{
  "schema": "architectureiq-public-quality-summary-v1",
  "domain": "generation",
  "source_sha256": "8f0f9c236e74e6685918496d0ec78ef209bafdd72fb13ba6aaaa1b50ae3f9e39",
  "plan_sha256": "7b480b16becfa17a529174169e0a5e6fb709799a9ce31de9e414057da1a55cf4",
  "status": "complete_with_cross_device_limitation",
  "coverage": {
    "completed_seed_runs": 192,
    "expected_seed_runs": 192,
    "verified_selected_checkpoints": 3,
    "expected_selected_checkpoints": 4
  },
  "metric_scope": "Conditional synthetic image energy score. Lower is better; this is not natural-image FID.",
  "caveat": "A plausible silhouette may still have collapsed appearance variability. Compare energy, conditional geometry and diversity jointly. MC mean baseline is estimated from 16 independent oracle draws, not fitted to evaluation references.",
  "checkpoint_recomputations": [
    {
      "node_id": "candidate_ae7b6a4c1c102ac65308",
      "status": "verified",
      "seed": 240,
      "device": "cpu",
      "checkpoint_sha256": "e3c95e572e2a183cc2862fceae19025e7048db2d4f488766e5cf581ffae1a1b5",
      "diagnostics_sha256": "2ddcf9d59df1a0a56f4595f6b2535e6b7c16fdb11b37798f846703862cb3ef4c",
      "saved_samples_sha256": "98c5519123842634a86e20938250dc71db9be373cdde0041da181f7ce4caa1cb",
      "metric_absolute_error": 7.450580596923828e-09,
      "sample_max_absolute_error": 2.7418136596679688e-06
    },
    {
      "node_id": "candidate_e9681c5d795db8946e1c",
      "status": "verified",
      "seed": 240,
      "device": "cpu",
      "checkpoint_sha256": "54a6f4d908715b8bc0c2e56a94045ef59b43ef83e3cdec756df320a0577f4ed0",
      "diagnostics_sha256": "6fcf9e853dd0ab073fe3dc1480405abba5afd86512f182a6102df836c71a0459",
      "saved_samples_sha256": "8b1f9e05545b5bef793a946526a6bf5440a1967260e845ec4412c92c209efa17",
      "metric_absolute_error": 0.0,
      "sample_max_absolute_error": 4.6193599700927734e-06
    },
    {
      "node_id": "candidate_b83c1445855d4b8e3b87",
      "status": "verified",
      "seed": 240,
      "device": "cpu",
      "checkpoint_sha256": "3e222439047519d3868bd1f56563adf1fb6042a223331b0df28088b1d8127854",
      "diagnostics_sha256": "d2e030828ef7b4d63ce131d73090a1cf9f1ffe534183eb1393fca0c062e70649",
      "saved_samples_sha256": "00974d3e9b84b22fbbf84907ee948e2a11818d540019b6f1a460a66a14d15b23",
      "metric_absolute_error": 7.450580596923828e-09,
      "sample_max_absolute_error": 2.086162567138672e-06
    },
    {
      "node_id": "candidate_c154ceb7fdd2523d4ae3",
      "status": "recompute_mismatch",
      "seed": 240,
      "device": "cpu",
      "checkpoint_sha256": "0fccab4a09745cee0fa8e1404d281182cc4bb3d025be3d1c525733bbd7d3afc7",
      "diagnostics_sha256": "f7a117538ea92af9bc71d744ec5e285dc89cf297d76f69625c77a650e6233585",
      "saved_samples_sha256": "72a31fd8dfb51554f04d001c9aedfc554788b5139c3385eac4a056b517837ce2",
      "metric_absolute_error": 6.2212347984313965e-06,
      "sample_max_absolute_error": 0.001063227653503418
    }
  ],
  "cross_device_protocol": {
    "type": "post_hoc_cross_device_verification",
    "predeclared": false,
    "reason": "The preselected Transformer CPU replay exceeded the existing 0.001 pixel tolerance (0.00106323) while its energy error stayed within 0.0001. A same-checkpoint replay on the original GPU stack was then added as diagnostic evidence.",
    "original_cpu_thresholds": {
      "energy_absolute_error_max": 0.0001,
      "pixel_absolute_error_max": 0.001
    },
    "unchanged": [
      "dataset/condition/draw/seed selection",
      "training/model source",
      "all confirmation metrics and original samples",
      "CPU mismatch label",
      "CPU tolerance"
    ],
    "interpretation": "Three selected checkpoints passed independent CPU replay. Transformer failed the CPU pixel tolerance and exactly reproduced its saved results on the original GPU backend. This is not a CPU pass; the hardware/PyTorch/kernel cause is not uniquely isolated."
  }
}