{
  "schema": "marker-gpu-trace-public-plan/v1",
  "campaign_id": "2026-10-26-marker-gpu-trace",
  "status": "frozen_before_lifecycle",
  "question": "For each accepted second closing marker, was its exact verification row budget-forced or sampled without a budget force?",
  "identity": {
    "source_commit": "626f6d8abe8f987a49e09ceab24651684e28ef4f",
    "candidate_image": "sha256:789be89fc1f16c28673fad1dbb539974f7d363c6c04ee1372a22fd12b21090d3",
    "runtime_source": "local-inference-lab/vllm@8244d1c44348ae8a5bbfc9f1b607021652283ed4",
    "upstream_split_top_p_commit": "0a1276e5e3d9000488e7590cc0957edd709f2152"
  },
  "constants": [
    "model and revision",
    "all model flags and MTP3",
    "GPU allocation and order",
    "runtime/model cache volumes",
    "TP2 and configured C4 capacity",
    "C1 request with 64-token reasoning budget, 512-token completion cap, maximum reasoning effort, temperature 0.6 and top-p 0.95"
  ],
  "candidate_changes": [
    "preceding upstream split-top-p patch unchanged",
    "bounded actual-GPU force/cache/acceptance/commit tracing",
    "preweight trace-kernel fixture before the existing sampler gate"
  ],
  "preweight_trace_gate": {
    "timeout_seconds": 120,
    "weights_loaded": false,
    "gpu_memory_limit_bytes": 67108864,
    "criteria": "Actual patched cache/budget kernel matches the uninstrumented baseline and expected force events on bounded marker-boundary fixtures."
  },
  "sampling_gate": {
    "forced_cases": 8,
    "ordinary_cases": 8,
    "ordinary_rows": 272,
    "seeds": [58231, 42],
    "vocabularies": [131072, 154856],
    "batch_sizes": [4, 64],
    "top_p": 0.95,
    "dtype": "float32",
    "policy": "upstream-count-v1",
    "one_predeclared_attempt": true,
    "child_timeout_seconds": 180,
    "tensor_peak_limit_bytes": 1073741824
  },
  "inference": {
    "maximum_probe_invocations": 1,
    "maximum_requests": 2,
    "request_timeout_seconds": 120,
    "total_window_seconds": 300,
    "retries": 0,
    "second_case_only": "The first response naturally closes below the boundary and answers correctly."
  },
  "trace_gate": {
    "required": "Exact output-token reconciliation for every TP rank, including explicit force writes, GPU marker-cache values, row mapping and draft positions, acceptance and postcommit boundary metadata.",
    "fail_closed": "Missing events, overflow, rank disagreement or nonreproduction remains unknown.",
    "scope": "Synchronized device-state snapshots are diagnostic-only and cannot support performance claims or retroactively prove an older request's internal state."
  },
  "restoration": "Restore the exact copied incumbent registry through the managed mode boundary and verify identity, readiness, router admission, resources and one frozen smoke.",
  "window": {
    "target_seconds": 2700,
    "candidate_startup_seconds": 900,
    "request_seconds": 300,
    "restoration_reserve_seconds": 1200
  },
  "ignis": "hold",
  "promoted": false,
  "projection": {
    "native_source_sha256": "383c1a90ea02d04b01da62f8f4d4e8fb7424fe4147f065d015d79a2655ffb539",
    "transformations": [
      "Removed private paths and operator-only command context.",
      "Preserved frozen limits, gates, identities and no-promotion boundary."
    ]
  }
}