{
  "artifact_version": 1,
  "status": "accepted_negative_rescue_result",
  "source_commit": "0582756d01c7381713b3c290cc01a370002bd0f8",
  "comparison_id": "HXO-EXP-0008",
  "publication": {
    "source_artifact_sha256": "c40661d2ae7fb3a4c5007a17a8c334b4fb76b6d94d077b058f26456a2064f569",
    "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": [31418, 31419],
    "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": 2048,
    "tokens_per_run": 1048576,
    "final_validation_scope": "full_stream",
    "final_validation_targets": 217645,
    "intermediate_eval_interval": 256,
    "intermediate_eval_batches": 32,
    "ordinary_step_filter": "step > 128 and step % 256 != 0",
    "pair_order": [
      "seed31418:dense_then_omega_then_diagonal",
      "seed31419:diagonal_then_omega_then_dense"
    ],
    "initialization_manifests": {
      "31418": "219be9239df740388015d21d491c6350b66c89c62283e1ecd13ed03fee11cbdf",
      "31419": "b2decb4508446a7a0bad6eb3f750ee3f81e94476e409a645dfb4f9b931756fd9"
    }
  },
  "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
    },
    "omega_mlp_diagonal": {
      "physical_parameter_count": 7672920,
      "estimated_active_parameters": 7476696,
      "dense_operator_dimensionality": 8355840,
      "estimated_flops_per_token": 16614144,
      "implemented_variant": "one-stage TensorAxis MLP with an identity-initialized FP32 post-diagonal on each gate, value, and output projection"
    }
  },
  "runs": [
    {
      "order": 1,
      "seed": 31418,
      "model": "dense",
      "run_id": "20260825T004828.805805-0500_dense_8732d8c9",
      "config_hash": "8732d8c94bfc4447b402604ff87178cbb2111eb010293af1f80c88cec7cfe593",
      "final_validation_loss": 5.728849528766921,
      "final_validation_perplexity": 307.61516232694004,
      "total_wall_time_seconds": 34.83292062499095,
      "ordinary_compute_median_tokens_per_sec": 35007.25087262146,
      "ordinary_step_median_tokens_per_sec": 34252.26519712062,
      "peak_memory_bytes": 685944873,
      "console_log_sha256": "19d42e7c95263b6b3d103e823fc61de48444b876a58a335d30d265eec360e6a7",
      "metrics_jsonl_sha256": "9385473ed727b7f36300435cdddf102863e7fa53f31265e0404f6f6d24be4b81",
      "manifest_sha256": "2e855c92cfab9c0f9acb356774d27c686a6fe924589667aa69892238ee6f3248",
      "final_model_sha256": "06eb9b8215e35f4d4b3cce98c1fdacb436c0056f65b58fe1bd345fc343067a58"
    },
    {
      "order": 2,
      "seed": 31418,
      "model": "omega_mlp",
      "run_id": "20260825T004941.947214-0500_omega_mlp_2852f5be",
      "config_hash": "2852f5be1e80f0a62df63a801de21cf3bf358718509c2f1e7977a086faec5deb",
      "final_validation_loss": 5.841490006235965,
      "final_validation_perplexity": 344.2919558314117,
      "total_wall_time_seconds": 34.48856058399542,
      "ordinary_compute_median_tokens_per_sec": 35870.826500395444,
      "ordinary_step_median_tokens_per_sec": 35154.77405413873,
      "peak_memory_bytes": 467143812,
      "console_log_sha256": "567fdd1df94a70d340f2be75282c1775f2221479abe87d028b10e463a54339dc",
      "metrics_jsonl_sha256": "8805298052ec8de381833fa12fb60d6c3137749ab8a9388c84c2f8663f62e82f",
      "manifest_sha256": "f5166eb52c2e3688f9f54fc86560007696d8531fa7458f05652f16eb084ea216",
      "final_model_sha256": "8227c48421ee391613c2a50ea8d17f7aa9e70cc4788ea4086e71705d3c886954"
    },
    {
      "order": 3,
      "seed": 31418,
      "model": "omega_mlp_diagonal",
      "run_id": "20260825T005034.494488-0500_omega_mlp_diagonal_f0ad3a19",
      "config_hash": "f0ad3a19695bedab08364da0c6aa521f359546a893a01c4934b5e97a36ada7f4",
      "final_validation_loss": 5.841077302210928,
      "final_validation_perplexity": 344.14989447211593,
      "total_wall_time_seconds": 38.492498084000545,
      "ordinary_compute_median_tokens_per_sec": 32459.763172497725,
      "ordinary_step_median_tokens_per_sec": 31203.33977705211,
      "peak_memory_bytes": 481657136,
      "diagonal_delta_rms": 0.011628348380327225,
