paper v2.31: matched 30-epoch BP/DFA controls (was unsourced 0.609/0.308)

The §5 ¶3 BP-no-penalty value of 0.609 ± 0.004 and DFA-no-penalty value
of 0.308 ± 0.014 turned out to be unsourced — they were carried over
from a hardcoded comment in experiments/bp_with_penalty_control.py
("BP-trainable (3-seed mean): 0.609") that nobody had actually measured
with a matched 30-epoch run.

Ran the missing matched controls under the same recipe as BP+pen
(lam=0, 30 epochs, AdamW 1e-3, wd 0.01, cosine schedule, batch 128,
3 seeds 42/123/456):

  BP no-pen 30ep: per-seed 0.5851, 0.5845, 0.5863  →  0.585 ± 0.001
                  (paper said 0.609 ± 0.004, off by 0.024)
  DFA no-pen 30ep: per-seed 0.3070, 0.2985, 0.2966 →  0.301 ± 0.005
                  (paper said 0.308 ± 0.014)

Also re-grounded DFA+penalty 30ep using the dfa_pen_short 3-seed run
(0.3593, 0.3610, 0.3604 → 0.360 ± 0.001), which is what the deep-cosine
+0.155 figure was computed on. The paper had 0.363 ± 0.001 — that came
from the 100-epoch run, not the 30-epoch run, so it was an apples-to-
oranges comparison with BP+pen 30-ep.

Paper changes (§5 ¶3):
  BP penalty cost:  -8 pp  →  -5.5 pp
  DFA pen rescue:   +5.5 → +5.9 pp
  DFA+pen margin vs frozen: +1.4 → +1.1 pp
  BP-to-DFA gap:     17 → 17.0 pp (unchanged)
  BP-to-SB gap:      7.7 → 7.7 pp (unchanged)
  BP-to-DFA gap is still the lower-bound credit-quality cost claim;
  17 pp gap is unchanged in magnitude.

Also updated:
- §5 ¶1 prose: 0.363 → 0.360, 0.308 → 0.301
- §4 ¶4 prose: DFA+pen 0.363 → 0.360
- Appendix J Table 9 caption: 0.363 → 0.360, +9.0 → +9.3 pp gap to SB
- Appendix L paragraph: +5.5 → +5.9 pp DFA penalty rescue
- Figure 3 panel C bar values + title pen-cost annotation
- New results/matched_30ep_control_summary.json as auditable record

Page layout preserved: 9 main pages + refs p10, 18 total, 0 overfull.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

7 files changed, 841 insertions, 0 deletions

diff --git a/results/bp_no_penalty_30ep/bp_pen_lam0.0_s123.json b/results/bp_no_penalty_30ep/bp_pen_lam0.0_s123.json
new file mode 100644
index 0000000..26065b0
--- /dev/null
+++ b/results/bp_no_penalty_30ep/bp_pen_lam0.0_s123.json
@@ -0,0 +1,42 @@
+{
+  "config": {
+    "seed": 123,
+    "epochs": 30,
+    "lr": 0.001,
+    "wd": 0.01,
+    "lam": 0.0,
+    "output_dir": "results/bp_no_penalty_30ep"
+  },
+  "final_acc": 0.5845,
+  "log": [
+    {
+      "epoch": 1,
+      "test_acc": 0.3958
+    },
+    {
+      "epoch": 5,
+      "test_acc": 0.4851
+    },
+    {
+      "epoch": 10,
+      "test_acc": 0.5345
+    },
+    {
+      "epoch": 15,
+      "test_acc": 0.5564
+    },
+    {
+      "epoch": 20,
+      "test_acc": 0.5748
+    },
+    {
+      "epoch": 25,
+      "test_acc": 0.5827
+    },
+    {
+      "epoch": 30,
+      "test_acc": 0.5845
+    }
+  ],
+  "margin_pp": 23.550000000000004
+}
\ No newline at end of file
diff --git a/results/bp_no_penalty_30ep/bp_pen_lam0.0_s42.json b/results/bp_no_penalty_30ep/bp_pen_lam0.0_s42.json
new file mode 100644
index 0000000..b55aee7
--- /dev/null
+++ b/results/bp_no_penalty_30ep/bp_pen_lam0.0_s42.json
@@ -0,0 +1,42 @@
+{
+  "config": {
+    "seed": 42,
+    "epochs": 30,
+    "lr": 0.001,
+    "wd": 0.01,
+    "lam": 0.0,
+    "output_dir": "results/bp_no_penalty_30ep"
+  },
+  "final_acc": 0.5851,
+  "log": [
+    {
+      "epoch": 1,
+      "test_acc": 0.392
+    },
+    {
+      "epoch": 5,
+      "test_acc": 0.495
+    },
+    {
+      "epoch": 10,
+      "test_acc": 0.5296
+    },
+    {
+      "epoch": 15,
