analysis: localize early-layer feedback bottlenecks

6 files changed, 343 insertions, 4 deletions

diff --git a/RESULTS.md b/RESULTS.md
index 4fc6793..50377ba 100644
--- a/RESULTS.md
+++ b/RESULTS.md
@@ -663,6 +663,26 @@ per full hidden field, whereas structured targets are scored per channel-basis
 moment. No empirical 1133x variance-reduction claim is made from their raw
 ratio.
 
+A no-training oracle audit then separates feedback capacity from estimator
+efficiency on disjoint 32-example BatchNorm batches. Early-third alignment is
+`0.05494` when each example/channel receives unconstrained optimal coefficients
+for the existing `[1,tanh(h)]` spatial fields, but falls to `0.02397` when those
+coefficients must be predicted by the actual output-error-linear A/G family.
+The learned structured estimator reaches `0.00721`, about 30% of that
+cross-validated family oracle. Thus the current basis is not incapable of
+crossing the v2 threshold, but causal regression leaves a substantial gap and
+the output-error-only coefficient map itself discards over half of the spatial
+oracle's directional alignment.
+
+Adding fixed local-average and channel-mean somatic fields does not help
+(`0.02246` early-third cross-validated alignment). An independent spatial
+template reaches `0.04797`, but its prediction/target energy ratio is `0.458`
+despite low cosine, a warning that the high-capacity map amplifies
+out-of-sample error rather than solving credit assignment cleanly. These are
+post-failure oracle diagnostics, not trainable-method results. They rule out a
+naive basis expansion and point toward both better causal sample efficiency
+and a more informative hierarchical/high-level feedback context.
+
 ## How to run
 `experiments/run.py --mode {bp,fa,dfa,sdil} --dataset {mnist,fmnist,cifar10} --depth D --residual {0,1} --act {tanh,gelu,silu,relu}`
 Batteries: `experiments/run_v2.sh <ds> "<depths>" <res> <act> "<seeds>" <ep> <pfx>`.
diff --git a/REVIEW_SCORECARD.md b/REVIEW_SCORECARD.md
index 846cb22..3bbb0c8 100644
--- a/REVIEW_SCORECARD.md
+++ b/REVIEW_SCORECARD.md
@@ -74,6 +74,7 @@ retroactively reopened by success on standard vision benchmarks.
 | 2026-07-22 / frozen Oral-A A1--A3 | A1 and A2 pass; full A3 SDIL becomes nonfinite and fails four of six checks; A4 untouched | 5 → 5 | Closes the standard-scale question negatively. The narrow mechanism paper survives, while any standard-ResNet or oral claim does not |
 | 2026-07-22 / native C4 | BurstCCN and Dual Prop author-code records pass the strict audit; Dual Prop reproduces 92.46% test in 23119.8 s | 5 → 5 | Closes a baseline-fidelity objection and confirms a strong expensive comparator, but does not repair SDIL's failed A3 evidence |
 | 2026-07-22 / Oral-A-v2 V2-1 | Six clean frozen-forward records: structured calibration raises early/all-layer alignment to 0.0072/0.0527 but fails two frozen advancement checks | 5 → 5 | Confirms representable-subspace variance was real, while showing that early-layer causal credit remains below the standard-depth gate; V2-2 and confirmation stay closed |
+| 2026-07-22 / post-failure representation oracle | Current family has a 0.0240 cross-validated early-alignment ceiling on the fixed probe versus 0.0072 learned; unconstrained basis coefficients reach 0.0549 | 5 → 5 | Localizes both a causal-regression gap and an output-error-only capacity gap, but an oracle audit supplies no task-performance evidence |
 
 Future rows are appended only after an audited frozen stage. A score staying flat is informative:
 engineering, theory exposition, or visualization may make the paper more defensible without
diff --git a/ROADMAP.md b/ROADMAP.md
index 75e606a..2ad919f 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -364,6 +364,15 @@ was not launched; test and confirmation seeds remain untouched. This localizes
 the remaining bottleneck to early-layer credit rather than merely spatial
 projection variance.
 
