paper: report audited standard-depth scaling

3 files changed, 226 insertions, 31 deletions

diff --git a/paper/CLAIM_LEDGER.json b/paper/CLAIM_LEDGER.json
index 93fcc90..ca0f900 100644
--- a/paper/CLAIM_LEDGER.json
+++ b/paper/CLAIM_LEDGER.json
@@ -4,7 +4,8 @@
     "../results/figs/figure2_scaling.png",
     "../results/figs/figure3_innovation.png",
     "../results/figs/figure4_resnet_confirmation.png",
-    "../results/figs/figure5_bci_v2.png"
+    "../results/figs/figure5_bci_v2.png",
+    "../results/figs/figure6_standard_depth_scaling.png"
   ],
   "gate_statuses": [
     {
@@ -18,6 +19,10 @@
     {
       "expected": "passed",
       "source": "results/bci_v2_calibrated_confirmation_gate.json"
+    },
+    {
+      "expected": "passed",
+      "source": "results/oral_a_dynamic_scaling_v2_gate.json"
     }
   ],
   "manuscript": "paper/MANUSCRIPT.md",
@@ -212,6 +217,76 @@
       "token": "1.47"
     },
     {
+      "format": "fixed3_percent",
+      "id": "oral_a_dynamic_d32_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/dynamic/32",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "92.254%"
+    },
+    {
+      "format": "fixed3_percent",
+      "id": "oral_a_dynamic_d56_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/dynamic/56",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "92.760%"
+    },
+    {
+      "format": "fixed3_percent",
+      "id": "oral_a_bp_d56_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/bp/56",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "92.632%"
+    },
+    {
+      "format": "fixed3_percent",
+      "id": "oral_a_clean_kp_d56_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/clean_kp/56",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "92.670%"
+    },
+    {
+      "format": "fixed3_percent",
+      "id": "oral_a_dfa_d56_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/dfa/56",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "30.850%"
+    },
+    {
+      "format": "fixed3",
+      "id": "oral_a_dynamic_depth_gain",
+      "pointer": "/statistics/dynamic_d20_to_d56_gain_points_mean",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "1.176"
+    },
+    {
+      "format": "fixed3",
+      "id": "oral_a_dynamic_dfa_d56_gap",
+      "pointer": "/statistics/dynamic_minus_dfa_points_mean/56",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "61.910"
+    },
+    {
+      "format": "fixed6",
+      "id": "oral_a_dynamic_d32_alignment",
+      "pointer": "/statistics/dynamic_early_alignment_mean/32",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "0.999613"
+    },
+    {
+      "format": "fixed6",
+      "id": "oral_a_dynamic_d56_alignment",
+      "pointer": "/statistics/dynamic_early_alignment_mean/56",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "0.999423"
+    },
+    {
+      "format": "fixed3",
+      "id": "oral_a_dynamic_d56_mac_ratio",
+      "pointer": "/statistics/dynamic_mac_ratio_to_bp_mean/56",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "token": "1.331"
+    },
+    {
       "format": "percent3_percent",
       "id": "r2_final_performance",
       "pointer": "/metrics/intact_final/mean",
@@ -384,7 +459,7 @@
     "Neither mechanism is claimed as new.",
     "we do not infer that cortex implements BP;",
     "we do not claim arbitrary top-down traffic removal;",
-    "It does not support a ResNet-20-to-56 depth claim.",
+    "The four-method panel does not establish dominance over the broader local-learning literature.",
     "The joint biological gate nevertheless fails.",
     "It does not establish that the same plasticity rule operates in cortex.",
     "we do not call this variant single-phase."
diff --git a/paper/MANUSCRIPT.md b/paper/MANUSCRIPT.md
index 733284e..b92a074 100644
--- a/paper/MANUSCRIPT.md
+++ b/paper/MANUSCRIPT.md
@@ -24,21 +24,23 @@ that subtraction changes direction rather than merely gain. In a frozen
 10.31% MNIST accuracy under strong soma-predictable traffic, whereas innovation
 retains 97.348%. Across a 12-fold increase in hidden depth on flattened
 CIFAR-10, the learned-feedback backbone changes by -0.214 points while direct
-feedback alignment loses early-layer alignment. On a standard ResNet-20, a
-dynamic paired-neutral innovation rule reaches 91.584% mean CIFAR-10 test
-accuracy across five untouched seeds, versus 91.388% for clean reciprocal
-credit; the one-sided 95% upper bound on its deficit is 0.131 points. It uses
-zero task-loss queries and 1.326 times the matched BP MAC estimate, but pays for
-one instruction-off neutral observation per training example. In a separate
-six-task by five-model synthetic BCI confirmation, a local actor--critic
-innovation reaches 100.000% task success, 99.833% terminal outcome decoding,
-and 30/30 predicted causal-role signs. An acute outcome lesion reduces
-role-aligned separation by 0.400. The supported conclusion remains
-algorithmic: somato-dendritic innovation can protect local credit from
-soma-predictable traffic, remain stable on ResNet-20, and multiplex performance
+feedback alignment loses early-layer alignment. Across a frozen 60-record
+standard ResNet-20/32/56 panel, a dynamic paired-neutral innovation rule rises
+from 91.584% to 92.254% and 92.760% mean CIFAR-10 test accuracy. All five
+paired seeds improve from depth 20 to 56, with a 1.176-point mean gain; at
+ResNet-56 it matches BP (92.632%) and clean reciprocal credit (92.670%) while
+retaining 0.999423 early-third teaching alignment. Its MAC estimate remains at
+most 1.331 times matched BP, but it pays for one instruction-off neutral
+observation per training example. In a separate six-task by five-model
+synthetic BCI confirmation, a local actor--critic innovation reaches 100.000%
+task success, 99.833% terminal outcome decoding, and 30/30 predicted
+causal-role signs. An acute outcome lesion reduces role-aligned separation by
+0.400. The supported conclusion remains algorithmic: somato-dendritic
+innovation can protect local credit from soma-predictable traffic, gain
+accuracy with standard ResNet depth on CIFAR-10, and multiplex performance
 change with outcome surprise in a controlled dynamical task. These results do
-not establish a cortical learning rule or positive utility from added
-standard-network depth.
+not establish a cortical learning rule or broad superiority over
+local-learning alternatives.
 
