diff options
| -rw-r--r-- | ORAL_A_V6.md | 29 | ||||
| -rw-r--r-- | RESULTS.md | 14 | ||||
| -rw-r--r-- | results/oral_a_v6_calibration/result.json | 529 | ||||
| -rw-r--r-- | results/oral_a_v6_calibration_gate.json | 178 |
4 files changed, 750 insertions, 0 deletions
diff --git a/ORAL_A_V6.md b/ORAL_A_V6.md index 7663f27..6e597f4 100644 --- a/ORAL_A_V6.md +++ b/ORAL_A_V6.md @@ -94,3 +94,32 @@ recovery after the endpoint. Failure closes V6 and the present no-KP causal vectorizer route. Passing opens only a separately frozen short task gate; capture alone cannot raise the reviewer score. +## Audited outcome (2026-08-06) + +V6-1 fails on clean source revision `98bfda7`. The run executes the exact +frozen 19 stages, 380 edge events, 760 batch-loss queries, and 48,640 +per-example observations. All forward weights, readout weights, BatchNorm +state/affines, and forward optimizer tensors remain bitwise fixed; the local +fits use zero forward-weight reads and zero reverse-mode learning operations. + +The readout stage succeeds. Its final feedback/forward parameter cosine is +`0.81319`, and final-hidden teaching alignment is `0.87219`, compared with +`0.00377` and `0.07787` under fixed HFA. This confirms the post-failure V5 +diagnosis: per-example causal observations have enough signal to identify the +small dense feedback map when fitted appropriately. + +The convolutional diagonal approximation fails. Feedback norm ratios already +reach `10.17--49.08` on finite late edges. The predicted parent-field MSE +grows from `0.0258` at edge 18 to `578` at edge 16, `3.12e7` at edge 15, and +`6.84e30` at edge 9. Edge 9 and all earlier stages become nonfinite. The +child contexts are strongly correlated, so treating their covariance as +diagonal creates a multiplicative error cascade rather than a stable local +inverse. + +Consequently every alignment and norm gate fails, no task-accuracy run is +opened, and validation/test endpoints remain untouched. Per the frozen rule, +there is no ridge, clipping, or full-covariance recovery. V6 closes the +present task-loss-only no-KP causal vectorizer route. Recovering convolutional +feedback would require a materially different information source or an +expensive full local system-identification method, which would need to beat +the already measured KP, response-mirror, and BP cost/accuracy baselines. @@ -1147,6 +1147,20 @@ observations are informative near the output, not evidence that the no-KP vectorizer scales; the existing positive ResNet scaling result still belongs to the inherited KP substrate. +The final bounded no-KP follow-up, Oral-A-v6, tests whether strict stagewise +timing and locally whitened fits repair V5 without increasing its 760-query +budget. The dense readout fit works: feedback/forward cosine reaches +`0.81319` and last-hidden teaching alignment reaches `0.87219`. The +convolutional diagonal fit does not. Correlated child contexts make the +diagonal inverse unstable: finite feedback norm ratios reach `49.08`, parent +field MSE grows from `0.0258` at edge 18 to `6.84e30` at edge 9, and edge 9 +through the input become nonfinite. All forward and BatchNorm state remains +bitwise fixed, and no validation or test endpoint is evaluated. This closes +the present task-loss-only no-KP causal-vectorizer route. A full covariance +solver or local response measurement would be system identification; the +latter is already represented by the response-mirror baseline, while KP is +the substrate behind the positive standard-depth scaling result. + ## 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/results/oral_a_v6_calibration/result.json b/results/oral_a_v6_calibration/result.json new file mode 100644 index 0000000..223f8ff --- /dev/null +++ b/results/oral_a_v6_calibration/result.json @@ -0,0 +1,529 @@ +{ + "architecture": { + "adaptive_feedback_parameters": 267904, + "family": "CIFAR 6n+2 ResNet, option-A shortcuts", + "feedback_macs_per_example": 