+The subsequent no-training representation oracle further decomposes that
+bottleneck. The existing spatial fields reach `0.05494` early alignment with
+unconstrained per-example coefficients; the actual output-error-conditioned
+family reaches `0.02397` on a disjoint 32-example batch, while learned v2
+reaches `0.00721`. Naive local-average/channel-context fields fall slightly to
+`0.02246`. The next algorithmic branch should therefore address both causal
+sample efficiency and feedback context; simply lowering `eta_A`, adding the
+tested fields, or reopening V2-2 is not justified.
+
 Prepare convolutional local-update primitives and ResNet-20/32/56 protocols early. Queue frozen
 runs opportunistically on authorized idle GPUs. Because BurstCCN already reports CIFAR-10 and
 ImageNet scaling, dataset scale alone is not novel. The oral-level target is a memorable joint
diff --git a/experiments/analyze_oral_a_failure.py b/experiments/analyze_oral_a_failure.py
index bf68d12..ce2e2b0 100644
--- a/experiments/analyze_oral_a_failure.py
+++ b/experiments/analyze_oral_a_failure.py
@@ -26,6 +26,9 @@ def main():
     parser.add_argument(
         "--v2_gate", default="results/oral_a_v2_calibration_gate.json")
     parser.add_argument(
+        "--representation",
+        default="results/oral_a_representation_diagnosis.json")
+    parser.add_argument(
         "--out", default="results/oral_a_failure_diagnosis.json")
     args = parser.parse_args()
 
@@ -33,6 +36,7 @@ def main():
     short = load(args.short)
     gate = load(args.gate)
     v2_gate = load(args.v2_gate)
+    representation = load(args.representation)
     expected = {
         "mode": "sdil", "depth": 20, "epochs": 200,
         "vectorizer_mode": "channel_gated", "a_scale": 1.0,
@@ -48,6 +52,8 @@ def main():
     if (v2_gate["status"] != "failed"
             or v2_gate["confirmation_test_seeds_touched"]):
         raise ValueError("expected a failed v2 gate with untouched confirmation")
+    if representation["probe"]["test_examples_touched"] != 0:
+        raise ValueError("representation diagnosis touched test examples")
 
