{
  "schema": "lmruntime.capability-matrix.v3",
  "release": "3.4.0",
  "canonical_url": "https://lmruntime.com/capabilities/",
  "status_url": "https://lmruntime.com/status/",
  "evidence_url": "https://lmruntime.com/evidence-benchmarks/",
  "getting_started_url": "https://lmruntime.com/getting-started/",
  "packages_url": "https://lmruntime.com/packages/",
  "api_reference_url": "https://lmruntime.com/api-reference/",
  "distribution_status": "public-nuget-package-family",
  "package_ids": [
    "UAIX.LmRuntime.Abstractions",
    "UAIX.LmRuntime.Tensors",
    "UAIX.LmRuntime.Acceleration",
    "UAIX.LmRuntime.Gguf",
    "UAIX.LmRuntime.Sampling",
    "UAIX.LmRuntime.Kernels.Cpu",
    "UAIX.LmRuntime.Tokenization",
    "UAIX.LmRuntime.Models.Llama",
    "UAIX.LmRuntime.Backends.CpuManaged",
    "UAIX.LmRuntime.Backends.Cuda",
    "UAIX.LmRuntime.Backends.DirectML",
    "UAIX.LmRuntime.Backends.Vulkan",
    "UAIX.LmRuntime.Backends.Rocm",
    "UAIX.LmRuntime.Backends.Metal",
    "UAIX.LmRuntime.Backends.Cuda.Native.win-x64",
    "UAIX.LmRuntime.Backends.Cuda.LegacyK80.win-x64",
    "UAIX.LmRuntime.LocalEndpoint"
  ],
  "target": {
    "language": "C#",
    "runtime_form": "managed-cpu-local-with-explicit-backend-contracts",
    "model_container": "GGUF",
    "model_path": "LLaMA-family",
    "application_entry": "UAIX.LmRuntime.LocalEndpoint",
    "backend_control_plane": "UAIX.LmRuntime.Acceleration",
    "managed_cpu_backend": "UAIX.LmRuntime.Backends.CpuManaged"
  },
  "documented_capabilities": [
    "runtime-neutral inference, tokenizer, diagnostics, and governance contracts",
    "tensor shapes, storage traits, quantized block metadata, and reference vector math",
    "bounded GGUF parsing, validation, hashing, sharding, and mapped access",
    "GGUF-driven tokenization, special-token handling, chat templates, truncation, and parity tools",
    "greedy and probability sampling, logit processing, deterministic state, stops, and generation control",
    "managed CPU reference, vector, half-precision, quantized, dispatcher, and parity surfaces",
    "LLaMA-family configuration, tensor binding, reference sessions, KV cache, persistence, and parity surfaces",
    "backend registration, capability declarations, local probes, deterministic selection policies, diagnostics, devices, and visible CPU fallback",
    "managed CPU backend identity and no-native-asset availability",
    "CUDA, DirectML, Vulkan, ROCm, and Metal registration and fail-closed diagnostic surfaces",
    "separate CUDA native-asset package identities with no managed API surface",
    "verified local model loading, isolated LocalEndpoint sessions, UAIX context evidence, and deterministic greedy generation"
  ],
  "backend_boundary": {
    "selectable_now": [
      "UAIX.LmRuntime.Backends.CpuManaged"
    ],
    "registration_and_diagnostics": [
      "UAIX.LmRuntime.Backends.Cuda",
      "UAIX.LmRuntime.Backends.DirectML",
      "UAIX.LmRuntime.Backends.Vulkan",
      "UAIX.LmRuntime.Backends.Rocm",
      "UAIX.LmRuntime.Backends.Metal"
    ],
    "native_asset_identities": [
      "UAIX.LmRuntime.Backends.Cuda.Native.win-x64",
      "UAIX.LmRuntime.Backends.Cuda.LegacyK80.win-x64"
    ],
    "rule": "Package presence, declared compatibility, successful probing, executable adapter availability, and model execution evidence are separate states."
  },
  "host_owned": [
    "model acquisition and licensing",
    "native driver and adapter deployment",
    "backend execution and parity evidence",
    "profile or memory import",
    "prompt assembly",
    "tools and commands",
    "network and provider access",
    "transport and server hosting",
    "persistence and audit",
    "user interface and approval",
    "production operations and support"
  ],
  "claim_boundary": "Published APIs and package documentation do not create blanket model, platform, GPU execution, performance, security, support, or production-readiness guarantees.",
  "write_capabilities": [],
  "generated_utc": "2026-06-24T01:36:14Z"
}