analysis: identify hierarchical spatial credit bottleneck

2 files changed, 248 insertions, 2 deletions

diff --git a/results/oral_a_failure_diagnosis.json b/results/oral_a_failure_diagnosis.json
index b26d81e..79feb33 100644
--- a/results/oral_a_failure_diagnosis.json
+++ b/results/oral_a_failure_diagnosis.json
@@ -52,10 +52,11 @@
       "matched frozen-forward unit targets later yield only 0.0011 early-layer alignment",
       "structured learning reaches 0.0072 versus a 0.0240 cross-validated family oracle",
       "unconstrained coefficients raise the same spatial basis oracle only to 0.0549",
-      "direct A/G estimation leaves early alignment unchanged at 0.0071"
+      "direct A/G estimation leaves early alignment unchanged at 0.0071",
+      "held-out gated 3x3 maps from exact child errors reach 0.9998"
     ],
     "interpretation_limit": "instantaneous target metrics alone do not prove zero conditional learning; structured v2 has low instantaneous cosine yet positive exact-gradient alignment",
-    "next_design_constraint": "do not add the tested naive local-average/channel-mean bases; do not repeat the closed A/G rate grid; use a more informative hierarchical or high-level contextual feedback signal and audit remaining cross-layer noise",
+    "next_design_constraint": "do not add the tested activation contexts or repeat the closed A/G rate grid; learn spatial hierarchical feedback without weight copying and audit against hierarchical FA/BurstCCN",
     "next_test_outcome": "representable-subspace perturbation improved alignment but failed; direct vectorizer-space estimation then improved only middle/late layers and also failed the frozen early-layer gate"
   },
   "frozen_outcome": {
@@ -76,6 +77,12 @@
       "loss": null
     }
   },
+  "post_failure_hierarchical_refinement": {
+    "direct_output_channel_gated": 0.031122580851844408,
+    "downstream_activation_context": 0.032068526988258954,
+    "hierarchical_1x1_gated": 0.50521034930192,
+    "hierarchical_3x3_gated": 0.9997667354251036
+  },
   "post_failure_representation_refinement": {
     "channel_gated_cv_oracle_early_third_alignment": 0.02396580002561188,
     "naive_local_context_cv_early_third_alignment": 0.022457713873793972,
@@ -110,6 +117,7 @@
   "source_paths": {
     "full": "results/oral_a_dev/sdil_full_r20_s0.json",
     "gate": "results/oral_a_full_gate.json",
+    "hierarchical_oracle": "results/oral_a_hierarchical_oracle.json",
     "representation": "results/oral_a_representation_diagnosis.json",
     "short": "results/oral_a_short/sdil_channel_gated_lr0.03.json",
     "v2_gate": "results/oral_a_v2_calibration_gate.json",
diff --git a/results/oral_a_hierarchical_oracle.json b/results/oral_a_hierarchical_oracle.json
new file mode 100644
index 0000000..545a492
--- /dev/null
+++ b/results/oral_a_hierarchical_oracle.json
@@ -0,0 +1,238 @@
+{
+  "cosine": {
+    "direct_output_channel_gated": {
+      "early_third_mean": 0.031122580851844408,
+      "per_layer": [
+        0.014921543495746436,
+        0.019894676282936723,
+        0.0282271051699598,
+        0.03533547855645916,
+        0.04383726627872673,
+        0.0445194153272376
+      ]
+    },
+    "downstream_activation_context": {
+      "early_third_mean": 0.032068526988258954,
+      "per_layer": [
+        0.015172893156451723,
+        0.02093329244734575,
+        0.03032952347243866,
+        0.036240372930205864,
+        0.043887546626221505,
+        0.04584753329689023
+      ]
+    },
+    "hierarchical_1x1": {
+      "early_third_mean": 0.44052470743135946,
+      "per_layer": [
+        0.46788420886369614,
+        0.3231655357294116,
+        0.5899724506620518,
+        0.3257336277626957,
+        0.6199696177846528,
+        0.3164228037856487
+      ]
+    },
+    "hierarchical_1x1_gated": {
+      "early_third_mean": 0.50521034930192,
+      "per_layer": [
+        0.5667462015924819,
+        0.3611245555418099,
+        0.6843671689570183,
+        0.3570338150608624,
+        0.7093727314925211,
+        0.35261762316682654
+      ]
+    },
+    "hierarchical_3x3": {
+      "early_third_mean": 0.8199172772242803,
+      "per_layer": [
+        0.7459063715286154,
+        0.8262732802772361,
+        0.793522765424401,
+        0.8599287817089369,
+        0.8139193036048705,
+        0.879953160801622
+      ]
+    },
+    "hierarchical_3x3_gated": {
+      "early_third_mean": 0.9997667354251036,
