 1,555 / 1,544 / 1,522 / 1,582 / 1,702 s | 35.91 / 13.55 / 8.11 / 4.19 / 3.12% |
| joint optimizer, 1 / 2 / 4 / 8 / 16 starts | 3,092 / 3,084 / 3,068 / 3,196 / 3,451 s | 9.13 / 4.39 / 0.92 / 0.00 / 0.00% |
| C-only generator 16 + 1 F-only start | 1,565 s | 2.60% |
| C+phi generator 16 + 1 joint start | 1,566 s | 0.01% |
| HIFOO, best of 16 | 15.3 s | 50.75% |
| C-only kNN / C+phi kNN, b = 16 | 9.5 s / 7.6 s | 10.19% / 11.32% |
| C-only twin / C+phi twin, b = 16 | 1.9 s / 17.2 s | 20.27% / 32.47% |

- **Compute ratio ([38]):** C-only, the F-only optimizer at 8 starts (1,582 s) against the generator's 11.5 s: 138 times. C+phi, the joint optimizer at 2 starts (3,084 s) against the generator's 29.1 s: 106 times.
- **Why the C+phi ratio is only somewhat smaller:** the joint optimizer needs 2 starts against 8, but the optimizer's cost is nearly flat in the number of starts (the k starts run in one vectorized call, X24's convention; the F-only optimizer at 1 start ranges from 1,503 to 1,610 s over the timing plants, wider than the differences between start counts). The joint optimizer is also charged the F-only optimizer it starts from, which doubles its cost. And the C+phi generator costs 2.5 times C-only's: sampling takes the same time (median 8.7 s for 16 samples), but scoring C+phi candidates takes a median 20.4 s against 2.8 s, from 10 to 110 s in total over the timing plants.
- **Batched on the GPU** (sampling only, 199 plants in one call, `timing_batched.json`): generator 0.019 s per plant for both variants, twin 0.003 s, kNN retrieval 0.013 s (C-only) and 0.005 s (C+phi).
- The optimizer tasks ran 6 at a time on 8 cores (load about 6); the MIMO timings ran one plant at a time (load 0.7 to 1.0). Ratios are within one setting, so this does not change them, but absolute seconds across the two settings are not measured under the same load.

## 4. Conclusion
- **SISO:** C+phi and C-only generators tie at 16 samples on the common J*, as X28 S4 found, and both have lower median regret than kNN, the regression twins and the flat references. C+phi's own optimizer is stronger (2 joint starts reach the C+phi generator's regret; the F-only optimizer needs 8 to reach C-only's), but each optimizer costs about the same at any start count, so the compute ratio moves only from 138 times (C-only) to 106 times (C+phi). At 1 and 4 samples C+phi certifies fewer plants.
- **MIMO spline setting:** C+phi has lower median regret than C-only for each method family at 16 samples (difference of median regrets at m = 3 and m = 2: generator 3.7 and 2.8 points, kNN 7.9 and 4.7, twin 3.0 and 1.5), at equal certified share with repair. C-only certifies more at 1 sample. The C+phi generator at 16 samples has lower median regret than C-only's per-plant Adam optimizer at m = 2 (3.59% against 5.07%), not at m = 3 (6.56% against 4.33%), at about 1/1,500 of its CPU per plant on one thread.
- **Actuator setting:** generating phi gains nothing at the per-plant optimum. Step 3B is not run (Matteo's decision, 2026-11-17): the actuator slides draw C+phi at the C-only values, labeled "= C-only, not trained separately", with the 3A numbers as the reason.

**Deviations and caveats**
- SISO C-only twin rows are X21's stored outputs re-scored on J*. The re-sampled seed 0 missed its 1-point self-check on the certified share per candidate (+1.57) because the MC-dropout draws differ across machines; this was registered before any twin regret (`PREREG_siso.md`, 19:05 UTC). The re-sampled rows agree within 1 point at b = 16 (85.76% / 20.73%).
- The C+phi SISO twins reached their best validation loss early (step 4,000 of 60,000), so part of their gap to the generator may come from the training run.
- The actuator cost first used the trapezoid rule, which failed the tr(P) self-check at every horizon (median 0.25%); it was replaced by RK4 on an extra cost state before any 3A number (`PREREG.md` addendum 18:20 UTC).
- Actuator step 3B was planned only if Matteo asked for it after 3A, and was not run (`PREREG.md` addendum 2026-11-17). The C+phi actuator values on the deck are C-only's values, not a measurement.
- The joint optimizer and the C+phi Adam optimizer define J* in their settings, so their regret falls to 0 by construction.
- Regrets here are on each setting's common J* and are higher than earlier C-only numbers on J_Fonly (SISO) or J_ref (actuator); the two scales are not mixed.
- The MIMO spline J*_T2 uses id and so3 labels only, so T2 regrets are not on the scale of the held-out X28 regrets.
- The SISO and MIMO timings ran on two different nodes (move-hpc-06-gpu and move-hpc-08-gpu, 8 cores and an NVIDIA L40S each), each with nothing else of this project running. Ratios are computed within one setting's timing job.

**Sources:** the PREREG and addenda in `results/icml_r3/X29/`; the report files named in each section; HF commits 8f92ad1d (SISO), d7d0258 (MIMO spline setting), 09f35d5 (actuator 3A) and 374b6ec (SISO CPU, X27 re-score, this write-up).