{
  "experiment": "h3_dual_trajectory_step3_probe",
  "seed": 33,
  "commit": "eb7840bb6ad59c0025dc7c655a9dd9e22f950500",
  "rl_kernel_repo": "RL-Kernel-worktrees/h3-rebuild",
  "execution_status": "complete",
  "device": "NVIDIA H100",
  "compute_capability": "sm90",
  "driver": "580.173.02",
  "torch": "2.13.0+cu130",
  "load_model_seconds": 50.788915812037885,
  "D_delta_at_step2": {
    "max_abs_latent_video_at_step2": 2.6226043701171875e-06,
    "max_abs_latent_audio_at_step2": 0.0
  },
  "arms": {
    "A": {
      "max_abs_velocity_video": 5.7220458984375e-06,
      "max_abs_velocity_audio": 0.0,
      "name": "A_replay_step3_entry"
    },
    "B": {
      "max_abs_velocity_video": 5.7220458984375e-06,
      "max_abs_velocity_audio": 0.0,
      "name": "B_shared_single_step2_latent"
    },
    "C": {
      "max_abs_velocity_video": 5.7220458984375e-06,
      "max_abs_velocity_audio": 0.0,
      "name": "C_shared_batched_step2_latent"
    }
  },
  "optional_delta_scale_sweep": [],
  "historical_b200_A_replay_from_pr_assets": {
    "max_abs_velocity_video_step3": 0.6215000152587891,
    "source": "pending/h3-full-transformer/reports/experiments/h3-full-transformer-b200/dual_trajectory_seed33.json",
    "note": "Use sm100 GPU run of this script to reproduce A/B on device; not substitutable with H100 forward."
  },
  "conclusion_b200_style": "inconclusive",
  "interpretation": {
    "supports_latent_accumulation": "B_shared max_abs << A_replay max_abs on video velocity at step 3",
    "supports_step3_layout_sensitivity": "B_shared max_abs still large (~same as A) with identical step-2 latent"
  },
  "retest_timestamp": "2026-10-23T15:24:00+02:00",
  "retest_note": "H100 rerun on mcml; safetensors from full_runs/c_run/seed33"
}