{
  "study_id": "AM-RP-001",
  "plan_version": "0.1.0",
  "prepared_date": "2026-10-28",
  "author": "Anant Mishra",
  "status": "draft_preparation_not_executed",
  "assumption_note": "All workload sizes are proposed synthetic engineering assumptions, not observed demand or results.",
  "command_reference_version": "vLLM v0.31.0; reconcile installed help before execution",
  "sources": [
    "https://docs.vllm.ai/en/v0.31.0/cli/bench/serve/",
    "https://docs.vllm.ai/en/latest/benchmarking/cli/",
    "https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf"
  ],
  "capacity_profiles": [
    {
      "id": "short",
      "input_tokens": 512,
      "output_token_limit": 128
    },
    {
      "id": "medium",
      "input_tokens": 2048,
      "output_token_limit": 256
    },
    {
      "id": "long",
      "input_tokens": 8192,
      "output_token_limit": 256
    }
  ],
  "max_in_flight_requests": [
    1,
    4,
    8
  ],
  "attempts_per_trial": 100,
  "trial_seeds": [
    17,
    29,
    43
  ],
  "arrival_pattern": "saturated, bounded concurrency; not a model of observed user arrivals",
  "capacity_cache_policy": "Prefix caching disabled for the baseline, verified in the actual launch configuration. Warm-cache experiments require a distinct plan version and workload.",
  "warmup_requests_per_trial": 10,
  "warmup_in_measured_counts": false,
  "service_objectives": {
    "client_ttft_ms_max": null,
    "client_e2e_ms_max": null,
    "request_timeout_seconds": null,
    "quality_min_usable_fraction": null,
    "owner": null,
    "note": "Unfilled by design. Set and justify before collecting acceptance evidence. No default threshold implies suitability."
  },
  "quality_cases": "synthetic-quality-tasks.jsonl",
  "quality_rubric": {
    "0": "Wrong, unsupported, incomplete for the task, or an inappropriate answer instead of abstention.",
    "1": "Requires material human correction before use; no release-blocking failure.",
    "2": "Usable without material correction; expected behaviour, support, citations, and access boundary respected.",
    "release_blockers": [
      "restricted_content_disclosure",
      "invented_policy_presented_as_approved"
    ],
    "note": "Record dimension-level scores and critical failures separately. A mean cannot override a release blocker."
  },
  "required_trial_evidence": [
    "plan_sha256",
    "environment_manifest_sha256",
    "case_set_sha256",
    "command_and_client_version",
    "utc_start_and_end",
    "seed",
    "warmup_and_cache_record",
    "attempted_and_outcome_counts",
    "raw_per_request_records",
    "latency_percentile_method",
    "actual_input_and_output_token_counts",
    "quality_review_record",
    "failures_and_limitations"
  ],
  "execution_dependencies": [
    "Authorised hardware and test-window access",
    "Verified model and environment manifest",
    "Predeclared service objectives and failure criteria",
    "Installed CLI help and measurement definitions",
    "Representative permitted tasks and accountable quality reviewer"
  ]
}