{
  "$schema": "../../config/schema/capability-transform-policy.schema.json",
  "schemaVersion": 1,
  "source": "doppler",
  "capabilityTransforms": [
    {
      "kind": "lane-mismatch-guard",
      "dtypeEffect": "fail-closed",
      "match": {
        "modelId": "gemma-4-e2b-it-q4k-ehf16-af16-int4ple",
        "platformVendor": "apple"
      },
      "transforms": ["failClosedLaneMismatch"],
      "reason": "af16-int4ple lane requires retainQ4KMaterialization=true to dispatch the fused-q4k+f16 widetile pool, which produces intermittent NaN at L1.attn_core_out on Apple Metal - load the af32 manifest variant instead. Bug is intermittent and depends on buffer-pool aliasing patterns; confirmed re-emerged 2026-05-07 after e5c9a6a tightened bucket defaults.",
      "evidence": [
        "tests/demo/demo-models.test.js",
        "tests/inference/gemma4-e2b-q4k-apple-retain-regression.test.js"
      ]
    },
    {
      "kind": "platform-workaround",
      "dtypeEffect": "none",
      "match": {
        "modelId": {
          "in": [
            "gemma-4-e2b-it-q4k-ehf16-af32",
            "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
            "gemma-4-e2b-it-q4k-ehf16-af16-int4ple"
          ]
        },
        "platformVendor": "apple",
        "retainQ4KMaterialization": true
      },
      "transforms": ["disableRetainQ4KMaterialization"],
      "reason": "Fused-q4k FFN kernels under retain=mixed corrupt subsequent-layer attention compute on Apple Metal (intermittent NaN at L1.attn_core_out and beyond). Standalone-q4k projection kernels are clean; only the fused gate+up+down dispatch is the perturbation source.",
      "evidence": [
        "tests/config/capability-retain-q4k-policy.test.js",
        "tests/config/gemma4-e2b-conversion-config-contract.test.js",
        "tests/inference/gemma4-e2b-q4k-apple-retain-regression.test.js"
      ]
    },
    {
      "kind": "explicit-lane",
      "dtypeEffect": "selective-f16",
      "match": {
        "modelId": {
          "in": [
            "qwen-3-5-0-8b-q4k-ehaf16",
            "qwen-3-5-2b-q4k-ehaf16"
          ]
        },
        "requestedActivationDtype": "f16",
        "hasSubgroups": true,
        "hasF16": true
      },
      "transforms": ["useQwenF16PrimaryMatmuls"],
      "reason": "Runtime requested the promoted f16 primary lane for Qwen 3.5 small execution-v1 Q4K models with the shared linear/full attention graph.",
      "evidence": [
        "tests/integration/qwen-execution-v1-linear-decode-f16.test.js",
        "tests/integration/execution-graph-transforms.test.js"
      ]
    },
    {
      "kind": "explicit-lane",
      "dtypeEffect": "narrow-to-f16",
      "match": {
        "modelId": "qwen-3-6-27b-q4k-eaf16",
        "activationDtype": "f16",
        "mathDtype": "f16",
        "accumDtype": "f16",
        "hasSubgroups": true,
        "hasF16": true
      },
      "transforms": ["useQwen36F16Activations"],
      "reason": "Runtime requested the experimental all-f16 lane for Qwen 3.6 27B.",
      "evidence": [
        "tests/config/qwen36-f16-variant-identity.test.js",
        "tests/integration/execution-graph-transforms.test.js"
      ]
    },
    {
      "kind": "explicit-lane",
      "dtypeEffect": "selective-f16",
      "match": {
        "modelId": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
        "requestedActivationDtype": "f16",
        "hasSubgroups": true,
        "hasF16": true
      },
      "transforms": ["useGemma4Int4PleSelectiveF16Decode"],
      "reason": "Runtime requested the experimental selective f16 decode lane for Gemma 4 E2B INT4 PLE.",
      "evidence": [