 ## 1. Introduction
 
@@ -87,8 +89,8 @@ Our contributions are:
 2. a conditional-projection analysis, a descent/gain boundary, and explicit
    perturbation variance and resource accounting;
 3. frozen evidence that the innovation operation is load-bearing under
-   soma-predictable traffic, including an independently confirmed ResNet-20
-   endpoint; and
+   soma-predictable traffic, including a complete ResNet-20/32/56 panel with
+   positive paired depth gains; and
 4. an independently confirmed local actor--critic instantiation in which
    learned causal roles vectorize performance change and outcome surprise,
    together with retained failed protocols that delimit the result.
@@ -312,6 +314,14 @@ and dynamic innovation from scratch for untouched seeds 10--14, uses no
 validation examples, and evaluates the 10,000-example test set once at the
 endpoint. Both conditions use the same initialization/data seed pairing.
 
+After the independently frozen biological prerequisite passed, a separate
+standard-depth protocol reused those ten records and opened exactly 50 new
+endpoints: BP and DFA at ResNet-20, and BP, DFA, clean reciprocal credit, and
+dynamic innovation at ResNet-32 and ResNet-56, all for seeds 10--14. The
+resulting 60-record panel uses the same 200 epochs, 50,000 training examples,
+one final test evaluation, and paired initialization/data seeds at every
+depth. Hyperparameters are copied across depth without depth-specific tuning.
+
 ### Synthetic BCI confirmation
 
