{ "created_utc": "2026-10-07T06:58:37.442670+00:00", "purpose": "Independent numerical verification of theory Eq. 24-26; no science run changes", "device": "cpu", "dtype": "torch.float64", "python": "3.10.16", "torch": "2.6.0+cu126", "transformers": "4.40.0", "spec": { "initialization": 20261007, "base_unique_blocks": 1, "repeats": 2, "initialization_reference_layers": 8, "model": { "vocab_size": 11, "positions": 16, "hidden_size": 8, "attention_heads": 2, "dropout": 0.0 } }, "scalar_selections": { "query": { "tokens": [ [ 1, 4, 2 ] ], "position": 2, "token_logit": 7 }, "train": { "tokens": [ [ 3, 1, 5 ] ], "position": 2, "token_logit": 4 } }, "absolute_error_tolerance": 1e-10, "initial_full_logits_max_abs_errors": { "shared": 0.0, "untied": 0.0, "mlp_shared": 0.0 }, "checks": { "shared": { "arm": "shared", "scalar_outputs": { "query": -0.05500031918167053, "train": 0.02101119074724663 }, "unique_parameters": 1104, "parameter_aliases": { "transformer.wte.weight": [ "transformer.wte.weight" ], "transformer.wpe.weight": [ "transformer.wpe.weight" ], "transformer.h.0.ln_1.weight": [ "transformer.h.0.ln_1.weight", "transformer.h.1.ln_1.weight" ], "transformer.h.0.ln_1.bias": [ "transformer.h.0.ln_1.bias", "transformer.h.1.ln_1.bias" ], "transformer.h.0.attn.c_attn.weight": [ "transformer.h.0.attn.c_attn.weight", "transformer.h.1.attn.c_attn.weight" ], "transformer.h.0.attn.c_attn.bias": [ "transformer.h.0.attn.c_attn.bias", "transformer.h.1.attn.c_attn.bias" ], "transformer.h.0.attn.c_proj.weight": [ "transformer.h.0.attn.c_proj.weight", "transformer.h.1.attn.c_proj.weight" ], "transformer.h.0.attn.c_proj.bias": [ "transformer.h.0.attn.c_proj.bias", "transformer.h.1.attn.c_proj.bias" ], "transformer.h.0.ln_2.weight": [ "transformer.h.0.ln_2.weight", "transformer.h.1.ln_2.weight" ], "transformer.h.0.ln_2.bias": [ "transformer.h.0.ln_2.bias", "transformer.h.1.ln_2.bias" ], "transformer.h.0.mlp.c_fc.weight": [ "transformer.h.0.mlp.c_fc.weight", "transformer.h.1.mlp.c_fc.weight" ], "transformer.h.0.mlp.c_fc.bias": [ "transformer.h.0.mlp.c_fc.bias", "transformer.h.1.mlp.c_fc.bias" ], "transformer.h.0.mlp.c_proj.weight": [ "transformer.h.0.mlp.c_proj.weight", "transformer.h.1.mlp.c_proj.weight" ], "transformer.h.0.mlp.c_proj.bias": [ "transformer.h.0.mlp.c_proj.bias", "transformer.h.1.mlp.c_proj.bias" ], "transformer.ln_f.weight": [ "transformer.ln_f.weight" ], "transformer.ln_f.bias": [ "transformer.ln_f.bias" ] }, "kernel": -9.253421669347837, "untied_kernel": -5.12471454354214, "common_parameter_kernel": -0.9955788057977971, "common_tied_embedding_kernel": 0.04231565484872117, "block_diagonal_kernel": -4.129135737744343, "block_cross_call_kernel": -4.128707125805699, "per_parameter_jacobian_errors": { "query": { "transformer.wte.weight": 0.0, "transformer.wpe.weight": 0.0, "transformer.h.0.ln_1.weight": 0.0, "transformer.h.0.ln_1.bias": 0.0, "transformer.h.0.attn.c_attn.weight": 