      "diagonal_scale_min": 0.9570760726928711,
      "diagonal_scale_max": 1.041029691696167,
      "console_log_sha256": "b8e55c31ae87cc8ef66fd84416d4230305d2f28a2e30edc878f84b97bffb1424",
      "metrics_jsonl_sha256": "66afc5f851664b62d17feb4d982328301a5e9efc2e95e075b4e7bb02ed23d794",
      "manifest_sha256": "0ce4a8f34e505701aa667a7fc65fc3a4b07d2a5ecf06d8af04e850661c7bf24e",
      "final_model_sha256": "42e14bae4fc38f803d5487ce7d2d26a7fccf7626805cd014567d34cfa00b42d7"
    },
    {
      "order": 4,
      "seed": 31419,
      "model": "omega_mlp_diagonal",
      "run_id": "20260825T005149.273960-0500_omega_mlp_diagonal_d7b29cd3",
      "config_hash": "d7b29cd3473853e49d9e82776a90ebac925c1803461def8c0a3e12d66aca1a88",
      "final_validation_loss": 5.824723961288178,
      "final_validation_perplexity": 338.5676623963649,
      "total_wall_time_seconds": 38.92341912499978,
      "ordinary_compute_median_tokens_per_sec": 32090.673209970017,
      "ordinary_step_median_tokens_per_sec": 30842.9860988672,
      "peak_memory_bytes": 481657136,
      "diagonal_delta_rms": 0.011195043101906776,
      "diagonal_scale_min": 0.961900532245636,
      "diagonal_scale_max": 1.0371925830841064,
      "console_log_sha256": "b37e1a7ffb3f376c0e9587adb340b9c9a3cfe71c94d6115e79f444519996e86e",
      "metrics_jsonl_sha256": "792da0844a526e6c9963881202884314fbd21f22aef6180bada3c5d651e88b04",
      "manifest_sha256": "f2eb6a56f4f392defe5ec816bd312a0380b4ae1df4b1f61577192080c4c8bb2a",
      "final_model_sha256": "5f35c55a3845414c0d88f999b18441e2e57ec91af0bc428ecfdfa27b97ed9b17"
    },
    {
      "order": 5,
      "seed": 31419,
      "model": "omega_mlp",
      "run_id": "20260825T005236.415403-0500_omega_mlp_a58bbadf",
      "config_hash": "a58bbadfe0624eb324a1566bfeee25751de462bd0631c5e6237f5be87398f238",
      "final_validation_loss": 5.824466112396199,
      "final_validation_perplexity": 338.48037435380235,
      "total_wall_time_seconds": 33.730669040989596,
      "ordinary_compute_median_tokens_per_sec": 36174.58434260325,
      "ordinary_step_median_tokens_per_sec": 35447.856838341744,
      "peak_memory_bytes": 467143812,
      "console_log_sha256": "bcb2449decd15d84a2ea5921bba10b74a32eb33e3c3dbe5f4aaec295d434209e",
      "metrics_jsonl_sha256": "81427e2bb0389f0154da5f5f9ad4c8bb515776a52bdbdbd4d0b8fc7598c8db13",
      "manifest_sha256": "cbf17d48e8e72884ad9ebc7d66301b3362d5cb9e0b38bdd523c9638887728b9d",
      "final_model_sha256": "9abc655f3c7c49d033105902e57cc49b819196b06137e7325f8e02c1f5c57335"
    },
    {
      "order": 6,
      "seed": 31419,
      "model": "dense",
      "run_id": "20260825T005322.358768-0500_dense_25552a80",
      "config_hash": "25552a801917316d1f248ed2aed0735a921b986565addca2455c9cdeb812ef1a",
      "final_validation_loss": 5.743103905664453,
      "final_validation_perplexity": 312.0314255532815,
      "total_wall_time_seconds": 34.90163383301115,
      "ordinary_compute_median_tokens_per_sec": 35015.83032270805,
      "ordinary_step_median_tokens_per_sec": 34257.23033055544,
      "peak_memory_bytes": 685944873,
      "console_log_sha256": "5bd2a890913375a3823e75191223e033fa1d68f2c17986de6ff3fc1e91cedbf6",
      "metrics_jsonl_sha256": "823a4968cbdbc60b3eeaf4a90408dea5c214de65ec6b2cf7e5121e8be382cc46",
      "manifest_sha256": "4c5249c5689bf1968b836b3f448a3686485e577a6ff3c234935f4087f0a4247c",
      "final_model_sha256": "721e6f410ce8d97dc31b954a1adb1947f9dcead9f4117c35a1acbf3833e87808"
    }
  ],
  "aggregate": {
    "dense": {
      "final_validation_loss_mean": 5.735976717215687,
      "final_validation_perplexity_mean": 309.8232939401108,
      "total_wall_time_seconds_mean": 34.86727722900105,
      "ordinary_compute_median_tokens_per_sec_mean": 35011.54059766475,
      "ordinary_step_median_tokens_per_sec_mean": 34254.74776383803,
      "peak_memory_bytes_mean": 685944873
    },
    "omega_mlp": {
      "final_validation_loss_mean": 5.832978059316082,
