{
  "description": "strict single vs batched; diag arm uses --diag-video-projections (existing JSON, no GPU rerun for diag)",
  "pr_assets_pin": "4219a4e2de120647e0c80012e396d04822ca9f3e",
  "seeds": [
    11,
    22,
    33
  ],
  "rlkernel_strict_bitwise_equal_all_steps_with_diag": {
    "h100": true,
    "b200": true
  },
  "without_diag_strict_seed33_step3_max_abs_video_velocity": {
    "h100": 5.7220458984375e-06,
    "b200": 0.6215000152587891
  },
  "without_diag_provider_seed33_step3_max_abs_video_velocity": {
    "h100": 2.577500343322754,
    "b200": 0.960169792175293
  },
  "auditable_per_step_max": "pending/h3-full-transformer/reports/experiments/h3-full-transformer-h100/dual_trajectory_step_max.json",
  "mechanism_note": "Observed: unbound video proj_in/proj_out correlate with ~1e-6 layout drift; diag video-proj binding -> 0/16 on H100 and B200.",
  "euler_step3_probe": {
    "conclusion": "one_bf16_step_in_step3_input_then_spread_over_depth",
    "conclusion_detail": "B200 seed 33: the FP32 step-2 video latents of the two layouts differ by 1.9e-06, and after the bfloat16 cast at the step-3 input exactly 1 of 384 video elements differs (index [0,1,8], 0.40625 vs 0.408203125, 1.95e-03); audio and every other checked entry latent are identical after the cast. That single bf16 step reproduces the whole step-3 gap (arm P = arm A on every block, velocity 0.6215 video / 0.305 audio); with identical inputs (arms B, C) the layouts agree to 4e-06 to 5e-06. The relative difference of the block outputs grows about 300x over the 50 blocks (7.1e-04 after block 0, 0.22 after block 49), spread over the depth with no single sensitive layer. H100 seed 33: fp32 fork 2.6e-06 but 0 of 384 video elements differ after the cast, so the step-3 inputs are identical and the 5.7e-06 gap (arms A, B, C alike) is the layout-only gap. The layout-only gap is of the same size on both devices (B200 4.3e-06 / 5.4e-06). Do not summarise as a device sensitivity difference: the B200 spike needs the input element that straddles a bf16 rounding boundary. Script field conclusion_b200_style is a B<<A threshold only.",
    "h100_seed33_step3_entry_bf16": {
      "audit_json": "pending/h3-full-transformer/reports/experiments/h3-euler-step3-probe-h100/seed33_step3_entry_bf16_audit.json",
      "full_per_step_json": "pending/h3-full-transformer/reports/experiments/h3-euler-step3-probe-h100/step_entry_bf16_audit_full.json",
      "video_bf16_cells_differing_at_step3_entry": 0,
      "video_fp32_differing_at_step3_entry": 337,
      "video_numel": 384,
      "max_abs_fp32_fork": 2.6226043701171875e-06,
      "retest": "2026-10-23 mcml H100"
    },
    "h100_layerwise": {
      "report": "pending/h3-full-transformer/reports/experiments/h3-euler-step3-probe-h100/layerwise_report.json",
      "arm_A_velocity_video_max": 5.7220458984375e-06,
      "arm_B_velocity_video_max": 5.7220458984375e-06,
      "arm_E_velocity_video_max": 0.0,
      "arm_P_velocity_video_max": 0.0,
      "note": "E/P zero because 0 bf16 cells differ at step-3 entry; arm A equals layout-only gap"
    },
    "b200_seed33_step3_entry_bf16": {
      "video_elements_differing_after_bf16_cast": 1,
      "video_numel": 384,
      "element": {
        "index": [
          0,
          1,
          8
        ],
        "single_bf16": 0.40625,
        "batched_bf16": 0.408203125,
        "difference": 0.001953125
      },
      "audio_elements_differing_after_bf16_cast": 0,
      "audio_numel": 128,
      "source": "h3-euler-step3-probe-b200/README.md (Correction) and layerwise_report.json"
    },
    "b200_layerwise": {
      "report": "pending/h3-full-transformer/reports/experiments/h3-euler-step3-probe-b200/layerwise_report.json",
      "arm_P_relative_difference_block0": 0.000708,
      "arm_P_relative_difference_block49": 0.2245,
      "largest_single_block_growth": {
        "block 39": 1.77,
        "block 1": 1.74
      }
    },
    "b200_seed33_arms": {
      "step2_latent_video_max_abs": 1.9073486328125e-06,
      "step2_latent_audio_max_abs": 0.0,
      "A_replay_video_velocity_max": 0.6215000152587891,
      "A_replay_audio_velocity_max": 0.3051609992980957,
      "B_shared_single_video_velocity_max": 4.291534423828125e-06,
      "C_shared_batched_video_velocity_max": 5.424022674560547e-06
    },
    "reports": {
      "h100": "pending/h3-full-transformer/reports/experiments/h3-euler-step3-probe-h100/seed33_report.json",
      "b200": "pending/h3-full-transformer/reports/experiments/h3-euler-step3-probe-b200/seed33_report.json",
      "b200_readme": "pending/h3-full-transformer/reports/experiments/h3-euler-step3-probe-b200/README.md",
      "b200_layerwise": "pending/h3-full-transformer/reports/experiments/h3-euler-step3-probe-b200/layerwise_report.json",
      "h100_layerwise": "pending/h3-full-transformer/reports/experiments/h3-euler-step3-probe-h100/layerwise_report.json"
    },
    "limitations": [
      "single seed (33)",
      "one run per arm",
      "B200 one-element bf16 perturbation (arm P); H100 has 0 bf16 cells to perturb at step-3 entry (arms E/P trivially zero)",
      "H100 delta-scale K sweep not repeated on B200"
    ]
  }
}