40108672, + "forward_macs_per_example": 40551040, + "forward_parameters": 269722 + }, + "finite": false, + "fixed_hfa": { + "all_layer_alignment": 0.011278868932430143, + "early_third_alignment": -0.0002631605093483813, + "feedback_forward_cosine": [ + 0.0008809715509414673, + -0.002320552244782448, + 0.03730485588312149, + 0.02497659996151924, + -0.03438456356525421, + -0.001257734838873148, + 0.023045765236020088, + 0.010684739798307419, + -0.013966654427349567, + -0.03158111870288849, + 0.009583890438079834, + -9.433389641344547e-05, + 0.004490766208618879, + -0.004784357734024525, + 0.004104882013052702, + 0.004543168004602194, + -0.005818712059408426, + 0.0014950999757274985, + 0.003767170011997223 + ], + "feedback_forward_norm_ratio": [ + 0.9996462464332581, + 1.0550107955932617, + 0.9593725800514221, + 1.0017613172531128, + 0.999921441078186, + 0.9631232619285583, + 0.9992203712463379, + 0.9976226687431335, + 1.0018895864486694, + 1.0030434131622314, + 1.0180330276489258, + 1.0107296705245972, + 0.9962885975837708, + 0.9891077876091003, + 1.0009011030197144, + 1.0021207332611084, + 1.0043002367019653, + 1.0021206140518188, + 0.9992877244949341 + ], + "max_feedback_forward_norm_ratio": 1.0550107955932617, + "mean_feedback_forward_cosine": 0.0016142042954207252, + "min_feedback_forward_norm_ratio": 0.9593725800514221, + "per_layer": [ + -0.0005718570901080966, + -0.0006265088450163603, + 0.0012021202128380537, + 3.7642239476554096e-05, + 0.00048483084538020194, + -0.0021051904186606407, + 0.00329246511682868, + 0.00027297029737383127, + 0.007543372455984354, + -0.0053173573687672615, + 0.009830434806644917, + -0.0010977087076753378, + 0.02368866838514805, + 0.007433340419083834, + 0.04359162598848343, + 0.0057456037029623985, + 0.039965368807315826, + 0.0030560996383428574, + 0.07787258923053741 + ] + }, + "hardware": { + "cuda_device_name": "NVIDIA GeForce GTX 1080", + "cuda_visible_devices": "5", + "device": "cuda", + "peak_memory_allocated_bytes": 1430980096, + "torch_version": "2.3.1+cu118" + }, + "learned_scib": { + "all_layer_alignment": NaN, + "early_third_alignment": NaN, + "feedback_forward_cosine": [ + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + -0.006127675995230675, + -0.011246882379055023, + 0.021258916705846786, + 0.00929059274494648, + -0.001609913888387382, + -0.0008538670372217894, + 0.012138523161411285, + 0.002846281975507736, + 0.052370935678482056, + 0.8131900429725647 + ], + "feedback_forward_norm_ratio": [ + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + 7.744247913360596, + 7.140369415283203, + 11.147778511047363, + 32.04150390625, + 49.08123016357422, + 48.10276412963867, + 33.58113479614258, + 10.170685768127441, + 12.821575164794922, + 1.120253562927246 + ], + "max_feedback_forward_norm_ratio": NaN, + "mean_feedback_forward_cosine": NaN, + "min_feedback_forward_norm_ratio": NaN, + "per_layer": [ + NaN, + 0.0, + NaN, + 0.0, + NaN, + 0.0, + NaN, + 0.0, + 0.0, + -0.018046237528324127, + 0.009331459179520607, + 0.006288703065365553, + -0.0008923492860049009, + 0.024693015962839127, + 0.04579439014196396, + 0.011086630634963512, + -0.025804810225963593, + 0.3799460828304291, + 0.8721908926963806 + ] + }, + "method_audit": { + "causal_query_normalization_state": "evaluation_running_statistics", + "forward_state_max_absolute_difference": 0.0, + "forward_weight_reads_in_feedback_fit": 0, + "ordinary_task_normalization_state": "not_run_forward_frozen", + "reverse_mode_learning_operations": 0, + "stage_order": [ + "readout", + 18, + 17, + 16, + 15, + 14, + 13, + 12, + 11, + 10, + 9, + 8, + 7, + 6, + 5, + 4, + 3, + 2, + 1 + ] + }, + "protocol": "oral_a_v6_stagewise_whitened_causal_capture_v1", + "provenance": { + "git_commit": "98bfda74f7ee0457799b86465c867b6943c7be91", + "git_tracked_dirty": false + }, + "schema_version": 1, + "settings": { + "alignment_probe": 64, + "batch_size": 128, + "calibration_augmentation": false, + "conv_diagonal_relative_ridge": 0.001, + "depth": 