0.0, "transformer.h.0.attn.c_attn.bias": 0.0, "transformer.h.0.attn.c_proj.weight": 0.0, "transformer.h.0.attn.c_proj.bias": 0.0, "transformer.h.0.ln_2.weight": 0.0, "transformer.h.0.ln_2.bias": 0.0, "transformer.h.0.mlp.c_fc.weight": 0.0, "transformer.h.0.mlp.c_fc.bias": 0.0, "transformer.h.0.mlp.c_proj.weight": 0.0, "transformer.h.0.mlp.c_proj.bias": 0.0, "transformer.ln_f.weight": 0.0, "transformer.ln_f.bias": 0.0 }, "train": { "transformer.wte.weight": 0.0, "transformer.wpe.weight": 0.0, "transformer.h.0.ln_1.weight": 0.0, "transformer.h.0.ln_1.bias": 0.0, "transformer.h.0.attn.c_attn.weight": 0.0, "transformer.h.0.attn.c_attn.bias": 0.0, "transformer.h.0.attn.c_proj.weight": 0.0, "transformer.h.0.attn.c_proj.bias": 0.0, "transformer.h.0.ln_2.weight": 0.0, "transformer.h.0.ln_2.bias": 0.0, "transformer.h.0.mlp.c_fc.weight": 0.0, "transformer.h.0.mlp.c_fc.bias": 0.0, "transformer.h.0.mlp.c_proj.weight": 0.0, "transformer.h.0.mlp.c_proj.bias": 0.0, "transformer.ln_f.weight": 0.0, "transformer.ln_f.bias": 0.0 } }, "errors": { "scalar_jacobian_vs_alias_sum_max_abs": 0.0, "kernel_vs_common_plus_diagonal_plus_cross_abs": 8.881784197001252e-16, "untied_kernel_vs_common_plus_diagonal_abs": 0.0, "kernel_difference_vs_cross_terms_abs": 1.7763568394002505e-15 }, "passed": true, "plain_sgd_taylor_check": { "loss": "0.5 * (selected train token logit - 0.5)**2", "optimizer": "plain SGD, no momentum, clipping, decay or optimizer state", "kernel_directional_derivative_error": 8.881784197001252e-16, "steps": [ { "step_size": 0.0001, "actual_query_logit_change": -0.0004394824778334444, "kernel_first_order_change": -0.0004432285426914547, "absolute_taylor_remainder": 3.746064858010294e-06, "absolute_remainder_divided_by_step_squared": 374.60648580102935 }, { "step_size": 5e-05, "actual_query_logit_change": -0.00022068102755029573, "kernel_first_order_change": -0.00022161427134572735, "absolute_taylor_remainder": 9.33243795431623e-07, "absolute_remainder_divided_by_step_squared": 373.2975181726492 }, { "step_size": 2.5e-05, "actual_query_logit_change": -0.00011057423824875623, "kernel_first_order_change": -0.00011080713567286368, "absolute_taylor_remainder": 2.3289742410744602e-07, "absolute_remainder_divided_by_step_squared": 372.6358785719136 } ], "remainder_ratio_when_step_halved": [ 4.0140259987238895, 4.007102264046836 ], "accepted_ratio_interval": [ 3.0, 5.0 ], "passed": true } }, "untied": { "arm": "untied", "scalar_outputs": { "query": -0.05500031918167053, "train": 0.02101119074724663 }, "unique_parameters": 1976, "parameter_aliases": { "transformer.wte.weight": [ "transformer.wte.weight" ], "transformer.wpe.weight": [ "transformer.wpe.weight" ], "transformer.h.0.ln_1.weight": [ "transformer.h.0.ln_1.weight" ], "transformer.h.0.ln_1.bias": [ "transformer.h.0.ln_1.bias" ], "transformer.h.0.attn.c_attn.weight": [ "transformer.h.0.attn.c_attn.weight" ], "transformer.h.0.attn.c_attn.bias": [ "transformer.h.0.attn.c_attn.bias" ], "transformer.h.0.attn.c_proj.weight": [ "transformer.h.0.attn.c_proj.weight" ], "transformer.h.0.attn.c_proj.bias": [ "transformer.h.0.attn.c_proj.bias" ], "transformer.h.0.ln_2.weight": [ "transformer.h.0.ln_2.weight" ], "transformer.h.0.ln_2.bias": [ "transformer.h.0.ln_2.bias" ], "transformer.h.0.mlp.c_fc.weight": [ "transformer.h.0.mlp.c_fc.weight" ], "transformer.h.0.mlp.c_fc.bias": [ "transformer.h.0.mlp.c_fc.bias" ], "transformer.h.0.mlp.c_proj.weight": [ "transformer.h.0.mlp.c_proj.weight" ], "transformer.h.0.mlp.c_proj.bias": [ "transformer.h.0.mlp.c_proj.bias" ], "transformer.h.1.ln_1.weight": [ "transformer.h.1.ln_1.weight" ], "transformer.h.1.ln_1.bias": [ "transformer.h.1.ln_1.bias" ], "transformer.h.1.attn.c_attn.weight": [ "transformer.h.1.attn.c_attn.weight" ], "transformer.h.1.attn.c_attn.bias": [ "transformer.h.1.attn.c_attn.bias" ], "transformer.h.1.attn.c_proj.weight": [ "transformer.h.1.attn.c_proj.weight" ], "transformer.h.1.attn.c_proj.bias": [ "transformer.h.1.attn.c_proj.bias" ], "transformer.h.1.ln_2.weight": [ "transformer.h.1.ln_2.weight" ], "transformer.h.1.ln_2.bias": [ "transformer.h.1.ln_2.bias" ], "transformer.h.1.mlp.c_fc.weight": [ "transformer.h.1.mlp.c_fc.weight" ], "transformer.h.1.mlp.c_fc.bias": [ "transformer.h.1.mlp.c_fc.bias" ], "transformer.h.1.mlp.c_proj.weight": [ "transformer.h.1.mlp.c_proj.weight" ], "transformer.h.1.mlp.c_proj.bias": [ "transformer.h.1.mlp.c_proj.bias" ], "transformer.ln_f.weight": [ "transformer.ln_f.weight" ], "transformer.ln_f.bias": [ "transformer.ln_f.bias" ] }, "kernel": -5.12471454354214, "untied_kernel": -5.12471454354214, "common_parameter_kernel": -0.9955788057977971, "common_tied_embedding_kernel": 0.04231565484872117, "block_diagonal_kernel": -4.129135737744343, "block_cross_call_kernel": 0.0, "per_parameter_jacobian_errors": { "query": { "transformer.wte.weight": 0.0, "transformer.wpe.weight": 0.0, "transformer.h.0.ln_1.weight": 0.0, "transformer.h.0.ln_1.bias": 0.0, "transformer.h.0.attn.c_attn.weight": 0.0, "transformer.h.0.attn.c_attn.bias": 0.0, "transformer.h.0.attn.c_proj.weight": 0.0, "transformer.h.0.attn.c_proj.bias": 0.0, "transformer.h.0.ln_2.weight": 0.0, "transformer.h.0.ln_2.bias": 0.0, "transformer.h.0.mlp.c_fc.weight": 0.0, "transformer.h.0.mlp.c_fc.bias": 0.0, "transformer.h.0.mlp.c_proj.weight": 0.0, "transformer.h.0.mlp.c_proj.bias": 0.0, "transformer.h.1.ln_1.weight": 0.0, "transformer.h.1.ln_1.bias": 0.0, "transformer.h.1.attn.c_attn.weight": 0.0, "transformer.h.1.attn.c_attn.bias": 0.0, "transformer.h.1.attn.c_proj.weight": 0.0, "transformer.h.1.attn.c_proj.bias": 0.0, "transformer.h.1.ln_2.weight": 0.0, "transformer.h.1.ln_2.bias": 0.0, "transformer.h.1.mlp.c_fc.weight": 0.0, "transformer.h.1.mlp.c_fc.bias": 0.0, "transformer.h.1.mlp.c_proj.weight": 0.0, "transformer.h.1.mlp.c_proj.bias": 0.0, "transformer.ln_f.weight": 