{ "artifact_version": 1, "status": "accepted_initial_negative_architecture_result", "source_commit": "5c758b2aa68ee2b3d73e8a23c5bd187fefd280fd", "comparison_id": "HXO-EXP-0007", "publication": { "source_artifact_sha256": "128b3d4b8d8326fe73fb01082ed9a75af0e5a02b00776659ae6a5ae95ec71a76", "note": "Public edge-hardware training-efficiency export; measurements and provenance are preserved." }, "environment": { "chip": "Apple M3 Max", "memory_bytes": 38654705664, "macos_version": "26.4.1", "python_version": "3.12.9", "mlx_version": "0.32.0" }, "shared_controls": { "precision": "float32", "seeds": [31415, 31416, 31417], "data_seed": 1729, "data_manifest_hash": "1cb42267c085347ed103a77b2d44c2b65e4f7585727972bdbce64f6d3f8d0da3", "dataset_type": "wikitext2_raw_v1", "tokenizer": "wikitext2_word_v1", "vocabulary_size": 16384, "sequence_length": 512, "batch_size": 1, "steps": 4096, "tokens_per_run": 2097152, "train_stream_tokens": 2088628, "validation_stream_tokens": 217646, "final_validation_targets": 217645, "intermediate_eval_interval": 256, "intermediate_eval_batches": 32, "final_validation_scope": "full_stream", "test_split_used": false, "compile": false, "pair_order": [ "seed31415:dense_then_omega", "seed31416:omega_then_dense", "seed31417:dense_then_omega" ] }, "models": { "dense": { "physical_parameter_count": 9881472, "estimated_active_parameters": 9685248, "dense_operator_dimensionality": 9682944, "estimated_flops_per_token": 20938752 }, "omega_mlp": { "physical_parameter_count": 7670616, "estimated_active_parameters": 7474392, "dense_operator_dimensionality": 8355840, "estimated_flops_per_token": 16609536, "implemented_variant": "one-stage pure TensorAxis MLP with no permutation, gates, routing, low-rank residual, spectral operator, or custom Metal kernel" } }, "runs": [ { "order": 1, "model": "dense", "seed": 31415, "run_id": "20260824T184056.146724-0500_dense_1eae8908", "config_hash": "1eae890844eacd9c17109821875e6ebc066a4f25e58221f4e88adfdca85f3367", "final_validation_loss": 5.563161737810569, "final_validation_perplexity": 260.64562804903215, "total_wall_time_seconds": 65.3388961669989, "ordinary_compute_median_tokens_per_sec": 35405.217372234314, "ordinary_step_median_tokens_per_sec": 34785.61109728934, "peak_memory_bytes": 689973809, "first_fixed_validation_loss_at_most_5_8": {"step": 1536, "elapsed_seconds": 23.829409416997805}, "first_fixed_validation_loss_at_most_5_7": {"step": 2560, "elapsed_seconds": 40.00615529200877}, "console_log_sha256": "721a827d21ee34102ba8f3a1c044798adb8526b6f7eca7ec2d048ebeaa5d09d6", "metrics_jsonl_sha256": "522b45053496c2392ceb15f5e507e7d2d0a676de6aaa0ec7801952bbf47d83dc", "manifest_sha256": "6e53a72c53552e63b2b993a426d6a6e67ff7311eecaffb6eeea514b9197ec2c4", "final_model_sha256": "7ec2a455e74e8586478ad52b37c0fc5f6c0bcf955d668518b7cb59ccc297f258" }, { "order": 2, "model": "omega_mlp", "seed": 31415, "run_id": "20260824T184210.261632-0500_omega_mlp_13280633", "config_hash": "132806336b87f8b413a4fa2f08eb5cebef492a9cea0da93afcc4732968c85978", "final_validation_loss": 5.717567560613551, "final_validation_perplexity": 304.1641615070916, "total_wall_time_seconds": 64.22028791700723, "ordinary_compute_median_tokens_per_sec": 36117.5942362817, "ordinary_step_median_tokens_per_sec": 35476.769995276234, "peak_memory_bytes": 467143740, "first_fixed_validation_loss_at_most_5_8": {"step": 2560, "elapsed_seconds": 39.56809879199136}, "first_fixed_validation_loss_at_most_5_7": null, "console_log_sha256": "90c08ff241bf34f96456fe1f8eecd7a59b25bc8f56b5fb3b07f5819cbbb68cb4", "metrics_jsonl_sha256": "3b1a9792b000cd5866d56a4cbc66e1d3a947b839f56ee9750854b2aa19f923f8", "manifest_sha256": "93ba83c72a09b31c4abb1de67961cbf1c09a4994e6e8e9d9f804ffe027fc0854", "final_model_sha256": "5567ee058389b1a25b07dd55a876309d198c5ad843f2089067fc10112d49a641" }, { "order": 3, "model": "omega_mlp", "seed": 