{
  "approval_note": "handle bucket b. we have a doradusAI/doradusresearhc hugging face with consul kv crds already, with templates we wuse for releases, find and use it (+ prior: make it a doradus research release once its done and confirmed woring)",
  "approval_timestamp_utc": "2026-06-13T21:41:33Z",
  "source_dir": "/home/ogg130/open-source/evoquality-iqa-gguf",
  "target_repo": "Doradus-AI/EvoQuality-IQA-GGUF",
  "conversion": {
    "base_model": "ByteDance/EvoQuality",
    "base_license": "apache-2.0",
    "tool": "llama.cpp/convert_hf_to_gguf.py",
    "tool_commit": "b9010-d05fe1d7d",
    "quant": "Q8_0 (language tower) + f16 (mmproj sidecar)",
    "verified": "live llama-server load + image scoring smoke (placehold.co 512x512 png \u2192 '3', 371 tokens)",
    "host": "ai-backend (AB1) RTX PRO 6000"
  },
  "bench_run_timestamp_utc": "2026-06-13T22:17:48Z",
  "bench_results": {
    "BF16": {
      "plcc": 0.8033119711341278,
      "srcc": 0.7176936958279068,
      "n": 98,
      "size": "14.19 GB"
    },
    "Q8_0": {
      "plcc": 0.8037431282120499,
      "srcc": 0.7183392241452152,
      "n": 99,
      "size": "7.54 GB"
    },
    "Q6_K": {
      "plcc": 0.7924349784750757,
      "srcc": 0.7124182725088029,
      "n": 99,
      "size": "5.82 GB"
    },
    "Q5_K_M": {
      "plcc": 0.799887498812388,
      "srcc": 0.7158252217581657,
      "n": 99,
      "size": "5.07 GB"
    },
    "Q4_K_M": {
      "plcc": 0.7940155875576109,
      "srcc": 0.7137834695157875,
      "n": 99,
      "size": "4.36 GB"
    },
    "IQ4_XS": {
      "plcc": 0.7737659072872037,
      "srcc": 0.7028965143765948,
      "n": 99,
      "size": "3.96 GB"
    }
  },
  "bench_dataset": "AGIQA-3K test split (strawhat/agiqa-3k), n=99 stratified",
  "verification": "All 6 quants benchmarked. Q8_0 +0.05% vs BF16 PLCC (lossless). Q5_K_M -0.43% (sweet spot). All variants exceed upstream paper claims.",
  "private_to_public_transition": {
    "timestamp_utc": "2026-06-13T22:50:53Z",
    "authorization": "operator: we can deploy the huggingface link if weve loaded the quant and have proven it works for inference and its given role",
    "verification_evidence": [
      "Q8_0 single-shot smoke test on AB1 GPU4: placehold.co/512x512/png returns \"3\" with 371 tokens, 115 tok/s decode",
      "All 6 quants benched against AGIQA-3K (n=99 stratified): per-quant PLCC + SRCC documented in benchmarks-summary.json",
      "Q5_K_M deployed to swap-pool-32gb federated across ai-backend-2 (ab2-gpu2, ab2-gpu5) + ai-backend-3 (ab3-gpu1)",
      "Live smoke through Consul DNS swap-pool-32gb-evoquality-iqa.service.consul:8095 returns \"3\" with model alias swap-pool-32gb-evoquality-iqa, 371 tokens",
      "Bench-driven quant selection: Q5_K_M chosen as Pareto sweet spot (-0.43% PLCC vs BF16, 33% smaller than Q8_0)"
    ]
  },
  "reverted_to_private_at": "2026-06-13T22:53:09Z",
  "reverted_to_private_reason": "operator request \u2014 bench-validated but holding for additional verification (blog post + fleet wiring) before going public",
  "private_to_public_transitions": [
    {
      "timestamp_utc": "2026-06-13T23:02:47Z",
      "authorization": "its ok to make both public....do it",
      "target": "hf"
    }
  ]
}