0.0, "transformer.ln_f.bias": 0.0 }, "train": { "transformer.wte.weight": 0.0, "transformer.wpe.weight": 0.0, "transformer.h.0.ln_1.weight": 0.0, "transformer.h.0.ln_1.bias": 0.0, "transformer.h.0.attn.c_attn.weight": 0.0, "transformer.h.0.attn.c_attn.bias": 0.0, "transformer.h.0.attn.c_proj.weight": 0.0, "transformer.h.0.attn.c_proj.bias": 0.0, "transformer.h.0.ln_2.weight": 0.0, "transformer.h.0.ln_2.bias": 0.0, "transformer.h.0.mlp.c_fc.weight": 0.0, "transformer.h.0.mlp.c_fc.bias": 0.0, "transformer.h.0.mlp.c_proj.weight": 0.0, "transformer.h.0.mlp.c_proj.bias": 0.0, "transformer.h.1.ln_1.weight": 0.0, "transformer.h.1.ln_1.bias": 0.0, "transformer.h.1.attn.c_attn.weight": 0.0, "transformer.h.1.attn.c_attn.bias": 0.0, "transformer.h.1.attn.c_proj.weight": 0.0, "transformer.h.1.attn.c_proj.bias": 0.0, "transformer.h.1.ln_2.weight": 0.0, "transformer.h.1.ln_2.bias": 0.0, "transformer.h.1.mlp.c_fc.weight": 0.0, "transformer.h.1.mlp.c_fc.bias": 0.0, "transformer.h.1.mlp.c_proj.weight": 0.0, "transformer.h.1.mlp.c_proj.bias": 0.0, "transformer.ln_f.weight": 0.0, "transformer.ln_f.bias": 0.0 } }, "errors": { "scalar_jacobian_vs_alias_sum_max_abs": 0.0, "kernel_vs_common_plus_diagonal_plus_cross_abs": 0.0, "untied_kernel_vs_common_plus_diagonal_abs": 0.0, "kernel_difference_vs_cross_terms_abs": 0.0 }, "passed": true, "plain_sgd_taylor_check": { "loss": "0.5 * (selected train token logit - 0.5)**2", "optimizer": "plain SGD, no momentum, clipping, decay or optimizer state", "kernel_directional_derivative_error": 0.0, "steps": [ { "step_size": 0.0001, "actual_query_logit_change": -0.0002443138396955938, "kernel_first_order_change": -0.0002454680916971517, "absolute_taylor_remainder": 1.1542520015579228e-06, "absolute_remainder_divided_by_step_squared": 115.42520015579227 }, { "step_size": 5e-05, "actual_query_logit_change": -0.0001224460370929381, "kernel_first_order_change": -0.00012273404584857585, "absolute_taylor_remainder": 2.880087556377431e-07, "absolute_remainder_divided_by_step_squared": 115.20350225509723 }, { "step_size": 2.5e-05, "actual_query_logit_change": -6.129509043929782e-05, "kernel_first_order_change": -6.136702292428793e-05, "absolute_taylor_remainder": 7.19324849901093e-08, "absolute_remainder_divided_by_step_squared": 115.09197598417488 } ], "remainder_ratio_when_step_halved": [ 4.0076976097550965, 4.003876074590559 ], "accepted_ratio_interval": [ 3.0, 5.0 ], "passed": true } }, "mlp_shared": { "arm": "mlp_shared", "scalar_outputs": { "query": -0.05500031918167053, "train": 0.02101119074724663 }, "unique_parameters": 1424, "parameter_aliases": { "transformer.wte.weight": [ "transformer.wte.weight" ], "transformer.wpe.weight": [ "transformer.wpe.weight" ], "transformer.h.0.ln_1.weight": [ "transformer.h.0.ln_1.weight" ], "transformer.h.0.ln_1.bias": [ "transformer.h.0.ln_1.bias" ], "transformer.h.0.attn.c_attn.weight": [ "transformer.h.0.attn.c_attn.weight" ], "transformer.h.0.attn.c_attn.bias": [ "transformer.h.0.attn.c_attn.bias" ], "transformer.h.0.attn.c_proj.weight": [ "transformer.h.0.attn.c_proj.weight" ], "transformer.h.0.attn.c_proj.bias": [ "transformer.h.0.attn.c_proj.bias" ], "transformer.h.0.ln_2.weight": [ "transformer.h.0.ln_2.weight" ], "transformer.h.0.ln_2.bias": [ "transformer.h.0.ln_2.bias" ], "transformer.h.0.mlp.c_fc.weight": [ "transformer.h.0.mlp.c_fc.weight", "transformer.h.1.mlp.c_fc.weight" ], "transformer.h.0.mlp.c_fc.bias": [ "transformer.h.0.mlp.c_fc.bias", "transformer.h.1.mlp.c_fc.bias" ], "transformer.h.0.mlp.c_proj.weight": [ "transformer.h.0.mlp.c_proj.weight", "transformer.h.1.mlp.c_proj.weight" ], "transformer.h.0.mlp.c_proj.bias": [ "transformer.h.0.mlp.c_proj.bias", "transformer.h.1.mlp.c_proj.bias" ], "transformer.h.1.ln_1.weight": [ "transformer.h.1.ln_1.weight" ], "transformer.h.1.ln_1.bias": [ "transformer.h.1.ln_1.bias" ], "transformer.h.1.attn.c_attn.weight": [ "transformer.h.1.attn.c_attn.weight" ], "transformer.h.1.attn.c_attn.bias": [ "transformer.h.1.attn.c_attn.bias" ], "transformer.h.1.attn.c_proj.weight": [ "transformer.h.1.attn.c_proj.weight" ], "transformer.h.1.attn.c_proj.bias": [ "transformer.h.1.attn.c_proj.bias" ], "transformer.h.1.ln_2.weight": [ "transformer.h.1.ln_2.weight" ], "transformer.h.1.ln_2.bias": [ "transformer.h.1.ln_2.bias" ], "transformer.ln_f.weight": [ "transformer.ln_f.weight" ], "transformer.ln_f.bias": [ "transformer.ln_f.bias" ] }, "kernel": -7.18517398434009, "untied_kernel": -5.12471454354214, "common_parameter_kernel": -0.9955788057977971, "common_tied_embedding_kernel": 0.04231565484872117, "block_diagonal_kernel": -4.129135737744343, "block_cross_call_kernel": -2.060459440797952, "per_parameter_jacobian_errors": { "query": { "transformer.wte.weight": 0.0, "transformer.wpe.weight": 0.0, "transformer.h.0.ln_1.weight": 0.0, "transformer.h.0.ln_1.bias": 0.0, "transformer.h.0.attn.c_attn.weight": 0.0, "transformer.h.0.attn.c_attn.bias": 0.0, "transformer.h.0.attn.c_proj.weight": 0.0, "transformer.h.0.attn.c_proj.bias": 0.0, "transformer.h.0.ln_2.weight": 0.0, "transformer.h.0.ln_2.bias": 0.0, "transformer.h.0.mlp.c_fc.weight": 0.0, "transformer.h.0.mlp.c_fc.bias": 0.0, "transformer.h.0.mlp.c_proj.weight": 0.0, "transformer.h.0.mlp.c_proj.bias": 0.0, "transformer.h.1.ln_1.weight": 0.0, "transformer.h.1.ln_1.bias": 0.0, "transformer.h.1.attn.c_attn.weight": 0.0, "transformer.h.1.attn.c_attn.bias": 0.0, "transformer.h.1.attn.c_proj.weight": 0.0, "transformer.h.1.attn.c_proj.bias": 0.0, "transformer.h.1.ln_2.weight": 0.0, "transformer.h.1.ln_2.bias": 0.0, "transformer.ln_f.weight": 0.0, "transformer.ln_f.bias": 0.0 }, "train": { "transformer.wte.weight": 0.0, "transformer.wpe.weight": 0.0, "transformer.h.0.ln_1.weight": 0.0, "transformer.h.0.ln_1.bias": 0.0, "transformer.h.0.attn.c_attn.weight": 0.0, "transformer.h.0.attn.c_attn.bias": 0.0, "transformer.h.0.attn.c_proj.weight": 