      "final_validation_perplexity_mean": 341.38616509260703,
      "total_wall_time_seconds_mean": 34.10961481249251,
      "ordinary_compute_median_tokens_per_sec_mean": 36022.70542149935,
      "ordinary_step_median_tokens_per_sec_mean": 35301.31544624024,
      "peak_memory_bytes_mean": 467143812
    },
    "omega_mlp_diagonal": {
      "final_validation_loss_mean": 5.832900631749553,
      "final_validation_perplexity_mean": 341.35877843424043,
      "total_wall_time_seconds_mean": 38.70795860450016,
      "ordinary_compute_median_tokens_per_sec_mean": 32275.21819123387,
      "ordinary_step_median_tokens_per_sec_mean": 31023.162937959656,
      "peak_memory_bytes_mean": 481657136
    },
    "diagonal_relative_to_dense": {
      "physical_parameter_reduction_fraction": 0.22350435238798427,
      "estimated_active_parameter_reduction_fraction": 0.2280325707715486,
      "dense_operator_dimensionality_reduction_fraction": 0.13705583756345174,
      "estimated_flops_reduction_fraction": 0.20653609154929575,
      "mean_peak_memory_reduction_fraction": 0.2978194677752187,
      "ordinary_compute_median_throughput_change_fraction": -0.0781548700719954,
      "ordinary_step_median_throughput_change_fraction": -0.09433976417394274
    },
    "diagonal_relative_to_pure": {
      "ordinary_compute_median_throughput_change_fraction": -0.10403125435517324,
      "ordinary_step_median_throughput_change_fraction": -0.12118960594529982,
      "mean_final_validation_loss_change": -0.000077427566529
    }
  },
  "recovery": {
    "definition": "(pure_loss - diagonal_loss) / (pure_loss - dense_loss)",
    "seed31418": {
      "pure_dense_loss_gap": 0.11264047746904371,
      "recovery_fraction": 0.0036639051459157472
    },
    "seed31419": {
      "pure_dense_loss_gap": 0.08136220673174588,
      "recovery_fraction": -0.003169148211878982
    },
    "mean_recovery_fraction": 0.00024737846701838265,
    "mean_pure_dense_loss_gap": 0.0970013421003948,
    "mean_diagonal_dense_loss_gap": 0.09692391453386629
  },
  "gates": {
    "diagonal_beats_pure_both_seeds": false,
    "mean_recovery_at_least_0_50": false,
    "every_seed_recovery_at_least_0_25": false,
    "pooled_gap_at_most_half_pure_gap": false,
    "initialization_confound_triggered": false,
    "physical_parameter_reduction_at_least_0_20": true,
    "flops_reduction_at_least_0_18": true,
    "peak_memory_reduction_at_least_0_25": true,
    "diagonal_throughput_no_more_than_0_10_below_pure": false,
    "overall": "STOP"
  },
  "verification": {
    "metrics_rows_per_run": 2048,
    "manifests_completed": true,
    "console_and_structured_step_logging_present": true,
    "final_checkpoints_reloaded": true,
    "reloaded_full_validation_matches_recorded_exactly": true,
    "final_validation_targets_per_run": 217645,
    "prelaunch_test_suite": "70 passed",
    "focused_rescue_suite": "11 passed",
    "independent_prelaunch_qa_verdict": "pass"
  },
  "conclusion": {
    "confirmed_result": "At this matched-initialization two-seed screen, learned identity-post-diagonals recovered essentially none of the pure TensorAxis validation-loss gap and made the native-MLX structured arm slower.",
    "decision": "Stop the one-stage TensorAxis diagonal-rescue branch. Do not scale it or authorize custom Metal kernels for it.",
    "claim_ceiling": "This rejects only the tested single-stage post-diagonal rescue at this scale and budget; it does not reject the broader HELIX-Omega research program or support any capability-equivalence claim."
  },
  "limitations": [
    "This is a two-seed, 2,048-step early-convergence screen, not a scaling study.",
    "The project-specific word tokenizer makes perplexity non-comparable to other WikiText-2 tokenizations.",
    "The diagonal telemetry itself contributes some host-side per-step overhead; the measured throughput failure applies to the implemented instrumented path.",
    "This rescue experiment does not establish a broad edge-training efficiency claim."
  ]
}