20, + "events_per_stage": 20, + "feedback_scale": 1.0, + "loader_seed": 0, + "normalization": "batchnorm", + "perturb_seed": 5000, + "readout_relative_ridge": 1e-06, + "residual_scale": 1.0, + "seed": 0, + "sigma": 0.01, + "split_seed": 2027, + "train_limit": 10000, + "val_examples": 5000, + "width": 16 + }, + "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" + }, + "stage_fits": [ + { + "absolute_ridge": 0.00039648901367187497, + "after_mse": 1.4543522596359253, + "before_mse": 1.5032594203948975, + "examples": 2560, + "fit_target_cosine": 0.14180169892473388, + "kind": "readout", + "parameter_rms": 0.13902018685545237, + "relative_ridge": 1e-06 + }, + { + "absolute_ridge": 0.0007461613416671753, + "before_field_mse": 0.025809165835380555, + "edge": 18, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": 0.7461613416671753, + "minimum_diagonal": 0.0009920651791617274, + "observations": 20, + "parameter_rms": 0.753869841492034, + "parameter_update_rms": 0.756108481982398, + "relative_ridge": 0.001, + "target_rms": 0.16062293037888375 + }, + { + "absolute_ridge": 4.24671240234375, + "before_field_mse": 0.08534659817814827, + "edge": 17, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": 4246.71240234375, + "minimum_diagonal": 0.2649068832397461, + "observations": 20, + "parameter_rms": 0.6007491967588365, + "parameter_update_rms": 0.6027629071724261, + "relative_ridge": 0.001, + "target_rms": 0.20353771929970688 + }, + { + "absolute_ridge": 38247.564, + "before_field_mse": 578.0022918701172, + "edge": 16, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": 38247564.0, + "minimum_diagonal": 0.0, + "observations": 20, + "parameter_rms": 1.9794660509112565, + "parameter_update_rms": 1.9810645055729732, + "relative_ridge": 0.001, + "target_rms": 0.21561053942828906 + }, + { + "absolute_ridge": 2282466050.048, + "before_field_mse": 31215376.8, + "edge": 15, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": 2282466050048.0, + "minimum_diagonal": 0.0, + "observations": 20, + "parameter_rms": 2.828870306284442, + "parameter_update_rms": 2.8293681916124838, + "relative_ridge": 0.001, + "target_rms": 0.2689296525317363 + }, + { + "absolute_ridge": 147347425844002.8, + "before_field_mse": 1678965394636.8, + "edge": 14, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": 1.473474258440028e+17, + "minimum_diagonal": 2691794468864.0, + "observations": 20, + "parameter_rms": 2.899687546307522, + "parameter_update_rms": 2.9000687031993513, + "relative_ridge": 0.001, + "target_rms": 0.2908719252061969 + }, + { + "absolute_ridge": 9.83626578424372e+18, + "before_field_mse": 6.381369873326408e+16, + "edge": 13, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": 9.83626578424372e+21, + "minimum_diagonal": 0.0, + "observations": 20, + "parameter_rms": 2.68933867610627, + "parameter_update_rms": 2.6908325115291993, + "relative_ridge": 0.001, + "target_rms": 0.3458967232513598 + }, + { + "absolute_ridge": 2.310925825075764e+23, + "before_field_mse": 6.534132313707218e+20, + "edge": 12, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": 2.310925825075764e+26, + "minimum_diagonal": 4.643060183101878e+24, + "observations": 20, + "parameter_rms": 0.9222431094480169, + "parameter_update_rms": 0.9257461927512325, + "relative_ridge": 0.001, + "target_rms": 0.4134438229204942 + }, + { + "absolute_ridge": 5.315842145923256e+26, + "before_field_mse": 1.8040572061421976e+24, + "edge": 11, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": 5.315842145923256e+29, + "minimum_diagonal": 1.9429183075430184e+25, + "observations": 20, + "parameter_rms": 0.5912122703111267, + "parameter_update_rms": 0.596288099891043, + "relative_ridge": 0.001, + "target_rms": 0.5089343743931583 + }, + { + "absolute_ridge": 8.928898632191432e+29, + "before_field_mse": 2.5973411227762966e+27, + "edge": 10, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": 8.928898632191432e+32, + "minimum_diagonal": 