{
  "task": "Synthetic 32x32 visual answer+grounding policy",
  "rerun": "original CPU config regenerated 12 checkpoints, generated images, plots and GIF",
  "selection_rule": "all behavior GIFs use fixed seed 2 and hidden IDs 0/1/2/3/8/9; BoN RNG seed 902; counterfactual plot aggregates 3 seeds",
  "files": [
    {
      "file": "hw10_sft_demo.gif",
      "type": "image/gif",
      "method": "SFT",
      "seed": 2,
      "metric": "accuracy .552; grounding .523",
      "bytes": 39361
    },
    {
      "file": "hw10_grounding_demo.gif",
      "type": "image/gif",
      "method": "grounding-aware RL",
      "seed": 2,
      "metric": "accuracy .516; grounding .457",
      "bytes": 38843
    },
    {
      "file": "hw10_weak_demo.gif",
      "type": "image/gif",
      "method": "weak-verifier RL",
      "seed": 2,
      "metric": "accuracy .563; grounding .222; IoU .081",
      "bytes": 36992
    },
    {
      "file": "hw10_bon4_demo.gif",
      "type": "image/gif",
      "method": "Best-of-4 oracle",
      "seed": 2,
      "metric": "accuracy .743; 4x inference samples",
      "bytes": 38733
    },
    {
      "file": "hw10_counterfactuals.png",
      "type": "image/png",
      "method": "SFT/answer/grounding/weak paired counterfactuals",
      "seed": "0,1,2",
      "metric": "SFT original/no-image/occluded/swapped/conflict .552/.438/.370/.342/.236",
      "bytes": 35050
    },
    {
      "file": "hw10_shortcut_failure.png",
      "type": "image/png",
      "method": "SFT conflict shortcut",
      "seed": "0,1,2",
      "metric": "accuracy .552→.236",
      "bytes": 13478
    }
  ]
}