     calibration_rows = []
     first_nonfinite_epoch = None
@@ -94,7 +100,7 @@ def main():
         "protocol": "oral_a_A3_failure_diagnosis_v2",
         "source_paths": {
             "full": args.full, "short": args.short, "gate": args.gate,
-            "v2_gate": args.v2_gate,
+            "v2_gate": args.v2_gate, "representation": args.representation,
         },
         "source_commits": {
             "full": full["provenance"]["git_commit"],
@@ -135,13 +141,29 @@ def main():
                 structured["metrics"]["all_layer_alignment"]),
             "structured_gate_status": v2_gate["status"],
         },
+        "post_failure_representation_refinement": {
+            "trained_structured_early_third_alignment": (
+                structured["metrics"]["early_third_alignment"]),
+            "channel_gated_cv_oracle_early_third_alignment": (
+                representation["cosine"]["channel_gated_cv"]
+                ["early_third_mean"]),
+            "per_example_spatial_oracle_early_third_alignment": (
+                representation["cosine"]["per_example_spatial_oracle"]
+                ["early_third_mean"]),
+            "naive_local_context_cv_early_third_alignment": (
+                representation["cosine"]["local_context_cv"]
+                ["early_third_mean"]),
+        },
         "diagnosis": {
-            "classification": "unit_target_regression_snr_inadequate_before_runaway",
+            "classification": (
+                "early_credit_limited_by_estimator_efficiency_and_feedback_capacity"),
             "evidence": [
                 "instantaneous prediction-target cosine remains effectively zero",
                 "calibration MSE remains indistinguishable from stochastic target power",
                 "target power grows by orders of magnitude before nonfiniteness",
                 "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",
             ],
             "interpretation_limit": (
                 "instantaneous target metrics alone do not prove zero conditional "
@@ -150,6 +172,10 @@ def main():
             "next_test_outcome": (
                 "representable-subspace perturbation improved exact alignment but "
                 "failed the frozen early-layer causal-capture gate"),
+            "next_design_constraint": (
+                "do not add the tested naive local-average/channel-mean bases; "
+                "improve causal sample efficiency and use a more informative "
+                "hierarchical or high-level contextual feedback signal"),
         },
     }
     os.makedirs(os.path.dirname(os.path.abspath(args.out)), exist_ok=True)
diff --git a/results/oral_a_failure_diagnosis.json b/results/oral_a_failure_diagnosis.json
index 5239e8b..1b142df 100644
--- a/results/oral_a_failure_diagnosis.json
+++ b/results/oral_a_failure_diagnosis.json
@@ -44,14 +44,17 @@
     "target_power_growth_epoch1_to_87": 2126827049.3932126
   },
   "diagnosis": {
-    "classification": "unit_target_regression_snr_inadequate_before_runaway",
+    "classification": "early_credit_limited_by_estimator_efficiency_and_feedback_capacity",
     "evidence": [
       "instantaneous prediction-target cosine remains effectively zero",
       "calibration MSE remains indistinguishable from stochastic target power",
       "target power grows by orders of magnitude before nonfiniteness",
-      "matched frozen-forward unit targets later yield only 0.0011 early-layer alignment"
+      "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"
     ],
     "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; improve causal sample efficiency and use a more informative hierarchical or high-level contextual feedback signal",
     "next_test_outcome": "representable-subspace perturbation improved exact alignment but failed the frozen early-layer causal-capture gate"
   },
   "frozen_outcome": {
@@ -72,6 +75,12 @@
       "loss": null
     }
   },
+  "post_failure_representation_refinement": {
+    "channel_gated_cv_oracle_early_third_alignment": 0.02396580002561188,
+    "naive_local_context_cv_early_third_alignment": 0.022457713873793972,
+    "per_example_spatial_oracle_early_third_alignment": 0.05494018487684554,
+    "trained_structured_early_third_alignment": 0.007209055746595065
+  },
   "post_failure_v2_refinement": {
     "structured_all_layer_alignment": 0.05273995646520665,
     "structured_early_third_alignment": 0.007209055746595065,
@@ -93,6 +102,7 @@
   "source_paths": {
     "full": "results/oral_a_dev/sdil_full_r20_s0.json",
     "gate": "results/oral_a_full_gate.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_representation_diagnosis.json b/results/oral_a_representation_diagnosis.json
new file mode 100644
index 0000000..3c170c1
--- /dev/null
+++ b/results/oral_a_representation_diagnosis.json
@@ -0,0 +1,273 @@
+{
+  "cosine": {
+    "channel_gated_cv": {
+      "all_layer_mean": 0.26860839708781414,
+      "early_third_mean": 0.02396580002561188,
+      "per_layer": [
+        0.011761176795650953,
+        0.01509529795660324,
+        0.02110834681304405,
+        0.02648881734378023,
+        0.035379675557930944,
+        0.03396148568666185,
+        0.04737202878125393,
+        0.08733494968419549,
+        0.11412488007354467,
+        0.11111145838920782,
+        0.16484829973665457,
+        0.1705562681820029,
+        0.22815721287819002,
+        0.401197490658692,
+        0.5625015542517685,
+        0.5668841474011743,
+        0.7651559832644279,
+        0.7405204712136877,
+        0.999999999999998
+      ]
+    },
+    "local_context_cv": {
+      "all_layer_mean": 0.2657763621141079,
+      "early_third_mean": 0.022457713873793972,
+      "per_layer": [
+        0.011563062975842375,
+        0.013967016700256895,
+        0.019497958665444695,
+        0.025315213917818324,
+        0.03192140408252579,
+        0.032481626900875755,