+      "test_acc": 0.5539
+    },
+    {
+      "epoch": 20,
+      "test_acc": 0.5721
+    },
+    {
+      "epoch": 25,
+      "test_acc": 0.5813
+    },
+    {
+      "epoch": 30,
+      "test_acc": 0.5851
+    }
+  ],
+  "margin_pp": 23.61
+}
\ No newline at end of file
diff --git a/results/bp_no_penalty_30ep/bp_pen_lam0.0_s456.json b/results/bp_no_penalty_30ep/bp_pen_lam0.0_s456.json
new file mode 100644
index 0000000..de092b2
--- /dev/null
+++ b/results/bp_no_penalty_30ep/bp_pen_lam0.0_s456.json
@@ -0,0 +1,42 @@
+{
+  "config": {
+    "seed": 456,
+    "epochs": 30,
+    "lr": 0.001,
+    "wd": 0.01,
+    "lam": 0.0,
+    "output_dir": "results/bp_no_penalty_30ep"
+  },
+  "final_acc": 0.5863,
+  "log": [
+    {
+      "epoch": 1,
+      "test_acc": 0.3969
+    },
+    {
+      "epoch": 5,
+      "test_acc": 0.4957
+    },
+    {
+      "epoch": 10,
+      "test_acc": 0.5344
+    },
+    {
+      "epoch": 15,
+      "test_acc": 0.5612
+    },
+    {
+      "epoch": 20,
+      "test_acc": 0.5716
+    },
+    {
+      "epoch": 25,
+      "test_acc": 0.5845
+    },
+    {
+      "epoch": 30,
+      "test_acc": 0.5863
+    }
+  ],
+  "margin_pp": 23.730000000000008
+}
\ No newline at end of file
diff --git a/results/bp_nopen_30ep_3seed.log b/results/bp_nopen_30ep_3seed.log
new file mode 100644
index 0000000..f88e725
--- /dev/null
+++ b/results/bp_nopen_30ep_3seed.log
@@ -0,0 +1,58 @@
+=== BP no-pen 30ep seed=42 ===
+BP + ‖f‖² penalty: seed=42, lam=0.0, epochs=30
+  ep 1: test_acc=0.3920
+  ep 5: test_acc=0.4950
+  ep 10: test_acc=0.5296
+  ep 15: test_acc=0.5539
+  ep 20: test_acc=0.5721
+  ep 25: test_acc=0.5813
+  ep 30: test_acc=0.5851
+
+FINAL test acc: 0.5851
+Compare to:
+  BP-trainable (3-seed mean):   0.609
+  Penalized DFA lam=1e-2:       0.363
+  DFA-shallow:                  0.349
+
+Margin vs DFA-shallow baseline: +23.61 pp
+  → BP+penalty intermediate; partial capacity loss + residual mode 2
+Saved results/bp_no_penalty_30ep/bp_pen_lam0.0_s42.json
+=== BP no-pen 30ep seed=123 ===
+BP + ‖f‖² penalty: seed=123, lam=0.0, epochs=30
+  ep 1: test_acc=0.3958
+  ep 5: test_acc=0.4851
+  ep 10: test_acc=0.5345
+  ep 15: test_acc=0.5564
+  ep 20: test_acc=0.5748
+  ep 25: test_acc=0.5827
+  ep 30: test_acc=0.5845
+
+FINAL test acc: 0.5845
+Compare to:
+  BP-trainable (3-seed mean):   0.609
+  Penalized DFA lam=1e-2:       0.363
+  DFA-shallow:                  0.349
+
+Margin vs DFA-shallow baseline: +23.55 pp
+  → BP+penalty intermediate; partial capacity loss + residual mode 2
+Saved results/bp_no_penalty_30ep/bp_pen_lam0.0_s123.json
+=== BP no-pen 30ep seed=456 ===
+BP + ‖f‖² penalty: seed=456, lam=0.0, epochs=30
+  ep 1: test_acc=0.3969
+  ep 5: test_acc=0.4957
+  ep 10: test_acc=0.5344
+  ep 15: test_acc=0.5612
+  ep 20: test_acc=0.5716
+  ep 25: test_acc=0.5845
+  ep 30: test_acc=0.5863
+
+FINAL test acc: 0.5863
+Compare to:
+  BP-trainable (3-seed mean):   0.609
+  Penalized DFA lam=1e-2:       0.363
+  DFA-shallow:                  0.349
+
+Margin vs DFA-shallow baseline: +23.73 pp
+  → BP+penalty intermediate; partial capacity loss + residual mode 2
+Saved results/bp_no_penalty_30ep/bp_pen_lam0.0_s456.json
+=== ALL DONE ===
diff --git a/results/dfa_no_penalty_30ep/results_cifar10.json b/results/dfa_no_penalty_30ep/results_cifar10.json
new file mode 100644
index 0000000..b6607bf
--- /dev/null
+++ b/results/dfa_no_penalty_30ep/results_cifar10.json
@@ -0,0 +1,549 @@
+{
+  "42": {
+    "dfa": {
+      "log": {
+        "train_loss": [
+          2.041171367111206,
+          2.01035458278656,