+      "per_layer": [
+        0.9993640919952334,
+        0.999799566383427,
+        0.9999004439090606,
+        0.9998297253276784,
+        0.9998818831563995,
+        0.9998247017788229
+      ]
+    }
+  },
+  "fit": {
+    "hierarchical_relative_ridge": 1e-06,
+    "ridge_role": "fixed numerical conditioning, not endpoint selection"
+  },
+  "interpretation": {
+    "direct_output_channel_gated": "current output-error-only family",
+    "downstream_activation_context": "current family plus ordinary next-node activation map",
+    "hierarchical": "local linear feedback from exact child error fields following the true residual DAG; gated variants multiply by child ReLU state",
+    "status": "post_failure_oracle_not_a_trainable_method_or_gate"
+  },
+  "network": {
+    "depth": 20,
+    "forward_state": "deterministic_initialization_no_updates",
+    "normalization": "batchnorm",
+    "residual_scale": 1.0,
+    "seed": 0,
+    "width": 16
+  },
+  "prediction_energy_over_target_energy": {
+    "direct_output_channel_gated": {
+      "early_third_mean": 0.001537713589318464,
+      "per_layer": [
+        0.00048708346199335603,
+        0.0008494185513749271,
+        0.001249646051341693,
+        0.001762353268497944,
+        0.0023878024038859126,
+        0.00248997779881695
+      ]
+    },
+    "downstream_activation_context": {
+      "early_third_mean": 0.0017585368373614592,
+      "per_layer": [
+        0.0005894725033945322,
+        0.0011358088629614706,
+        0.0014933947759630931,
+        0.0019371272285281854,
+        0.002642552679428714,
+        0.0027528649738927592
+      ]
+    },
+    "hierarchical_1x1": {
+      "early_third_mean": 0.21093509531885588,
+      "per_layer": [
+        0.21368997518885788,
+        0.10759142591349577,
+        0.3482826378878522,
+        0.10767558816021679,
+        0.3861797202365838,
+        0.1021912245261288
+      ]
+    },
+    "hierarchical_1x1_gated": {
+      "early_third_mean": 0.2791378873040979,
+      "per_layer": [
+        0.3157016623047377,
+        0.1334361923772813,
+        0.4681671742151389,
+        0.1292411902995396,
+        0.5023064159214083,
+        0.12597468870648137
+      ]
+    },
+    "hierarchical_3x3": {
+      "early_third_mean": 0.6797097711665088,
+      "per_layer": [
+        0.5423770967465465,
+        0.6985505551295879,
+        0.6396710728617321,
+        0.7498588674920689,
+        0.6612808289892634,
+        0.7865202057798534
+      ]
+    },
+    "hierarchical_3x3_gated": {
+      "early_third_mean": 1.0027014963437466,
+      "per_layer": [
+        0.9733409516366273,
+        1.024496384636999,
+        1.0067263938527884,
+        1.008659614236897,
+        0.9967422111022111,
+        1.0062434225969583
+      ]
+    }
+  },
+  "probe": {
+    "evaluation_examples": 128,
+    "fit_examples": 128,
+    "layers": [
+      0,
+      1,
+      2,
+      3,
+      4,
+      5
+    ],
+    "residual_dag_children": [
+      [
+        1,
+        2
+      ],
+      [
+        2
+      ],
+      [
+        3,
+        4
+      ],
+      [
+        4
+      ],
+      [
+        5,
+        6
+      ],
+      [
+        6
+      ]
+    ],
+    "separate_batchnorm_graphs": true,
+    "source": "unaugmented first training-prefix examples",
+    "test_examples_touched": 0
+  },
+  "protocol": "oral_a_post_failure_hierarchical_oracle_v1",
+  "provenance": {
+    "git_commit": "45d2b57310d556c4d1e78ee185279eda14189585",
+    "git_tracked_dirty": false
+  },
+  "split": {
+    "dataset": "cifar10",
+    "input_layout": "NCHW",
+    "input_shape": [
+      3,
+      32,
+      32
+    ],
+    "loader_seed": 0,
+    "normalization_mean": [
+      0.49140000343322754,
+      0.4821999967098236,
+      0.4465000033378601
+    ],
+    "normalization_std": [
+      0.24699999392032623,
+      0.2434999942779541,
+      0.26159998774528503
+    ],
+    "split_from_training_only": true,
+    "split_seed": 2027,
+    "test_examples": 10000,
+    "train_examples": 10000,
+    "training_augmentation": "none",
+    "validation_class_counts": {
+      "0": 500,
+      "1": 500,
+      "2": 500,
+      "3": 500,
+      "4": 500,
+      "5": 500,
+      "6": 500,
+      "7": 500,
+      "8": 500,
+      "9": 500
+    },
+    "validation_examples": 5000,
+    "validation_index_sha256": "8328b206a97c420e49e54e3eca4abe3274c4756b084355784ea3fb8059e4515b"
+  }
+}