+        0.04422040389025903,
+        0.08206046492345852,
+        0.11003918030189887,
+        0.10851704119726613,
+        0.1613744067292345,
+        0.167258052602825,
+        0.22673659297189666,
+        0.39712579796193304,
+        0.5577716450313811,
+        0.5599644189104344,
+        0.7632643600906147,
+        0.7366722323140862,
+        0.9999999999999978
+      ]
+    },
+    "per_example_spatial_oracle": {
+      "all_layer_mean": 0.3080300858793031,
+      "early_third_mean": 0.05494018487684554,
+      "per_layer": [
+        0.044284993661542316,
+        0.050289163913581,
+        0.054174960251254504,
+        0.056859289797845186,
+        0.06385232305504676,
+        0.06018037858180342,
+        0.07359532477033978,
+        0.1412692499078675,
+        0.16052161434051526,
+        0.15835344551911612,
+        0.20160746652301614,
+        0.20989211288493392,
+        0.2581259209822714,
+        0.4772294334388024,
+        0.6118273095794977,
+        0.6367033659113905,
+        0.7892394590882331,
+        0.8045658194996996,
+        1.0000000000000022
+      ]
+    },
+    "spatial_template_cv": {
+      "all_layer_mean": 0.2988933125374179,
+      "early_third_mean": 0.04796723142241297,
+      "per_layer": [
+        0.018794834494927203,
+        0.028104944902399245,
+        0.03589688012133678,
+        0.04943569370504626,
+        0.06487004742738658,
+        0.09070098788338177,
+        0.10403720218754312,
+        0.12806722415370214,
+        0.15473667864101381,
+        0.2100427890594658,
+        0.251471876051357,
+        0.29367024000272923,
+        0.3516688770876262,
+        0.36271913833514663,
+        0.48618596240775314,
+        0.5601553963417819,
+        0.7149574547592554,
+        0.77345671064909,
+        0.9999999999999973
+      ]
+    }
+  },
+  "interpretation": {
+    "channel_gated_cv": "actual v2 vectorizer family fit on 32 examples and evaluated on 32 disjoint examples",
+    "local_context_cv": "diagnostic four-field family adding local spatial average and cross-channel somatic context",
+    "per_example_spatial_oracle": "upper bound for the two [1,tanh(h)] fields with unconstrained per-example/channel coefficients",
+    "spatial_template_cv": "output-error linear map with an independent coefficient at every hidden unit",
+    "status": "post_failure_diagnosis_not_a_learning_result_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": {
+    "channel_gated_cv": {
+      "all_layer_mean": 0.18153600823060526,
+      "early_third_mean": 0.002141493001862695,
+      "per_layer": [
+        0.0013493838251117957,
+        0.0018821866339346622,
+        0.001696421773313066,
+        0.001957989772753797,
+        0.002779034484324995,
+        0.003183941521737851,
+        0.004221306553680124,
+        0.016149756754974574,
+        0.021920567376265642,
+        0.02088059044942729,
+        0.0365591553150065,
+        0.0398603168872378,
+        0.0655479631539235,
+        0.21288315903697133,
+        0.36735600776651145,
+        0.39189274768920795,
+        0.6184532017563165,
+        0.6406104266618678,
+        0.9999999989689331
+      ]
+    },
+    "local_context_cv": {
+      "all_layer_mean": 0.19307878388438796,
+      "early_third_mean": 0.0034804679550285975,
+      "per_layer": [
+        0.002811484345442635,
+        0.0030526487624646416,
+        0.0029164019568946173,
+        0.0033475132507695907,
+        0.004236770387682787,
+        0.004517989026917315,
+        0.005602177120465608,
+        0.020577920086456147,
+        0.026298887646854674,
+        0.026779671005441024,
+        0.04293551458003673,
+        0.046946976827938976,
+        0.07304993437673593,
+        0.25110846414727345,
+        0.39983404016800456,
+        0.4403499134572238,
+        0.6400042870848195,
+        0.6741263003721254,
+        0.999999999199824
+      ]
+    },
+    "per_example_spatial_oracle": {
+      "all_layer_mean": 0.18659338577180437,
+      "early_third_mean": 0.0030446135866956468,
+      "per_layer": [
+        0.0019590616394833775,
+        0.0024993103774871434,
+        0.002925141382872601,
+        0.0031998702722740354,
+        0.004054197420553255,
+        0.0036301004275034673,
+        0.005254602524406208,
+        0.019885848610928476,
+        0.025921735678432582,
+        0.024952912936957894,
+        0.04109786003346419,
+        0.04484580079931531,
+        0.06838933561711134,
+        0.2341008355593867,
+        0.3779443971902448,
+        0.4102824117515338,
+        0.6264783686368899,
+        0.6478525388054378,
+        1.0
+      ]
+    },
+    "spatial_template_cv": {
+      "all_layer_mean": 0.578876932893507,
+      "early_third_mean": 0.457647070414877,
+      "per_layer": [
+        0.4541521195275675,
+        0.47896384166837286,
+        0.448121193317785,
+        0.44093294951602213,
+        0.45548109382727925,
+        0.4682312246322352,
+        0.47915529592237,
+        0.48488773706669025,
+        0.49772630973811416,
+        0.5115627889679282,
+        0.5303103557041301,
+        0.556252488968008,
+        0.5816302553670893,
+        0.6011432068873668,
+        0.6639513115023615,
+        0.7067060521061842,
+        0.7903486506759895,
+        0.8491048527148057,
+        0.9999999968663323
+      ]
+    }
+  },
+  "probe": {
+    "batchnorm_cv": "fit and evaluation halves use separate 32-example batch statistics and exact-gradient graphs",
+    "evaluation_examples": 32,
+    "examples": 64,
+    "fit_examples": 32,
+    "source": "unaugmented first training-prefix examples",
+    "test_examples_touched": 0
+  },
+  "protocol": "oral_a_post_failure_representation_diagnosis_v1",
+  "provenance": {
+    "git_commit": "ba50c8fc286f6071c688a12606216cfb871a7ee0",
+    "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"
+  }
+}