+          2.0089993319702146,
+          2.0087824996566774,
+          2.005417049255371,
+          2.004202005844116,
+          2.004135177154541,
+          2.0013649726104736,
+          2.0012059982299806,
+          2.002799745788574,
+          1.999473673324585,
+          1.9972123648834228,
+          1.9990896060180665,
+          1.9961500553894043,
+          1.9974410807800294,
+          1.997062333984375,
+          1.9967610731506347,
+          1.9935473257827758,
+          1.9938662228393556,
+          1.9911954043197633,
+          1.9918721434783935,
+          1.9906449029922486,
+          1.9902287648773194,
+          1.9899413529968262,
+          1.9901034127044677,
+          1.9888135684204102,
+          1.9867314432144165,
+          1.98775391166687,
+          1.9877222427749635,
+          1.986023156890869
+        ],
+        "train_acc": [
+          0.25212,
+          0.26684,
+          0.2673,
+          0.26968,
+          0.2702,
+          0.27032,
+          0.27152,
+          0.27198,
+          0.27294,
+          0.27146,
+          0.27256,
+          0.27506,
+          0.27462,
+          0.27678,
+          0.27724,
+          0.27692,
+          0.27528,
+          0.2775,
+          0.27616,
+          0.27902,
+          0.28,
+          0.28058,
+          0.28046,
+          0.28064,
+          0.27876,
+          0.27924,
+          0.2785,
+          0.28286,
+          0.2799,
+          0.28246
+        ],
+        "test_acc": [
+          0.2899,
+          0.3039,
+          0.297,
+          0.2862,
+          0.2876,
+          0.2963,
+          0.2947,
+          0.302,
+          0.29,
+          0.2862,
+          0.2878,
+          0.3067,
+          0.2853,
+          0.3071,
+          0.2948,
+          0.2992,
+          0.309,
+          0.298,
+          0.303,
+          0.3047,
+          0.3106,
+          0.3006,
+          0.3082,
+          0.2972,
+          0.3017,
+          0.3064,
+          0.305,
+          0.3064,
+          0.3061,
+          0.307
+        ]
+      },
+      "diagnostics": {
+        "bp_cosine": [
+          0.4062161147594452,
+          0.02404380589723587,
+          -0.01314563024789095,
+          -0.02864420786499977
+        ],
+        "perturbation_rho": [
+          0.02938488870859146,
+          0.0,
+          0.0,
+          -0.04040054604411125
+        ],
+        "nudging": {
+          "0.001": [
+            -1.2400560081005096e-06,
+            -9.313225746154785e-10,
+            9.313225746154785e-10,
+            9.313225746154785e-10
+          ],
+          "0.003": [
+            -3.570225089788437e-06,
+            -4.6566128730773926e-09,
+            1.862645149230957e-09,
+            3.725290298461914e-09
+          ],
+          "0.01": [
+            -1.1967029422521591e-05,
+            -1.210719347000122e-08,
+            1.3969838619232178e-09,
+            2.7939677238464355e-09
+          ]
+        },
+        "hidden_norms_per_layer": [
+          12514.5302734375,
+          10006967.0,
+          24731472.0,
+          48831016.0,
+          61899120.0
+        ],
+        "bp_grad_norms_per_layer": [
+          1.1740979743990465e-06,
+          7.62120411224032e-09,
+          7.4135684258180845e-09,
+          7.434150628427005e-09,
+          7.419798997432281e-09
+        ]
+      },
+      "drift": {
+        "embed.weight": 106.47232673325593,
+        "embed.bias": 88.63566200531763,
+        "blocks.0.ln.weight": 4.505657196044922,