0.0, "transformer.h.0.attn.c_proj.bias": 0.0, "transformer.h.0.ln_2.weight": 0.0, "transformer.h.0.ln_2.bias": 0.0, "transformer.h.0.mlp.c_fc.weight": 0.0, "transformer.h.0.mlp.c_fc.bias": 0.0, "transformer.h.0.mlp.c_proj.weight": 0.0, "transformer.h.0.mlp.c_proj.bias": 0.0, "transformer.h.1.ln_1.weight": 0.0, "transformer.h.1.ln_1.bias": 0.0, "transformer.h.1.attn.c_attn.weight": 0.0, "transformer.h.1.attn.c_attn.bias": 0.0, "transformer.h.1.attn.c_proj.weight": 0.0, "transformer.h.1.attn.c_proj.bias": 0.0, "transformer.h.1.ln_2.weight": 0.0, "transformer.h.1.ln_2.bias": 0.0, "transformer.ln_f.weight": 0.0, "transformer.ln_f.bias": 0.0 } }, "errors": { "scalar_jacobian_vs_alias_sum_max_abs": 0.0, "kernel_vs_common_plus_diagonal_plus_cross_abs": 2.220446049250313e-15, "untied_kernel_vs_common_plus_diagonal_abs": 0.0, "kernel_difference_vs_cross_terms_abs": 2.220446049250313e-15 }, "passed": true, "plain_sgd_taylor_check": { "loss": "0.5 * (selected train token logit - 0.5)**2", "optimizer": "plain SGD, no momentum, clipping, decay or optimizer state", "kernel_directional_derivative_error": 8.881784197001252e-16, "steps": [ { "step_size": 0.0001, "actual_query_logit_change": -0.00034189519899297666, "kernel_first_order_change": -0.00034416179310329224, "absolute_taylor_remainder": 2.2665941103155855e-06, "absolute_remainder_divided_by_step_squared": 226.65941103155853 }, { "step_size": 5e-05, "actual_query_logit_change": -0.00017151580147124756, "kernel_first_order_change": -0.00017208089655164612, "absolute_taylor_remainder": 5.650950803985568e-07, "absolute_remainder_divided_by_step_squared": 226.0380321594227 }, { "step_size": 2.5e-05, "actual_query_logit_change": -8.589937032372347e-05, "kernel_first_order_change": -8.604044827582306e-05, "absolute_taylor_remainder": 1.4107795209959315e-07, "absolute_remainder_divided_by_step_squared": 225.72472335934904 } ], "remainder_ratio_when_step_halved": [ 4.010996005693371, 4.005552051107399 ], "accepted_ratio_interval": [ 3.0, 5.0 ], "passed": true } } }, "source_sha256": { "docs/development-artifacts/loop-learning-development-v1/verify_loop_kernel.py": "60219e6639f69b2c85be9470f00bbc2234408694f2569e2737f1f662d0bea16d", "src/llm_memory_editability/loop_learning.py": "502153b1aaac4f988103aacc8f3106dc57b0a4d6495115628335c95fc46c64be", "src/llm_memory_editability/grokking_reproduction.py": "7e3a89d96f6fef7bd4665ac46dd1acb6e7cb461c6878647ecc7c6f425ac31117" }, "passed": true, "limits": "Tiny randomly initialized complete ReproductionGPT with causal attention, GELU MLP, residual paths, LayerNorm and tied input/output embedding. Kernel identities hold at the shared common-function point. Finite-difference steps use a fixed scalar squared loss and disposable CPU copies. No claim about AdamW dynamics, knowledge quality, generalization, sample efficiency, grokking, scientific training completion or an active GPU run is tested." }