        "docs/perf-investigations/gemma-4-int4ple-loop.md",
        "tests/config/gemma4-e2b-conversion-config-contract.test.js"
      ]
    },
    {
      "kind": "explicit-lane",
      "dtypeEffect": "narrow-to-f16",
      "match": {
        "modelId": "gemma-4-12b-it-text-q4k-ehf16-af16",
        "activationDtype": "f16",
        "mathDtype": "f16",
        "accumDtype": "f16",
        "hasSubgroups": true,
        "hasF16": true
      },
      "transforms": ["useGemma412BTextF16Activations"],
      "reason": "af16 manifest variant on shader-f16 GPU - dispatch the Gemma 4 12B text f16-residual lane with stable f32-Q/f16-KV prefill/decode attention boundaries and stable post-layer tail.",
      "evidence": [
        "tests/integration/gemma4-12b-f16-activations.test.js",
        "tests/integration/execution-graph-transforms.test.js"
      ]
    },
    {
      "kind": "lane-mismatch-guard",
      "dtypeEffect": "fail-closed",
      "match": {
        "modelId": "gemma-4-12b-it-text-q4k-ehf16-af32",
        "activationDtype": "f16",
        "mathDtype": "f16",
        "accumDtype": "f16",
        "hasSubgroups": true,
        "hasF16": true
      },
      "transforms": ["failClosedLaneMismatch"],
      "reason": "af32 manifest cannot dispatch f16 activations - load the af16 manifest variant or pick an f32 runtime profile.",
      "evidence": [
        "tests/integration/gemma4-12b-f16-activations.test.js",
        "tests/integration/execution-graph-transforms.test.js"
      ]
    },
    {
      "kind": "explicit-lane",
      "dtypeEffect": "narrow-to-f16",
      "match": {
        "modelId": "gemma-4-31b-it-text-q4k-ehf16-af16",
        "activationDtype": "f16",
        "mathDtype": "f16",
        "accumDtype": "f16",
        "hasSubgroups": true,
        "hasF16": true
      },
      "transforms": ["useGemma431BTextF16Activations"],
      "reason": "af16 manifest variant on shader-f16 GPU - dispatch the all-f16 lane.",
      "evidence": [
        "tests/integration/gemma4-31b-f16-activations.test.js",
        "tests/integration/execution-graph-transforms.test.js"
      ]
    },
    {
      "kind": "explicit-lane",
      "dtypeEffect": "narrow-to-f16",
      "match": {
        "modelId": "gemma-4-e2b-it-q4k-ehf16-af16-int4ple",
        "activationDtype": "f16",
        "mathDtype": "f16",
        "accumDtype": "f16",
        "hasSubgroups": true,
        "hasF16": true
      },
      "transforms": ["useGemma4Int4PleAf16Activations"],
      "reason": "af16-int4ple manifest variant on shader-f16 GPU - dispatch the all-f16 lane (q4_widetile_f16a + decode multicol_f16a + lm_head_gemv_subgroup_f16a).",
      "evidence": [
        "tests/config/gemma4-e2b-int4ple-f16-variant-identity.test.js",
        "tests/inference/gemma4-e2b-q4k-apple-retain-regression.test.js"
      ]
    },
    {
      "kind": "lane-mismatch-guard",
      "dtypeEffect": "fail-closed",
      "match": {
        "modelId": "gemma-4-31b-it-text-q4k-ehf16-af32",
        "activationDtype": "f16",
        "mathDtype": "f16",
        "accumDtype": "f16",
        "hasSubgroups": true,
        "hasF16": true
      },
      "transforms": ["failClosedLaneMismatch"],
      "reason": "af32 manifest cannot dispatch f16 activations - load the af16 manifest variant or pick an f32 runtime profile.",
      "evidence": [
        "tests/integration/gemma4-31b-f16-activations.test.js",
        "tests/integration/execution-graph-transforms.test.js"
      ]
    },
    {
      "kind": "capability-optimization",
      "dtypeEffect": "none",
      "match": {
        "modelId": {
          "in": [
            "gemma-4-e2b-it-q4k-ehf16-af32",
            "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
