{
  "schemaVersion": 1,
  "backend": "cpu",
  "platform": "linux",
  "arch": "x64",
  "sourceOfTruth": "Runtime-ready means a built llama.cpp fork graph route selects the in-tree CPU op and real graph execution produces a finite, non-zero, deterministic result. For the CPU backend the score/decode arithmetic IS the C reference (packages/inference/reference/ + verify/qjl_polar_ref.c), so the standalone-fixture parity gate is `make reference-test`; this evidence file additionally records that the ops are reachable through ggml graph execution.",
  "status": "runtime-ready",
  "runtimeReady": true,
  "target": "linux-x64-cpu",
  "device": "Intel Core Ultra 9 275HX (24 cores, AVX2 + AVX-VNNI)",
  "atCommit": "eae44e75e82f70e85d9192e2e3cfa05fcbe80ab7",
  "smokeTarget": "cpu-dispatch-smoke",
  "smokeCommand": "make -C packages/inference/verify cpu-dispatch-smoke",
  "smokeSource": "verify/qjl_mt_check.c",
  "evidenceDate": "2026-05-11",
  "kernels": {
    "qjl": {
      "runtimeCapabilityKey": "qjl_full",
      "status": "runtime-ready",
      "runtimeReady": true,
      "graphOp": "GGML_OP_ATTN_SCORE_QJL",
      "builder": "ggml_attn_score_qjl",
      "smokeTarget": "cpu-dispatch-smoke",
      "smokeCommand": "make -C packages/inference/verify cpu-dispatch-smoke",
      "smokeOutputs": 24576,
      "bitMismatchesST_vs_MT": 0,
      "maxAbsDiffST_vs_MT": 0.0,
      "anyNaN": false,
      "refAllZero": false,
      "evidenceDate": "2026-05-11",
      "maxDiff": 0.0
    },
    "fused_attn": {
      "runtimeCapabilityKey": "fused_attn",
      "status": "runtime-ready",
      "runtimeReady": true,
      "graphOp": "GGML_OP_FUSED_ATTN_QJL_TBQ",
      "builder": "ggml_fused_attn_qjl_tbq",
      "note": "QJL-K score + online softmax + TBQ3-V mix; covers GGML_TYPE_QJL1_256 (K) and GGML_TYPE_TBQ3_0 (V) on the CPU backend. The standalone TBQ score op (GGML_OP_ATTN_SCORE_TBQ for turbo3/turbo4/turbo3_tcq) and the standalone Polar score op (GGML_OP_ATTN_SCORE_POLAR) are not exposed as public ggml builders in the fork's ggml.h \u2014 they are selected via the KV cache-type path; their CPU correctness is `make reference-test` (the C ops are the references) + the fused op above.",
      "smokeTarget": "cpu-dispatch-smoke",
      "smokeCommand": "make -C packages/inference/verify cpu-dispatch-smoke",
      "smokeOutputs": 32768,
      "bitMismatchesST_vs_MT": 0,
      "maxAbsDiffST_vs_MT": 0.0,
      "anyNaN": false,
      "refAllZero": false,
      "evidenceDate": "2026-05-11",
      "capabilityKey": "fused_attn",
      "maxDiff": 0.0
    }
  },
  "remaining": [
    "Wire probeKernels() in packages/app-core/scripts/build-llama-cpp-mtp.mjs to read this file for the CPU backend (mirror readVulkanRuntimeDispatchEvidence / vulkanCapabilityRuntimeReady) so a fresh linux-x64-cpu build's CAPABILITIES.json reports qjl_full runtime-ready instead of false. NOTE: this does NOT make the CPU build `publishable` \u2014 turbo3_tcq / polarquant remain false on the CPU backend (no CPU graph-dispatch op for the standalone TBQ/Polar score; the \u00a73 CPU kernel-completeness gate still fails the build by design). build-llama-cpp-mtp.mjs has uncommitted in-flight changes by a sibling workstream; the CPU-evidence probe wiring is left to that owner.",
    "kernel-contract.json runtimeStatus.cpu is now `runtime-ready` for the `qjl` kernel and the `fusedAttn` section (this evidence + a passing `make cpu-dispatch-smoke` back it; `make kernel-contract` green). The standalone TBQ/Polar score ops stay `reference-only` on cpu (validated by `make reference-test` \u2014 the C ops ARE the references)."
  ],
  "atCommitNote": "The committed packages/inference/llama.cpp submodule pin. eae44e75 = the prior 08032d57 (\"merge: W4-B CUDA QJL + Polar + TBQ3_TCQ kernels\") + one follow-up commit; the ggml_attn_score_qjl / ggml_fused_attn_qjl_tbq public ggml builders are identical between the two. The cached linux-x64-cpu fork build under packages/inference/llama.cpp/build/linux-x64-cpu/bin (or $ELIZA_STATE_DIR/local-inference/bin/mtp/linux-x64-cpu) provides libggml-{cpu,base}.so for the smoke.",
  "smokeRerun": "make -C packages/inference/verify cpu-dispatch-smoke  (no env vars: CPU_BIN_DIR + GGML_INC_DIR default to the in-repo fork build/checkout)"
}