{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-10-05-swift-96gb-baseline",
  "complete": true,
  "checkpoint": "Closed: Swift 8 GiB/eight-slot host-cache profile promoted after independent review; tested Flash candidate rejected.",
  "identity": {
    "configuration": "configuration-identity.json",
    "launcher": "launcher-identity.json"
  },
  "measurement": {
    "path": "direct online endpoint; separate CPU-only official SWE worker",
    "warm_cold_state": "Functional cache state uncontrolled; failed strict stress separate from completed unique-prefix-intent service-v2 cells.",
    "sample_count": {
      "preflight_repetitions": 3,
      "agentic": 18,
      "context128": 27,
      "context262": 9,
      "images": 12,
      "swe": 5,
      "capacity": 12,
      "service_latency_v2": 48,
      "swe_paired": 5,
      "flash_initial_service_v2": 12,
      "flash_mmq_diagnostic": 3,
      "swift_cache_pilot_per_arm": 18
    },
    "statistics": [
      "Historical service-v2 and initial Flash fields named median retain native nearest-rank p50; 12 attempts each; variable output, no p99 claim.",
      "Swift cache pilot reports both ordinary average-middle-pair medians and native nearest-rank p50 explicitly."
    ]
  },
  "headline_results": {
    "preflight_all_passed": true,
    "agentic_passed": 18,
    "context128_passed": 27,
    "context128_actual_prompt_max": 122788,
    "context262_passed": 9,
    "context262_actual_prompt_min": 253838,
    "context262_actual_prompt_max": 253868,
    "context_output_headroom_tokens": 8192,
    "images_passed": 12,
    "swe": {
      "attempted": 5,
      "graded": 5,
      "resolve_rate": 0.8,
      "resolved": 4
    },
    "capacity_completed": 0,
    "capacity_failed": 12,
    "capacity_performance_eligible": false,
    "service_latency_v2": [
      {
        "cell": "8k-c1",
        "artifact": "swift-service-v2-8k-c1.json",
        "completed": 12,
        "ttft_median_ms": 1733.2922000005055,
        "e2e_median_ms": 1837.6459000010072,
        "actual_prompt_min": 7415.0,
        "actual_prompt_max": 7433.0,
        "actual_output_min": 186.0,
        "actual_output_max": 396.0
      },
      {
        "cell": "8k-c4",
        "artifact": "swift-service-v2-8k-c4.json",
        "completed": 12,
        "ttft_median_ms": 3453.0902999995305,
        "e2e_median_ms": 3588.9893999992637,
        "actual_prompt_min": 7415.0,
        "actual_prompt_max": 7433.0,
        "actual_output_min": 142.0,
        "actual_output_max": 1075.0
      },
      {
        "cell": "24k-c1",
        "artifact": "swift-service-v2-24k-c1.json",
        "completed": 12,
        "ttft_median_ms": 3116.5836000000127,
        "e2e_median_ms": 3256.4419000009366,
        "actual_prompt_min": 22467.0,
        "actual_prompt_max": 22495.0,
        "actual_output_min": 195.0,
        "actual_output_max": 425.0
      },
      {
        "cell": "24k-c4",
        "artifact": "swift-service-v2-24k-c4.json",
        "completed": 12,
        "ttft_median_ms": 9502.766399999018,
        "e2e_median_ms": 9814.59679999898,
        "actual_prompt_min": 22467.0,
        "actual_prompt_max": 22495.0,
        "actual_output_min": 217.0,
        "actual_output_max": 470.0
      }
    ],
    "swe_paired": {
      "attempted": 5,
      "graded": 5,
      "resolve_rate": 0.8,
      "resolved": 4
    },
    "flash_initial": {
      "preflight_passed": true,
      "image_passed": 12,
      "service_artifact": "flash-m50-32k-service-v2-8k-c1.json",
      "requests_completed": 12,
      "practical_latency_gate": "failed",
      "visible_ttft_median_ms": 30405.80200000113,
      "e2e_median_ms": 30851.82150000037
    },
    "flash_mmq_diagnostic": {
      "artifact": "flash-diagnostic-mmq-8k-c1.json",
      "completed": 3,
      "eligible_for_official_gate": false,
      "reason": "instrumented diagnostic; three requests rather than frozen twelve-request cell"
    },
    "flash_trace_diagnostic": {
      "artifact": "flash-trace-8k-c1.json",
      "preflight_passed": false,
      "tools_passed": 19,
      "tools_attempts": 20,
      "diagnostic_requests": 3,
      "qualification_eligible": false
    },
    "flash_fixed4096_diagnostic": {
      "artifact": "flash-prefill4096-8k-c1.json",
      "requests": 3,
      "completed": 3,
      "preflight_families_passed": 5,
      "visible_ttft_median_ms": 27812.978400001157,
      "e2e_median_ms": 28217.62469999885,
      "eligible_for_official_gate": false
    },
    "swift_host_cache_pilot": {