+        "blocks.0.w1.weight": 71.1190136511018,
+        "blocks.0.w1.bias": 65.27901182999418,
+        "blocks.0.w2.weight": 191.81664513772796,
+        "blocks.1.ln.weight": 3.699037790298462,
+        "blocks.1.w1.weight": 76.41161958621439,
+        "blocks.1.w1.bias": 75.09969389083356,
+        "blocks.1.w2.weight": 132.93780551811736,
+        "blocks.2.ln.weight": 3.838078737258911,
+        "blocks.2.w1.weight": 96.89058587597209,
+        "blocks.2.w1.bias": 96.42063073285577,
+        "blocks.2.w2.weight": 155.6341879797901,
+        "blocks.3.ln.weight": 3.6865062713623047,
+        "blocks.3.w1.weight": 91.12525240918814,
+        "blocks.3.w1.bias": 89.48750762910116,
+        "blocks.3.w2.weight": 144.11887720845996,
+        "out_ln.weight": 0.4002464711666107,
+        "out_head.weight": 4.1392195557143445,
+        "out_head.bias": 0.3521037802404755
+      }
+    }
+  },
+  "123": {
+    "dfa": {
+      "log": {
+        "train_loss": [
+          2.043456584434509,
+          2.015362589454651,
+          2.011732778701782,
+          2.0096331211090086,
+          2.008639169998169,
+          2.006175118484497,
+          2.0041568740081788,
+          2.0016602905273437,
+          1.9992998320770263,
+          1.9955639269638061,
+          1.997627447128296,
+          1.9954585150146484,
+          1.9922321643829346,
+          1.9908589698028565,
+          1.99274623046875,
+          1.987565859451294,
+          1.9855359014892577,
+          1.9876045575714112,
+          1.9856640274047852,
+          1.9843628873062134,
+          1.983776489906311,
+          1.9830582943344117,
+          1.9843373564910889,
+          1.9821271927642823,
+          1.9814285857391358,
+          1.9822291962051393,
+          1.9806529276275635,
+          1.9811888027191162,
+          1.981866326751709,
+          1.9813806748199463
+        ],
+        "train_acc": [
+          0.2482,
+          0.2599,
+          0.26138,
+          0.26296,
+          0.26512,
+          0.26796,
+          0.2672,
+          0.2701,
+          0.27262,
+          0.27384,
+          0.27248,
+          0.27358,
+          0.27466,
+          0.2741,
+          0.2742,
+          0.27818,
+          0.27998,
+          0.27902,
+          0.27798,
+          0.27852,
+          0.2805,
+          0.2829,
+          0.27792,
+          0.28224,
+          0.28406,
+          0.28142,
+          0.28238,
+          0.2844,
+          0.28192,
+          0.28278
+        ],
+        "test_acc": [
+          0.2683,
+          0.2831,
+          0.2816,
+          0.2732,
+          0.2737,
+          0.2935,
+          0.2862,
+          0.2764,
+          0.2816,
+          0.2943,
+          0.2797,
+          0.2884,
+          0.2935,
+          0.2935,
+          0.2918,
+          0.2941,
+          0.2962,
+          0.2928,
+          0.2944,
+          0.2986,
+          0.292,
+          0.301,
+          0.2926,
+          0.3011,
+          0.2988,
+          0.298,
+          0.2972,
+          0.2948,
+          0.2994,
+          0.2985
+        ]
+      },
+      "diagnostics": {
+        "bp_cosine": [
+          0.4047228693962097,
+          0.04104449227452278,
+          -0.011393947526812553,
+          -0.009042292833328247
+        ],
+        "perturbation_rho": [
+          0.0367465503513813,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "nudging": {
+          "0.001": [
+            -1.1739321053028107e-06,