 The biological-signature test uses 40 cells with known experimenter-only
@@ -420,8 +430,36 @@ The dynamic run uses 1.326 times the matched BP affine-MAC estimate, zero
 logical task-loss queries, one paired neutral observation per ordinary example,
 and 2.02 GiB mean peak allocated memory. On the same GTX 1080s its mean wall
 time is 1.47 times paired clean KP. The result supports ResNet-20 robustness
-under the audited traffic intervention. It does not support a ResNet-20-to-56
-depth claim.
+under the audited traffic intervention. The ResNet-20 result alone does not
+support a ResNet-20-to-56 depth claim; that claim is tested only by the
+separately frozen panel below.
+
+### 5.4 Dynamic innovation gains accuracy with standard ResNet depth
+
+![Accuracy, paired depth gain, alignment, and cost across standard ResNet depth.](../results/figs/figure6_standard_depth_scaling.png)
+
+**Figure 6: Standard-depth scaling.** The renderer reads and validates all 60
+raw records, verifies their hashes against the passed gate, and binds the gate
+to the historical training-source revision. Error bars are 95% normal
+intervals over the five paired seeds.
+
+Dynamic innovation improves from 91.584% at ResNet-20 to 92.254% at ResNet-32
+and 92.760% at ResNet-56. Every paired seed improves from depth 20 to 56, for a
+mean gain of 1.176 points. At ResNet-56, its mean accuracy is slightly above
+the 92.632% BP and 92.670% clean-reciprocal endpoints. Fixed DFA reaches only
+30.850% at the same depth, a 61.910-point gap, while dynamic innovation's mean
+early-third alignment remains 0.999613 at ResNet-32 and 0.999423 at
+ResNet-56.
+
+The affine-MAC estimate grows proportionally with the matched forward model:
+the dynamic-to-BP ratio is 1.331 at ResNet-56, below the predeclared 1.34
+ceiling. This accounting still excludes no phase: the method uses zero
+task-loss queries but one explicitly counted instruction-off neutral
+observation per example. The panel establishes positive standard-depth
+scaling for the reciprocal-feedback SDIL instantiation. Because it contains
+only BP, fixed DFA, clean reciprocal credit, and dynamic SDIL, it does not yet
+establish dominance over Forward--Forward, PEPITA, equilibrium propagation,
+Dual Propagation, or ordinary FA at matched depths.
 
 ## 6. Biological-signature test and outcome-surprise evidence
 
@@ -498,7 +536,7 @@ These negatives constrain interpretation:
 - we do not infer that dendritic residuals directly drive cortical plasticity;
 - we do not infer that cortex implements BP;
 - we do not claim arbitrary top-down traffic removal;
-- we do not claim positive utility from adding standard ResNet depth; and
+- we do not claim that every competing local rule fails to scale; and
 - we do not treat directly supplied terminal reward as an emergent error.
 
 ## 7. Related work
@@ -545,7 +583,8 @@ feedback, a two-state contrast, or an input perturbation.
 The central novelty is narrow. Conditional projection is standard, and both
 instructional pathways are inherited. The empirical contribution is that
 subtraction is load-bearing under a controlled mixed-channel intervention and
-that the dynamic version remains stable at ResNet-20 scale.
+that the dynamic version remains stable and gains accuracy from ResNet-20 to
+ResNet-56.
 
 The predictor's information class is a hard boundary. A diagonal per-cell
 predictor cannot remove population-only or latent traffic. Neutral periods are
@@ -559,11 +598,14 @@ MAC overhead. Hardware implementations may price local elementwise operations,
 state storage, and phases differently from GPUs; this is why we report several
 resource axes rather than one scalar cost.
 
-The strongest standard result uses only ResNet-20. The old separately frozen
-ResNet-20/32/56 panel remains unopened because its original biological
-prerequisite failed. The successful v2 gate permits only a new independently
-frozen depth protocol; it cannot retroactively open the old panel. Positive
-added-depth utility therefore remains unresolved.
+The standard-depth result is complete but narrow: one CIFAR-10 architecture
+family, five seeds, and four methods. Its clean-task scaling belongs
+substantially to the inherited reciprocal Kolen--Pollack substrate; the
+SDIL-specific evidence is preservation under controlled soma-predictable
+traffic. The four-method panel does not establish dominance over the broader
+local-learning literature. Matched Forward--Forward, PEPITA, equilibrium
+propagation, Dual Propagation, ordinary FA, and additional architecture
+families are required to support that stronger claim.
 
