{
  "schemaVersion": 1,
  "status": "complete",
  "generatedAt": "2026-08-21T09:39:18Z",
  "behaviorScreens": {
    "Q4_K_M": {
      "coverage": "100 harmful + 100 benign",
      "harmfulExplicitRefusals": 3,
      "harmfulFinalAnswers": 100,
      "benignExplicitRefusals": 0,
      "benignFinalAnswers": 100,
      "screenPasses": 197,
      "screenTotal": 200
    },
    "Q8_0": {
      "coverage": "100 harmful + 100 benign",
      "harmfulExplicitRefusals": 1,
      "harmfulFinalAnswers": 100,
      "benignExplicitRefusals": 0,
      "benignFinalAnswers": 100,
      "screenPasses": 199,
      "screenTotal": 200
    },
    "BF16": {
      "coverage": "12 harmful + 12 benign confirmation",
      "harmfulExplicitRefusals": 0,
      "harmfulFinalAnswers": 12,
      "benignExplicitRefusals": 0,
      "benignFinalAnswers": 12,
      "screenPasses": 24,
      "screenTotal": 24
    }
  },
  "matchedQuantizationDrift": {
    "protocol": {
      "reference": "abliterated BF16 GGUF",
      "context": "shared regular-BF16 greedy teacher trajectory",
      "cases": 36,
      "positions": 481,
      "vocabulary": 248320,
      "metric": "exact D_KL(P_BF16 || P_candidate) over all vocabulary logits",
      "backend": "llama.cpp Metal",
      "scope": "logit/output drift only; residual and KV/recurrent-state tensors are not exposed by this path"
    },
    "BF16SelfCheck": {
      "meanForwardKLNats": 2.9846421671368228e-09,
      "maximumForwardKLNats": 4.872053698777177e-08,
      "top1Agreement": 1.0
    },
    "Q8_0": {
      "meanForwardKLNats": 0.013586595574725264,
      "medianForwardKLNats": 0.0016423006845517534,
      "p95ForwardKLNats": 0.05919012966480697,
      "maximumForwardKLNats": 0.5282105315238043,
      "top1Agreement": 0.9625779625779626
    },
    "Q4_K_M": {
      "meanForwardKLNats": 0.07279847441333857,
      "medianForwardKLNats": 0.011983728648275512,
      "p95ForwardKLNats": 0.31770582342302545,
      "maximumForwardKLNats": 2.7962358145696853,
      "top1Agreement": 0.9064449064449065
    }
  },
  "interpretation": [
    "The phrase-based behavior screen is an early-refusal and answer-presence check, not a semantic safety, compliance, or answer-quality score.",
    "Most behavior-screen generations reached the 256-token cap, so completion counts should not be read as full-answer completion rates.",
    "Q8_0 remains substantially closer to BF16 than Q4_K_M on the exact matched-logit comparison."
  ],
  "repository": "PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-GGUF",
  "baseModel": "ornith-ai/Ornith-1.5-35B-A3B",
  "baseRevision": "e4dfb35a93d4b6822a811a7676f3488514abe7e2",
  "llamaCppRevision": "555881ebc8b0fc0402b30e09258a32a7bfd13c52",
  "conversionValidation": [
    {
      "variant": "BF16",
      "kind": "text_generation",
      "prompt": "What is the capital of France?",
      "response": "The capital of France is **Paris**.",
      "passed": true
    },
    {
      "variant": "Q4_K_M",
      "kind": "text_generation",
      "prompt": "What is the capital of France?",
      "response": "The capital of France is **Paris**.",
      "passed": true
    },
    {
      "variant": "Q8_0",
      "kind": "text_generation",
      "prompt": "What is the capital of France?",
      "response": "The capital of France is **Paris**.",
      "passed": true
    },
    {
      "variant": "Q4_K_M + mmproj F16",
      "kind": "image_generation",
      "image_sha256": "b35b975fbbfdea567e19862c9d1ecb70aa44348852b2de45e48ef4aaad2f87f5",
      "prompt": "What is the dominant color of this image? Answer with one color word.",
      "response": "Red",
      "passed": true
    }
  ],
  "capabilities": {
    "text": true,
    "visionProjectorIncluded": true,
    "q4VisionSmokePassed": true,
    "q8VisionSmokePassed": false,
    "bf16VisionSmokePassed": false,
    "nativeMtpHeadIncluded": false
  }
}