+            -9.313225746154785e-09,
+            0.0,
+            0.0
+          ],
+          "0.003": [
+            -3.5855919122695923e-06,
+            1.862645149230957e-09,
+            0.0,
+            -9.313225746154785e-10
+          ],
+          "0.01": [
+            -1.1938624083995819e-05,
+            -6.984919309616089e-09,
+            -7.450580596923828e-09,
+            -1.3969838619232178e-09
+          ]
+        },
+        "hidden_norms_per_layer": [
+          13284.138671875,
+          7033498.5,
+          39633128.0,
+          64930576.0,
+          66113756.0
+        ],
+        "bp_grad_norms_per_layer": [
+          1.1239109198868391e-06,
+          6.197617086201035e-09,
+          5.381650236557789e-09,
+          5.38055022758499e-09,
+          5.385091483844917e-09
+        ]
+      },
+      "drift": {
+        "embed.weight": 110.02475615987971,
+        "embed.bias": 68.65394532089363,
+        "blocks.0.ln.weight": 4.497658729553223,
+        "blocks.0.w1.weight": 70.59434597015027,
+        "blocks.0.w1.bias": 60.73836185221871,
+        "blocks.0.w2.weight": 193.2499470830069,
+        "blocks.1.ln.weight": 3.926071882247925,
+        "blocks.1.w1.weight": 92.10013575370482,
+        "blocks.1.w1.bias": 84.79144794996425,
+        "blocks.1.w2.weight": 160.80728575641794,
+        "blocks.2.ln.weight": 4.058661937713623,
+        "blocks.2.w1.weight": 100.55769394683692,
+        "blocks.2.w1.bias": 95.75447672997026,
+        "blocks.2.w2.weight": 159.847282269534,
+        "blocks.3.ln.weight": 2.8910417556762695,
+        "blocks.3.w1.weight": 61.4594216054129,
+        "blocks.3.w1.bias": 50.547957978671036,
+        "blocks.3.w2.weight": 99.05733623792827,
+        "out_ln.weight": 0.373835951089859,
+        "out_head.weight": 3.990617544364531,
+        "out_head.bias": 0.4551460205356602
+      }
+    }
+  },
+  "456": {
+    "dfa": {
+      "log": {
+        "train_loss": [
+          2.0538753495788575,
+          2.0328780548095704,
+          2.0271591313171387,
+          2.0249926682281494,
+          2.0220700995635985,
+          2.02148416885376,
+          2.018033724822998,
+          2.017232068099976,
+          2.0104196131134033,
+          2.011538788986206,
+          2.007870107727051,
+          2.009203465423584,
+          2.0076511237335204,
+          2.0058005546569824,
+          2.0033703365707396,
+          2.0026689519500733,
+          2.0036011488342287,
+          2.001192732009888,
+          1.9991068982315063,
+          1.9989625353240967,
+          1.997356651878357,
+          1.997875876312256,
+          1.9966684870147704,
+          1.9943733695983887,
+          1.9942921846008301,
+          1.9950503900146483,
+          1.995271727142334,
+          1.9939517930603028,
+          1.9949346865844726,
+          1.9955234482192994
+        ],
+        "train_acc": [
+          0.24646,
+          0.25364,
+          0.25594,
+          0.25786,
+          0.26122,
+          0.2595,
+          0.26164,
+          0.26184,
+          0.26632,
+          0.26368,
+          0.26816,
+          0.26846,
+          0.2676,
+          0.26972,
+          0.26778,
+          0.26982,
+          0.27154,
+          0.27196,
+          0.27306,
+          0.2728,
+          0.2749,
+          0.27498,
+          0.274,
+          0.27632,
+          0.27516,
+          0.27554,
+          0.274,
+          0.27776,
+          0.27618,
+          0.27674
+        ],
+        "test_acc": [
+          0.2591,