 Finally, the synthetic BCI evidence has a hard ecological boundary. Outcome
 reward is supplied to the critic, the psychometric targets are calibrated to
@@ -586,4 +628,5 @@ bash experiments/finalize_accept.sh
 It rechecks the main figures, theoretical identities, local-rule mechanics,
 baseline protocols, native-author records, standard-ResNet confirmation,
 failed biological protocols, the passed calibrated BCI confirmation, and the
-sealed old standard-depth boundary.
+passed 60-record standard-depth panel while preserving the sealed old
+standard-depth boundary.
diff --git a/paper/manuscript_audit.json b/paper/manuscript_audit.json
index 55ee55a..fe67713 100644
--- a/paper/manuscript_audit.json
+++ b/paper/manuscript_audit.json
@@ -163,6 +163,66 @@
       "source": "results/figs/figure4_resnet_confirmation_manifest.json"
     },
     {
+      "id": "oral_a_dynamic_d32_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/dynamic/32",
+      "rendered": "92.254%",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
+      "id": "oral_a_dynamic_d56_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/dynamic/56",
+      "rendered": "92.760%",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
+      "id": "oral_a_bp_d56_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/bp/56",
+      "rendered": "92.632%",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
+      "id": "oral_a_clean_kp_d56_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/clean_kp/56",
+      "rendered": "92.670%",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
+      "id": "oral_a_dfa_d56_accuracy",
+      "pointer": "/statistics/test_accuracy_mean_percent/dfa/56",
+      "rendered": "30.850%",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
+      "id": "oral_a_dynamic_depth_gain",
+      "pointer": "/statistics/dynamic_d20_to_d56_gain_points_mean",
+      "rendered": "1.176",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
+      "id": "oral_a_dynamic_dfa_d56_gap",
+      "pointer": "/statistics/dynamic_minus_dfa_points_mean/56",
+      "rendered": "61.910",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
+      "id": "oral_a_dynamic_d32_alignment",
+      "pointer": "/statistics/dynamic_early_alignment_mean/32",
+      "rendered": "0.999613",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
+      "id": "oral_a_dynamic_d56_alignment",
+      "pointer": "/statistics/dynamic_early_alignment_mean/56",
+      "rendered": "0.999423",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
+      "id": "oral_a_dynamic_d56_mac_ratio",
+      "pointer": "/statistics/dynamic_mac_ratio_to_bp_mean/56",
+      "rendered": "1.331",
+      "source": "results/figs/figure6_standard_depth_scaling_manifest.json"
+    },
+    {
       "id": "r2_final_performance",
       "pointer": "/metrics/intact_final/mean",
       "rendered": "99.531%",
@@ -327,6 +387,10 @@
     {
       "path": "results/figs/figure5_bci_v2.png",
       "sha256": "e43e07d56c28022e861682e7b7f68ef010e2754c64eb58cfc8079d8a59d57b75"
+    },
+    {
+      "path": "results/figs/figure6_standard_depth_scaling.png",
+      "sha256": "6d6dc2b752f8cf0f3868cc7801f721662720bc50eec6c960f5063b427c634adf"
     }
   ],
   "gates": [
@@ -352,16 +416,21 @@
       "false_checks": [],
       "source": "results/bci_v2_calibrated_confirmation_gate.json",
       "status": "passed"
+    },
+    {
+      "false_checks": [],
+      "source": "results/oral_a_dynamic_scaling_v2_gate.json",
+      "status": "passed"
     }
   ],
   "ledger": {
     "path": "paper/CLAIM_LEDGER.json",
-    "sha256": "bd65f0747aba2dbe327f7150b7112dbc560c6e9cb984b058343679fd0dd89ab9"
+    "sha256": "fed71c8f22742840b314bc41b91cd38fa81c6d14339c051e30741e5506dc9519"
   },
   "manuscript": {
     "path": "paper/MANUSCRIPT.md",
-    "sha256": "c0633b8ff9f7b267e12eade47fac7b01429b6dc9b9255b71938e14169ca7822f",
-    "word_count": 4019
+    "sha256": "20eb3c331f1caed4cb3ae0406d14fe43f0d1d59adf8debfdfea93662566ee7e9",
+    "word_count": 4408
   },
   "sources": [
     {
@@ -381,12 +450,20 @@
       "sha256": "dc24f1593b7e07b33b45e0e1dfec895cf7bdc5a953e60d933c5c3596f284738e"
     },
     {
+      "path": "results/figs/figure6_standard_depth_scaling_manifest.json",
+      "sha256": "fc7e15dbad052a944a11b25bfb944843dd6c17e4f2defa82c4089b14be6d27c9"
+    },
+    {
       "path": "results/figs/main_figure_manifest.json",
       "sha256": "f5099b29d7f83ad7dc7783b927f682d7af43a90072e11fc7da2e8be5e7685450"
     },
     {
       "path": "results/kp_dynamic_projection_confirmation_gate.json",
       "sha256": "636c587e47287338cdca7d9558dc3bfb99a2cec615badb23337025d85b8128ef"
+    },
+    {
+      "path": "results/oral_a_dynamic_scaling_v2_gate.json",
+      "sha256": "c0a4c0530dea26c0fffc7285e7893e6e321ba248d1e13abefbafad311a89b44f"
     }
   ],
   "status": "passed",