{
  "schema": "anvil-investigation.prefill-phase-interpretation/v1",
  "source_revision": "6f32ec070f23ced9f50e704d854d775da52591ab",
  "status": "source-supported interpretation; no additional live requests",
  "findings": [
    {
      "claim": "dequant is a timeline bucket, not evidence of FP16 fallback or isolatedkernel cost. MMQ compute marks samebucket before gather, with hostqueue gaps charged to previousmark.",
      "source": [
        "src/prefill/prefill.cpp:1460",
        "src/prefill/prefill.cpp:2711",
        "src/prefill/prefill.cpp:2733"
      ]
    },
    {
      "claim": "Batched MTP callback is included in run_impl wall and aftereachchunk counter; should not explain outergap unless differentpathproof.",
      "source": [
        "src/prefill/prefill.cpp:2955",
        "src/prefill/prefill.cpp:2963",
        "src/program/generate.cpp:5374"
      ]
    },
    {
      "claim": "prompt_ms startsatrequest and endsafter lending, each promptpart including shortverifywindows, checkpointatturnboundary and finalrefill; these extend beyond timedbatchedrun.",
      "source": [
        "src/program/generate.cpp:6981",
        "src/program/generate.cpp:7028",
        "src/program/generate.cpp:7044"
      ]
    },
    {
      "claim": "Exact numeric lend/refill timing available with STRATA_TRACE; not yet exposed by approveddiagnosticparser. This is nextnarrow attribution surface, not proofthoseoperations causeallgap.",
      "source": [
        "src/program/generate.cpp:6785",
        "src/program/generate.cpp:6840"
      ]
    },
    {
      "claim": "host_staging_ms and ple_ms are cumulative PrefillStats across run_impl calls in one engine; wall/GPU phases and chunk setup/wait/afterchunk are perrun. Difference cumulativefields only within sameprocess, accountingforinterveningcalls.",
      "source": [
        "src/prefill/prefill.cpp:607",
        "src/prefill/prefill.cpp:2685",
        "src/prefill/prefill.cpp:3012",
        "src/prefill/prefill.cpp:3015"
      ]
    }
  ],
  "limitations": [
    "No physicalSSDclaimfrom file_blob counters",
    "No changingR36 orprefillceiling basedonlyaggregatephase",
    "No inferencealllayersfusedfromenv; mixedQ8downlayerscanremainMMQ",
    "Diagnostic timing overhead excludesheadlineperformance"
  ],
  "observations": {
    "mmq_large_calls": [
      {
        "tokens": 7410,
        "wall_ms": 12532,
        "after_chunk_ms": 9
      },
      {
        "tokens": 7412,
        "wall_ms": 10285,
        "after_chunk_ms": 7
      },
      {
        "tokens": 7413,
        "wall_ms": 10581,
        "after_chunk_ms": 7
      }
    ],
    "interpretation": "Batched draft callback included and measured7–9ms; cannot aloneexplain reported~12s outergap."
  }
}