--- 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).