{
  "schema": "anvil-serving.capacity-comparison-summary/v1",
  "recorded_at": "2026-09-24T05:32:56.097178+00:00",
  "current_evidence": "native/mtp-uva3/capacity.json",
  "baseline_evidence": "native/baseline32k/capacity.json",
  "baseline_provenance": "baseline-provenance.json",
  "current": {
    "schema": "anvil-serving.benchmark/v1",
    "started_at": "2026-09-24T05:23:07Z",
    "finished_at": "2026-09-24T05:29:33Z",
    "requests": 100,
    "completed": 100,
    "failed": 0,
    "performance_eligible": true,
    "concurrency": 1,
    "context_tokens": 4096,
    "max_context_tokens": 131072,
    "max_tokens": 512,
    "controlled_output_policy": "strict",
    "prompt_cache_mode": "unique",
    "request_canaries": true,
    "response_words": 32,
    "serve_flags": {
      "no_thinking": false,
      "reasoning_effort": null,
      "shared_prefix_burst": false,
      "thinking_mode": "disabled"
    },
    "configuration_fingerprint": null,
    "measurement_protocol": "capacity-v3",
    "metric_population": {
      "aggregate_throughput_eligible": true,
      "attempted_requests": 100,
      "excluded_requests": 0,
      "failure_policy": "failed or contract-invalid responses retain sanitized timing and metadata but are excluded from performance distributions",
      "p99_interpretation": "tail-estimate",
      "p99_interpretation_by_metric": {
        "decode_tok_s": "tail-estimate",
        "e2e": "tail-estimate",
        "effective_prefill_tok_s": "tail-estimate",
        "tpot": "tail-estimate",
        "ttft": "tail-estimate"
      },
      "p99_minimum_recommended_samples": 100,
      "p99_samples_by_metric": {
        "decode_tok_s": 100,
        "e2e": 100,
        "effective_prefill_tok_s": 100,
        "tpot": 100,
        "ttft": 100
      },
      "performance_eligible_requests": 100,
      "successful_requests": 100
    }
  },
  "metrics": {
    "ttft_p50_ms": 3003.0448000179604,
    "ttft_p95_ms": 3043.130799982464,
    "ttft_p99_ms": 3067.83629997517,
    "decode_tok_s_p50": 51.88547695919587,
    "e2e_p50_ms": 3852.9262999945786,
    "e2e_p95_ms": 3894.42009999766,
    "e2e_p99_ms": 3927.072999998927,
    "throughput_tok_s": 11.678387852350253,
    "tpot_p50_ms": 19.272549999930728,
    "tpot_p95_ms": 19.420213636873417,
    "prompt_tokens_min": 3609.0,
    "prompt_tokens_max": 3687.0,
    "completion_tokens_min": 45.0,
    "completion_tokens_max": 45.0
  },
  "baseline_metrics": {
    "ttft_p50_ms": 455.44210000662133,
    "decode_tok_s_p50": 158.3551791480093,
    "e2e_p50_ms": 733.4115999983624,
    "throughput_tok_s": 61.14585265214693
  },
  "ratios_current_over_baseline": {
    "ttft_p50_ms": 6.593691711798934,
    "decode_tok_s_p50": 0.3276525418262464,
    "e2e_p50_ms": 5.253429724868249,
    "throughput_tok_s": 0.19099231339184222
  },
  "interpretation": [
    "Whole-recipe comparison: context, memory fraction and weight placement differ. No isolated causal claim.",
    "Nominal 4K/C1 short output with thinking off; not generalizable to long input, reasoning, vision or concurrency.",
    "TPOT/mean ITL are per-request derived proxies; p99 at n=100 is not a stable service SLA."
  ]
}