+          0.2786,
+          0.2747,
+          0.2853,
+          0.2705,
+          0.2732,
+          0.274,
+          0.2744,
+          0.2837,
+          0.2887,
+          0.2899,
+          0.2931,
+          0.2939,
+          0.2822,
+          0.2828,
+          0.2954,
+          0.2961,
+          0.2902,
+          0.2943,
+          0.2941,
+          0.2972,
+          0.2971,
+          0.2955,
+          0.2942,
+          0.2995,
+          0.2923,
+          0.3007,
+          0.2965,
+          0.2971,
+          0.2966
+        ]
+      },
+      "diagnostics": {
+        "bp_cosine": [
+          0.3774259090423584,
+          -0.00713147409260273,
+          -0.007417748216539621,
+          0.0045123836025595665
+        ],
+        "perturbation_rho": [
+          -0.001135501079261303,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "nudging": {
+          "0.001": [
+            -9.578652679920197e-07,
+            -9.313225746154785e-10,
+            9.313225746154785e-10,
+            0.0
+          ],
+          "0.003": [
+            -2.8796494007110596e-06,
+            -2.3283064365386963e-09,
+            9.313225746154785e-10,
+            -3.725290298461914e-09
+          ],
+          "0.01": [
+            -9.520910680294037e-06,
+            -3.725290298461914e-09,
+            2.3283064365386963e-09,
+            -9.313225746154785e-10
+          ]
+        },
+        "hidden_norms_per_layer": [
+          13071.4912109375,
+          11935114.0,
+          44992308.0,
+          80702592.0,
+          99592856.0
+        ],
+        "bp_grad_norms_per_layer": [
+          9.842210602073465e-07,
+          4.690550614583344e-09,
+          4.549102428086371e-09,
+          4.550036347694686e-09,
+          4.553873722557e-09
+        ]
+      },
+      "drift": {
+        "embed.weight": 112.63263586193521,
+        "embed.bias": 120.50774359993062,
+        "blocks.0.ln.weight": 4.934469223022461,
+        "blocks.0.w1.weight": 79.98242400105649,
+        "blocks.0.w1.bias": 71.77394207010907,
+        "blocks.0.w2.weight": 207.96745983686054,
+        "blocks.1.ln.weight": 3.650259494781494,
+        "blocks.1.w1.weight": 92.54979942916204,
+        "blocks.1.w1.bias": 86.83660111158788,
+        "blocks.1.w2.weight": 164.60563519153237,
+        "blocks.2.ln.weight": 4.2977728843688965,
+        "blocks.2.w1.weight": 116.55584641362735,
+        "blocks.2.w1.bias": 109.68765194306474,
+        "blocks.2.w2.weight": 173.76733209355965,
+        "blocks.3.ln.weight": 4.123578071594238,
+        "blocks.3.w1.weight": 106.41492266556777,
+        "blocks.3.w1.bias": 102.54546146383908,
+        "blocks.3.w2.weight": 164.09091953653103,
+        "out_ln.weight": 0.4199577569961548,
+        "out_head.weight": 4.122513850074516,
+        "out_head.bias": 0.3386189609154552
+      }
+    }
+  },
+  "config": {
+    "dataset": "cifar10",
+    "d_hidden": 256,
+    "num_blocks": 4,
+    "batch_size": 128,
+    "epochs": 30,
+    "lr": 0.001,
+    "lr_fb": 0.001,
+    "wd": 0.01,
+    "lam": 0.1,
+    "K": 4,
+    "sigma_bridge": 0.05,
+    "ema_momentum": 0.995,
+    "term_grad_weight": 1.0,
+    "seeds": [
+      42,
+      123,
+      456
+    ],
+    "gpu": 0,
+    "output_dir": "results/dfa_no_penalty_30ep",
+    "methods": [
+      "dfa"
+    ],
+    "random_targets": false,
+    "penalty_lam": 0.0,
+    "num_classes": 10
+  }
+}
\ No newline at end of file
diff --git a/results/dfa_nopen_30ep_3seed.log b/results/dfa_nopen_30ep_3seed.log
