{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-09-19-glm53-apc",
  "complete": true,
  "identity": {
    "model": "brandonmusic/GLM-5.3-Flash-tr3-4bpw",
    "model_revision": "a5fee929cf4888b1824323e33e8a19b60129e025",
    "baseline": "r9-vision8-nospec",
    "candidate": "r10-vision8-apc-nospec",
    "bindings": "recipe-bindings.json"
  },
  "workload": {
    "finalist_requests_per_cell": 32,
    "concurrency": 4,
    "nominal_context_tokens": 32768,
    "actual_prompt_tokens_mean": {
      "unique": 25287.28125,
      "shared": 25043.0
    },
    "strict_response_words": 16,
    "canaries": true,
    "quality_repetitions": 3
  },
  "measurement": {
    "path": "online-direct",
    "warm_cold_state": "shared APC pool warmed before candidate finalist; unique population separate; restored no-cache control",
    "sample_count": 128,
    "statistics": [
      "mean visible TTFT",
      "mean first output",
      "completed requests per second",
      "mean E2E"
    ]
  },
  "gates": [
    {
      "name": "unique_throughput_regression_le_5_percent",
      "passed": true
    },
    {
      "name": "shared_mean_ttft_improvement_ge_15_percent",
      "passed": true
    },
    {
      "name": "three-repetition diagnostic quality",
      "passed": true
    },
    {
      "name": "vision and long-context functional gates",
      "passed": true
    },
    {
      "name": "restoration",
      "passed": true
    }
  ],
  "headline_results": [
    {
      "population": "unique",
      "candidate_percent_change": {
        "ttft_mean_ms": 4.015537882139819,
        "time_to_first_output_mean_ms": 6.218455129249567,
        "e2e_mean_ms": 2.7345730425953008,
        "completion_tokens_mean": 5.41573033707865,
        "prompt_tokens_mean": 0.0,
        "wall_clock_ms": 2.4459449735309002,
        "requests_per_second": -2.3875468903750696
      }
    },
    {
      "population": "shared",
      "candidate_percent_change": {
        "ttft_mean_ms": -67.84031736513431,
        "time_to_first_output_mean_ms": -73.147356383015,
        "e2e_mean_ms": -68.6601377759984,
        "completion_tokens_mean": 12.557022809123652,
        "prompt_tokens_mean": 0.0,
        "wall_clock_ms": -68.40948168384548,
        "requests_per_second": 216.55067827381228
      }
    }
  ],
  "failures": [
    "Original 64-word baseline: 2/8 strict word-count failures; excluded from performance claims.",
    "Original two-repetition quality checks were below the minimum of three; retained as diagnostics and superseded by matched three-repetition runs.",
    "The negative eight-request unique-prefix scout is retained."
  ],
  "limitations": [
    "Synthetic controlled-format workload, not real repository SWE.",
    "Reasoning volume differs despite identical visible word counts.",
    "Largest measured context is 309422 tokens at C1; four full-context requests were not tested.",
    "No video qualification.",
    "One 32-request cell per arm and cache mode; p99 is descriptive.",
    "Candidate unique-prefix native fingerprint is missing; external hash and identity binding are retained.",
    "The original swap mismatch did not reproduce after recreation; its cause is unproven."
  ],
  "decision": {
    "evidence_labels": [
      "functional",
      "capacity",
      "quality"
    ],
    "decision_labels": [
      "challenger",
      "no-promotion"
    ],
    "promotion_authorized": false,
    "next_gate": "Human approval of exact bounded APC promotion, then managed cutover and required client acceptance."
  },
  "evidence": [
    "matched-comparison.json",
    "candidate-shared32k-cache-delta.json",
    "baseline-quality-final.json",
    "candidate-quality-final.json",
    "reasoning-control.json",
    "candidate-eight-images.json",
    "candidate-long-context.json",
    "restoration.json",
    "independent-review.md"
  ]
}