{
  "name": "Codex closed-model amodal decomposition baseline",
  "planner": {
    "model": "gpt-6-astra",
    "reasoning_effort": "max",
    "entrypoint": "codex exec"
  },
  "image_generation": {
    "tool": "image_gen__imagegen",
    "entrypoint": "codex exec",
    "backend_model": "not exposed by the tool",
    "requested_family": "OpenAI flagship image generation",
    "model_lock_available": false
  },
  "inputs": 48,
  "pilot": [
    "vi_001",
    "vi_033",
    "vi_019"
  ],
  "policy": {
    "inventory": "Only original poster, no existing candidate masks or Qwen outputs",
    "layer_count": "Determined by independent editable objects and text blocks; no fixed cap",
    "generation": "One original-poster-conditioned request per foreground layer, plus one clean background request",
    "compositing": "Back-to-front using predicted z order; preserve original relative coordinates; record any size normalization",
    "selection": "First RGB output and first opacity output are retained, including visually poor results. A separate minimal transparency probe is excluded from the main batch and retained as capability evidence. Infrastructure retries only, with logs.",
    "occlusion": "Plausible hidden-region generation, not verified recovery of ground truth"
  },
  "evaluation": [
    "Object separation and unwanted background",
    "Visible content and position preservation",
    "Occlusion completion quality",
    "Vietnamese glyph/diacritic fidelity",
    "Alpha channel validity",
    "Reconstruction compared with the input"
  ],
  "publication": "New English Hugging Face Space; existing Space remains available for comparison",
  "docs": [
    "https://developers.openai.com/api/docs/models/gpt-6-astra",
    "https://developers.openai.com/api/docs/models/gpt-image-2",
    "https://developers.openai.com/api/docs/guides/image-generation"
  ],
  "opacity_stage": {
    "reason": "Native image tool returned RGB with painted transparency grids in two pilot layers and a separate minimal-prompt probe.",
    "model": "Same Codex built-in image editing tool; exact model not exposed",
    "input": "The first generated RGB layer",
    "output": "A separate grayscale opacity mask",
    "packing": "RGB channels unchanged; grayscale mask copied to alpha. Bilinear mask resize only if native dimensions differ, explicitly recorded. No thresholding or morphology.",
    "publication": "Expose raw RGB, generated mask, and derived RGBA separately. Derived RGBA is not described as native model output.",
    "pilot_masks": "vi_001/L08 and vi_033/L07 used explicit subject descriptions during opacity-stage validation; prompts and results are retained."
  },
  "previous_qwen_reference": {
    "space": "Elfsong/poster-layer-decomposition-48",
    "sha": "6cc04008dd417615990eec337b72a6f91a4fa230",
    "comparison": "Different candidate inventories and resolutions; qualitative comparison only"
  },
  "quota_recovery": {
    "interrupted_invocations": 478,
    "retained_outputs_recovered": 15,
    "method": "Unique PNGs in the interrupted Codex session directory are retained unchanged and SHA-256 verified. Host-written result metadata is explicitly marked; exact submitted image prompt and tool-call count are unavailable for these outputs. Only stages without retained outputs are retried after the user replenished quota.",
    "record": "quota_recovery.json"
  }
}