1.7739228624702963e+31, + "observations": 20, + "parameter_rms": 0.6434051617818786, + "parameter_update_rms": 0.6484778518071089, + "relative_ridge": 0.001, + "target_rms": 0.5395306751768495 + }, + { + "absolute_ridge": NaN, + "before_field_mse": 6.838150174412349e+30, + "edge": 9, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": NaN, + "minimum_diagonal": NaN, + "observations": 20, + "parameter_rms": NaN, + "parameter_update_rms": NaN, + "relative_ridge": 0.001, + "target_rms": 0.6844558032335791 + }, + { + "absolute_ridge": NaN, + "before_field_mse": Infinity, + "edge": 8, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": NaN, + "minimum_diagonal": NaN, + "observations": 20, + "parameter_rms": NaN, + "parameter_update_rms": NaN, + "relative_ridge": 0.001, + "target_rms": 0.6733447699197179 + }, + { + "absolute_ridge": NaN, + "before_field_mse": NaN, + "edge": 7, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": NaN, + "minimum_diagonal": NaN, + "observations": 20, + "parameter_rms": NaN, + "parameter_update_rms": NaN, + "relative_ridge": 0.001, + "target_rms": 0.7612376527771475 + }, + { + "absolute_ridge": NaN, + "before_field_mse": Infinity, + "edge": 6, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": NaN, + "minimum_diagonal": NaN, + "observations": 20, + "parameter_rms": NaN, + "parameter_update_rms": NaN, + "relative_ridge": 0.001, + "target_rms": 0.9903483684468577 + }, + { + "absolute_ridge": NaN, + "before_field_mse": NaN, + "edge": 5, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": NaN, + "minimum_diagonal": NaN, + "observations": 20, + "parameter_rms": NaN, + "parameter_update_rms": NaN, + "relative_ridge": 0.001, + "target_rms": 1.2384468689881056 + }, + { + "absolute_ridge": NaN, + "before_field_mse": Infinity, + "edge": 4, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": NaN, + "minimum_diagonal": NaN, + "observations": 20, + "parameter_rms": NaN, + "parameter_update_rms": NaN, + "relative_ridge": 0.001, + "target_rms": 1.376840848200482 + }, + { + "absolute_ridge": NaN, + "before_field_mse": NaN, + "edge": 3, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": NaN, + "minimum_diagonal": NaN, + "observations": 20, + "parameter_rms": NaN, + "parameter_update_rms": NaN, + "relative_ridge": 0.001, + "target_rms": 1.7979800732137456 + }, + { + "absolute_ridge": NaN, + "before_field_mse": Infinity, + "edge": 2, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": NaN, + "minimum_diagonal": NaN, + "observations": 20, + "parameter_rms": NaN, + "parameter_update_rms": NaN, + "relative_ridge": 0.001, + "target_rms": 1.845518153372885 + }, + { + "absolute_ridge": NaN, + "before_field_mse": NaN, + "edge": 1, + "examples": 2560, + "kind": "convolution", + "mean_diagonal": NaN, + "minimum_diagonal": NaN, + "observations": 20, + "parameter_rms": NaN, + "parameter_update_rms": NaN, + "relative_ridge": 0.001, + "target_rms": 2.2282601185794344 + } + ], + "test_examples_touched": 0, + "validation_endpoints_observed": 0, + "wall_seconds": 17.61427068710327, + "work": { + "clean_forward_examples": 48640, + "edge_events": 380, + "feedback_fit_macs_conservative_estimate": 5852657418240, + "forward_macs": 5917207756800, + "logical_batch_loss_queries": 760, + "per_example_causal_observations": 48640, + "per_example_cross_entropy_terms": 97280, + "perturbation_forward_examples": 97280, + "stages": 19, + "total_macs_conservative_estimate": 11769865175040 + } +} diff --git a/results/oral_a_v6_calibration_gate.json b/results/oral_a_v6_calibration_gate.json new file mode 100644 index 0000000..08bd3fa --- /dev/null +++ b/results/oral_a_v6_calibration_gate.json @@ -0,0 +1,178 @@ +{ + "checks": { + "all_layer_at_least_0.20": false, + "early_gain_over_fixed_hfa_at_least_0.08": false, + "early_third_at_least_0.10": false, + "exactly_19_stages": true, + "exactly_380_edge_events": true, + "exactly_48640_per_example_observations": true, + "exactly_760_batch_loss_queries": true, + "feedback_norm_ratios_in_0.1_to_3": false, + "finite": false, + "forward_state_bitwise_fixed": true, + "stage_order_readout_then_18_to_1": true, + "zero_forward_weight_reads_in_fit": true, + "zero_reverse_mode_learning_operations": true + }, + "conditional_short_task_gate_open": false, + "confirmation_test_seeds_touched": false, + "fixed_hfa": { + "all_layer_alignment": 0.011278868932430143, + "early_third_alignment": -0.0002631605093483813, + "feedback_forward_cosine": [ + 0.0008809715509414673, + -0.002320552244782448, + 0.03730485588312149, + 0.02497659996151924, + -0.03438456356525421, + -0.001257734838873148, + 0.023045765236020088, + 0.010684739798307419, + -0.013966654427349567, + -0.03158111870288849, + 0.009583890438079834, + -9.433389641344547e-05, + 0.004490766208618879, + -0.004784357734024525, + 0.004104882013052702, + 0.004543168004602194, + -0.005818712059408426, + 0.0014950999757274985, + 0.003767170011997223 + ], + "feedback_forward_norm_ratio": [ + 0.9996462464332581, + 1.0550107955932617, + 0.9593725800514221, + 1.0017613172531128, + 0.999921441078186, + 0.9631232619285583, + 0.9992203712463379, + 0.9976226687431335, + 1.0018895864486694, + 1.0030434131622314, + 1.0180330276489258, + 1.0107296705245972, + 0.9962885975837708, + 0.9891077876091003, + 1.0009011030197144, + 1.0021207332611084, + 1.0043002367019653, + 1.0021206140518188, + 0.9992877244949341 + ], + "max_feedback_forward_norm_ratio": 1.0550107955932617, + "mean_feedback_forward_cosine": 0.0016142042954207252, + "min_feedback_forward_norm_ratio": 0.9593725800514221, + "per_layer": [ + -0.0005718570901080966, + -0.0006265088450163603, + 0.0012021202128380537, + 3.7642239476554096e-05, + 0.00048483084538020194, + -0.0021051904186606407, + 0.00329246511682868, + 0.00027297029737383127, + 0.007543372455984354, + -0.0053173573687672615, + 0.009830434806644917, + -0.0010977087076753378, + 0.02368866838514805, + 0.007433340419083834, + 0.04359162598848343, + 0.0057456037029623985, + 0.039965368807315826, + 0.0030560996383428574, + 0.07787258923053741 + ] + }, + "learned_scib": { + "all_layer_alignment": NaN, + "early_third_alignment": NaN, + "feedback_forward_cosine": [ + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + -0.006127675995230675, + -0.011246882379055023, + 0.021258916705846786, + 0.00929059274494648, + -0.001609913888387382, + -0.0008538670372217894, + 0.012138523161411285, + 0.002846281975507736, + 0.052370935678482056, + 0.8131900429725647 + ], + "feedback_forward_norm_ratio": [ + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + NaN, + 7.744247913360596, + 7.140369415283203, + 11.147778511047363, + 32.04150390625, + 49.08123016357422, + 48.10276412963867, + 33.58113479614258, + 10.170685768127441, + 12.821575164794922, + 1.120253562927246 + ], + "max_feedback_forward_norm_ratio": NaN, + "mean_feedback_forward_cosine": NaN, + "min_feedback_forward_norm_ratio": NaN, + "per_layer": [ + NaN, + 0.0, + NaN, + 0.0, + NaN, + 0.0, + NaN, + 0.0, + 0.0, + -0.018046237528324127, + 0.009331459179520607, + 0.006288703065365553, + -0.0008923492860049009, + 0.024693015962839127, + 0.04579439014196396, + 0.011086630634963512, + -0.025804810225963593, + 0.3799460828304291, + 0.8721908926963806 + ] + }, + "protocol": "oral_a_v6_stagewise_whitened_causal_capture_gate_v1", + "review_score_after": 5, + "review_score_before": 5, + "score_change_rule": "causal capture alone cannot raise score", + "source_commit": "98bfda74f7ee0457799b86465c867b6943c7be91", + "source_result": "results/oral_a_v6_calibration/result.json", + "status": "failed", + "work": { + "clean_forward_examples": 48640, + "edge_events": 380, + "feedback_fit_macs_conservative_estimate": 5852657418240, + "forward_macs": 5917207756800, + "logical_batch_loss_queries": 760, + "per_example_causal_observations": 48640, + "per_example_cross_entropy_terms": 97280, + "perturbation_forward_examples": 97280, + "stages": 19, + "total_macs_conservative_estimate": 11769865175040 + } +} |