31416, "run_id": "20260824T184323.487129-0500_omega_mlp_b6b1d7fa", "config_hash": "b6b1d7fa67c9c0f8a381c545c0bd4feda49e3c754b908d3d9cf6a605dff7c623", "final_validation_loss": 5.719609829556203, "final_validation_perplexity": 304.7859812730762, "total_wall_time_seconds": 64.44066620900412, "ordinary_compute_median_tokens_per_sec": 36053.32301974014, "ordinary_step_median_tokens_per_sec": 35417.15684190551, "peak_memory_bytes": 467143740, "first_fixed_validation_loss_at_most_5_8": {"step": 2560, "elapsed_seconds": 39.752645917003974}, "first_fixed_validation_loss_at_most_5_7": null, "console_log_sha256": "4f98ad70cf39d55fb427b5b667446a2674e9caf000890dffa475d5236a8e7109", "metrics_jsonl_sha256": "9b8167c8d8af35cf8df35acb0d8882dd547af5c3bbe0effee5f6a3af50345e82", "manifest_sha256": "dd9fcdf2e80bbbb159164cb03a7d47fa11228e2986635f89161ca5d145792551", "final_model_sha256": "ea388d1936614ff47afb22888d9c05a53d0dbce2f07fc526af5c3b84d50eecf9" }, { "order": 4, "model": "dense", "seed": 31416, "run_id": "20260824T184436.055693-0500_dense_9b3a8d84", "config_hash": "9b3a8d84b1498f3502ecf29b9aca47a87f622685ccaa84e4a41850d40ad52563", "final_validation_loss": 5.561861547713589, "final_validation_perplexity": 260.3069593991437, "total_wall_time_seconds": 74.10152591700898, "ordinary_compute_median_tokens_per_sec": 32486.493670416738, "ordinary_step_median_tokens_per_sec": 31954.23155736798, "peak_memory_bytes": 689975857, "first_fixed_validation_loss_at_most_5_8": {"step": 1536, "elapsed_seconds": 24.18441575000179}, "first_fixed_validation_loss_at_most_5_7": {"step": 2560, "elapsed_seconds": 41.848833750002086}, "console_log_sha256": "914580b382a564b236c6bca8706c0e10b4766713e748d36dab5912c571e4a352", "metrics_jsonl_sha256": "7805c08b6674e692def9f91795baf6c0d48294389c96233d61fc1f34e3b858d1", "manifest_sha256": "0be7c120066853d9ca2ee55dbf43838e1704e409469f3c21cd5465a849b52d37", "final_model_sha256": "0ee0c46689c2fa72a5b1b8d244a23a79611c8dc6f20679a2e9932f64550daf55" }, { "order": 5, "model": "dense", "seed": 31417, "run_id": "20260824T184600.674918-0500_dense_e315ca03", "config_hash": "e315ca0352b2bff91eba476ae2d451f97fd16b6197f752e228f4d674496a56d9", "final_validation_loss": 5.568321400669803, "final_validation_perplexity": 261.99394706000214, "total_wall_time_seconds": 65.9494499170105, "ordinary_compute_median_tokens_per_sec": 35233.59998174308, "ordinary_step_median_tokens_per_sec": 34619.64355038829, "peak_memory_bytes": 686827569, "first_fixed_validation_loss_at_most_5_8": {"step": 1536, "elapsed_seconds": 24.38770150000346}, "first_fixed_validation_loss_at_most_5_7": {"step": 2560, "elapsed_seconds": 40.71796712500509}, "console_log_sha256": "f057bea50edcbf4528185c6e4c64cc11d9257cabc2004698857218de597cd2b4", "metrics_jsonl_sha256": "e97f8b6c4f2ff46a6a703834b50ce5f66d98e9055aebe5a905c1215f527654ab", "manifest_sha256": "1aed57e14cc41ce3bcbd9525e73764d54e62a8f0affc96930786fdf1de090494", "final_model_sha256": "6ab72d739ffb41ea0db2e0f90e26d19b70e1687e49ceef157db0c6adae95d04f" }, { "order": 6, "model": "omega_mlp", "seed": 31417, "run_id": "20260824T184715.534109-0500_omega_mlp_57185edb", "config_hash": "57185edb10fc8cac7ab098fcdff6b6059082510bdcbdb09fa5814e2ef35e8999", "final_validation_loss": 5.727047002255546, "final_validation_perplexity": 307.06117727790974, "total_wall_time_seconds": 64.31220912499703, "ordinary_compute_median_tokens_per_sec": 36038.942891377825, "ordinary_step_median_tokens_per_sec": 35404.85875254069, "peak_memory_bytes": 467143740, "first_fixed_validation_loss_at_most_5_8": {"step": 2560, "elapsed_seconds": 39.62436441599857}, "first_fixed_validation_loss_at_most_5_7": null, "console_log_sha256": "6a8809c6227da820df12135d81522554e353557d0c56ba73a187a20a7086dc4f", "metrics_jsonl_sha256": "d572c051ef529313ac68227ead7a924667021182be1623a2287a4babcc3643c4", "manifest_sha256": "34300fd3bcce190b9093afcee289c3d572db0ec99edabf1a34547e4b4ced6650", "final_model_sha256": "7e592c70f7ad7853c6a8974626d23f6798e4653a6e28f4bcc5b2e002f6e31cf2" } ], "aggregate": { "dense": { "final_validation_loss_mean": 