new file mode 100644
index 0000000..ec99d88
--- /dev/null
+++ b/results/dfa_nopen_30ep_3seed.log
@@ -0,0 +1,37 @@
+Using device: cuda:0
+
+============================================================
+Seed 42
+============================================================
+
+--- DFA ---
+  [DFA] Epoch 1: loss=2.0412, train=0.2521, test=0.2899
+  [DFA] Epoch 10: loss=2.0028, train=0.2715, test=0.2862
+  [DFA] Epoch 20: loss=1.9912, train=0.2790, test=0.3047
+  [DFA] Epoch 30: loss=1.9860, train=0.2825, test=0.3070
+  Final test acc: 0.3070
+
+============================================================
+Seed 123
+============================================================
+
+--- DFA ---
+  [DFA] Epoch 1: loss=2.0435, train=0.2482, test=0.2683
+  [DFA] Epoch 10: loss=1.9956, train=0.2738, test=0.2943
+  [DFA] Epoch 20: loss=1.9844, train=0.2785, test=0.2986
+  [DFA] Epoch 30: loss=1.9814, train=0.2828, test=0.2985
+  Final test acc: 0.2985
+
+============================================================
+Seed 456
+============================================================
+
+--- DFA ---
+  [DFA] Epoch 1: loss=2.0539, train=0.2465, test=0.2591
+  [DFA] Epoch 10: loss=2.0115, train=0.2637, test=0.2887
+  [DFA] Epoch 20: loss=1.9990, train=0.2728, test=0.2941
+  [DFA] Epoch 30: loss=1.9955, train=0.2767, test=0.2966
+  Final test acc: 0.2966
+
+All results saved to results/dfa_no_penalty_30ep/results_cifar10.json
+=== DONE ===
diff --git a/results/matched_30ep_control_summary.json b/results/matched_30ep_control_summary.json
new file mode 100644
index 0000000..23acca3
--- /dev/null
+++ b/results/matched_30ep_control_summary.json
@@ -0,0 +1,71 @@
+{
+  "matched_30ep_control": {
+    "config": {
+      "arch": "4-block d=256 pre-LN ResMLP",
+      "dataset": "CIFAR-10",
+      "epochs": 30,
+      "batch": 128,
+      "optimizer": "AdamW lr=1e-3 wd=0.01",
+      "schedule": "cosine",
+      "penalty": "lam*mean(||f_l(h_l)||^2) on residual branches"
+    },
+    "bp_no_pen_3seed": {
+      "per_seed": {
+        "42": 0.5851,
+        "123": 0.5845,
+        "456": 0.5863
+      },
+      "mean": 0.5853,
+      "std_ddof0": 0.00075,
+      "std_ddof1": 0.00092,
+      "source_files": [
+        "results/bp_no_penalty_30ep/bp_pen_lam0.0_s{42,123,456}.json"
+      ]
+    },
+    "bp_with_pen_s42": {
+      "acc": 0.5303,
+      "lam": 0.01,
+      "source": "results/bp_with_penalty/bp_pen_lam0.01_s42.json",
+      "note": "single-seed (s42 only); v2.30 polish identified the matched 30-ep no-pen 3-seed control was missing"
+    },
+    "dfa_no_pen_3seed": {
+      "per_seed": {
+        "42": 0.307,
+        "123": 0.2985,
+        "456": 0.2966
+      },
+      "mean": 0.3007,
+      "std_ddof0": 0.0045,
+      "std_ddof1": 0.0055,
+      "source_files": [
+        "results/dfa_no_penalty_30ep/results_cifar10.json"
+      ]
+    },
+    "dfa_with_pen_3seed": {
+      "per_seed": {
+        "42": 0.3593,
+        "123": 0.361,
+        "456": 0.3604
+      },
+      "mean": 0.3602,
+      "std_ddof0": 0.0007,
+      "std_ddof1": 0.0009,
+      "source_files": [
+        "results/dfa_pen_short/dfa_pen_lam0.01_s{42,123,456}.json"
+      ],
+      "lam": 0.01
+    },
+    "penalty_costs": {
+      "bp_pen_cost_pp": 5.5,
+      "dfa_pen_rescue_pp": 5.9
+    },
+    "margins_vs_frozen_0.349": {
+      "bp_no_pen_pp": 23.6,
+      "bp_with_pen_pp": 18.1,
+      "dfa_no_pen_pp": -4.8,
+      "dfa_with_pen_pp": 1.1,
+      "sb_with_pen_pp": 10.4,
+      "cb_with_pen_pp": 1.1
+    }
+  }
+}
\ No newline at end of file