            "gemma-4-e2b-it-q4k-ehf16-af16-int4ple"
          ]
        },
        "activationDtype": "f32",
        "headDim": 256,
        "kvDtype": "f16",
        "hasF16": true,
        "maxWorkgroupStorageSize": { "gte": 16384 }
      },
      "transforms": ["useHead256SmallPrefillAttention"],
      "reason": "Gemma 4 E2B promotes only small-tile prefill attention onto the head256 kernel on capable GPUs.",
      "evidence": [
        "tests/config/gemma4-e2b-conversion-config-contract.test.js",
        "tests/integration/execution-graph-transforms.test.js"
      ]
    },
    {
      "kind": "explicit-lane",
      "dtypeEffect": "narrow-to-f16",
      "match": {
        "activationDtype": "f16",
        "hasSubgroups": false,
        "hasF16": true,
        "requiresF16ActivationNarrowing": true
      },
      "transforms": ["narrowToF16Activations", "removeSubgroups"],
      "reason": "Runtime requested f16 activations on a non-subgroup shader-f16 GPU.",
      "evidence": [
        "tests/integration/execution-graph-transforms.test.js",
        "tests/integration/execution-v1-capability-policy.test.js"
      ]
    },
    {
      "kind": "explicit-lane",
      "dtypeEffect": "narrow-to-f16",
      "match": {
        "activationDtype": "f16",
        "hasSubgroups": true,
        "hasF16": true,
        "requiresF16ActivationNarrowing": true
      },
      "transforms": ["narrowToF16Activations"],
      "reason": "Runtime requested f16 activations on a shader-f16 GPU.",
      "evidence": [
        "tests/integration/execution-graph-transforms.test.js",
        "tests/integration/execution-v1-capability-policy.test.js"
      ]
    },
    {
      "kind": "runtime-session-compatibility",
      "dtypeEffect": "full-f32",
      "match": { "kvDtype": "f32", "hasSubgroups": false, "hasF16": true },
      "transforms": ["removeSubgroups", "widenToF32Activations"],
      "reason": "Runtime KV cache resolved f32 on a non-subgroup shader-f16 GPU.",
      "evidence": [
        "tests/integration/execution-v1-capability-policy.test.js"
      ]
    },
    {
      "kind": "runtime-session-compatibility",
      "dtypeEffect": "full-f32",
      "match": { "kvDtype": "f32", "hasSubgroups": true, "hasF16": true },
      "transforms": ["widenToF32Activations"],
      "reason": "Runtime KV cache resolved f32.",
      "evidence": [
        "tests/integration/execution-v1-capability-policy.test.js"
      ]
    },
    {
      "kind": "hardware-compatibility",
      "dtypeEffect": "full-f32",
      "match": { "hasSubgroups": false, "hasF16": false },
      "transforms": ["removeSubgroups", "widenToF32Activations"],
      "reason": "GPU lacks subgroups and shader-f16.",
      "evidence": [
        "tests/integration/execution-v1-capability-policy.test.js"
      ]
    },
    {
      "kind": "hardware-compatibility",
      "dtypeEffect": "none",
      "match": { "hasSubgroups": false, "hasF16": true },
      "transforms": ["removeSubgroups"],
      "reason": "GPU lacks subgroups.",
      "evidence": [
        "tests/integration/execution-v1-capability-policy.test.js"
      ]
    },
    {
      "kind": "hardware-compatibility",
      "dtypeEffect": "full-f32",
      "match": { "hasSubgroups": true, "hasF16": false },
      "transforms": ["widenToF32Activations"],
      "reason": "GPU lacks shader-f16.",
      "evidence": [
        "tests/integration/execution-v1-capability-policy.test.js"
      ]
    },
    {
      "kind": "default-noop",
      "dtypeEffect": "none",
      "match": {},
      "transforms": [],
      "reason": "GPU supports all required features.",
      "evidence": []
    }
  ]
}