5.56444822873132, "final_validation_perplexity_mean": 260.9821781693926, "total_wall_time_seconds_mean": 68.46329066700612, "ordinary_compute_median_tokens_per_sec_mean": 34375.10367479804, "ordinary_step_median_tokens_per_sec_mean": 33786.49540168187, "peak_memory_bytes_mean": 688925745, "time_to_fixed_validation_loss_5_8_seconds_mean": 24.13384222233435, "time_to_fixed_validation_loss_5_7_seconds_mean": 40.857652055671984 }, "omega_mlp": { "final_validation_loss_mean": 5.721408130808434, "final_validation_perplexity_mean": 305.33710668602583, "total_wall_time_seconds_mean": 64.32438775033613, "ordinary_compute_median_tokens_per_sec_mean": 36069.953382466556, "ordinary_step_median_tokens_per_sec_mean": 35432.928529907476, "peak_memory_bytes_mean": 467143740, "time_to_fixed_validation_loss_5_8_seconds_mean": 39.6483697083313, "time_to_fixed_validation_loss_5_7_seconds_mean": null }, "omega_relative_to_dense": { "physical_parameter_reduction_fraction": 0.22373751603000036, "estimated_active_parameter_reduction_fraction": 0.22827045833002935, "dense_operator_dimensionality_reduction_fraction": 0.13705583756345174, "estimated_flops_reduction_fraction": 0.206756161971831, "mean_peak_memory_reduction_fraction": 0.32192439694643726, "ordinary_compute_median_throughput_gain_fraction": 0.04930457006624511, "ordinary_step_median_throughput_gain_fraction": 0.0487305092952508, "mean_end_to_end_elapsed_reduction_fraction": 0.06045433803059386, "mean_final_validation_loss_increase": 0.15695990207711397, "time_to_fixed_validation_loss_5_8_increase_fraction": 0.6428536054503262 } }, "verification": { "metrics_rows_per_run": 4096, "manifests_completed": true, "console_and_structured_step_logging_present": true, "final_checkpoints_reloaded": true, "reloaded_full_validation_matches_recorded_exactly": true, "final_validation_scope": "full_stream", "final_validation_targets_per_run": 217645, "host_metal_test_suite_before_launch": "58 passed in 0.33s", "independent_prelaunch_qa_verdict": "pass", "independent_benchmark_verdict": "pass_for_initial_negative_architecture_result" }, "conclusion": { "confirmed_result": "At this exact 10M-class, one-stage, equal-shape FP32 pilot, the pure TensorAxis MLP reduced stored and estimated execution resources but produced worse held-out validation loss in every seed and a worse validation-loss-versus-wall-clock frontier than the dense MLP control.", "decision": "Do not scale this current primitive or authorize a custom Metal kernel. Preserve the negative result and run a separately registered native-MLX capacity-rescue ablation before any optimization work.", "claim_ceiling": "This result applies only to the implemented one-stage TensorAxis MLP and tokenizer-specific WikiText-2 pilot. It does not falsify later HELIX-Omega components or establish general capability equivalence." }, "limitations": [ "This is a small, approximately one-token-pass pilot with three seeds, not a scaling study.", "The raw whitespace tokenizer and vocabulary are project-specific; the reported perplexity is not directly comparable to standard WikiText-2 tokenizations or prior byte-tokenizer runs.", "Architecture-dependent RNG consumption means same-seed shared-layer tensors were not forced to identical initialization.", "One dense run slowed materially late in the sequence, so the mean end-to-end elapsed reduction is order and thermal sensitive.", "The final full-stream validation point and intermediate fixed-window points have different scopes; time-to-threshold comparisons use only the fixed window.", "This pilot does not establish a broad edge-training efficiency claim." ] }