      "comparison_artifact": "swift-host-cache-comparison.json",
      "requests_per_arm": 18,
      "matched_revisits_per_arm": 12,
      "passed": true,
      "control": {
        "requests": 12,
        "median_visible_ttft_ms": 9100.137449999238,
        "median_e2e_ms": 9226.973349999753,
        "reused_prompt_tokens": 124648,
        "nearest_rank_p50_visible_ttft_ms": 8966.76590000061,
        "nearest_rank_p50_e2e_ms": 9129.797699999472
      },
      "expanded": {
        "requests": 12,
        "reused_prompt_tokens": 373954,
        "median_visible_ttft_ms": 5352.872649998972,
        "nearest_rank_p50_visible_ttft_ms": 1752.577299997938,
        "median_e2e_ms": 5497.665400000187,
        "nearest_rank_p50_e2e_ms": 1861.3442999994732
      },
      "final_selection": "promoted",
      "limits": [
        "First pass interval sums include preflight intervals; later passes sum only newly captured native interval records. Occupancy is a sampled gauge, not an exact continuous peak.",
        "Throughput interval attribution assumes root-reported quiet>=6s capture boundaries; native IDs prove no intervening inference.",
        "All normal records are private until sanitized; native errors/startup can contain identifying free text.",
        "Observed state_transfers.*.bytes numerically equal operation counts in control logs. Preserve raw field names but do not use these as physical state-byte volumes without producer-unit verification; main_kv_transfers bytes/pages are separately reported.",
        "Two passes and12matched revisits per arm; sequential control then expanded, no p99 or broad coding/decode speed claim.",
        "Benefit is retention branch;20percent median TTFT improvement is not required if measured retention increases and E2E nonregression holds."
      ]
    },
    "swift_expanded_regression": {
      "service_comparison": "swift-cache-expanded-service-comparison.json",
      "service_completed": 48,
      "service_failed": 0,
      "all_midpoint_latency_ratios_at_most": 1.1,
      "agentic_passed": 18,
      "context262_passed": 9,
      "context262_actual_prompt_range": [
        253838,
        253868
      ],
      "context_output_headroom_tokens": 8192,
      "images_passed": 12,
      "deployment_acceptance": "passed; exact final startup, protocol, Pi, OpenClaw and Hermes checks retained"
    },
    "swift_explicit_history": {
      "artifact": "swift-explicit-history-outcome.json",
      "independent_fresh_answer_branches": 4,
      "passed": 4,
      "native_cached_tokens_each": 56963,
      "pressure_prompt_tokens": 311646,
      "device_kv_pool_tokens": 282112,
      "changed_interval_main_kv_h2d_bytes": 1455587328,
      "scope": "Explicitly cache-aware marked prior history only; group-level host restore, no per-request tier attribution or implicit-client generalization. No performance comparison.",
      "prior_attempts": "Two functionally successful attempts had zero reuse and do not pass the cached-continuation gate."
    }
  },
  "limitations": [
    "Functional cells overlapped bounded build/download activity; timings are not comparative performance.",
    "Synthetic agentic assertions are not executed repository-patch evidence.",
    "Five explicit SWE cases are a bounded sample, not the full Verified score.",
    "SWE image IDs were observed after task stage and after grading, not continuously pinned before every historical task.",
    "C4 is configured, not established by the serial context or image runs.",
    "Failed strict 128-word capacity does not establish ordinary prose streaming failure.",
    "Paired SWE replay is a distinct bounded attempt; original boundary-only image observations remain historical.",
    "Trace preflight had one unresolved tool connection failure; subsequent3requests are diagnosis only, despite matching timing components.",
    "Fixed4096 is three instrumented diagnostic requests, not a replacement twelve-request comparison; Flash expansion stopped after scout failure.",
    "Paired SWE wrapper retains missing_promotion_boundary; the separate final promotion decision supplies the campaign boundary without rewriting native historical evidence.",
    "Hermes does not expose context, maximum output or fallback observations; no parity claim for those fields.",
    "Native workflow owner reconciliation and scheduler convergence are separate unresolved work; the final router handoff preserves its immediately observed owner state."
  ],
  "decision": {
    "evidence_labels": [
      "functional",
      "quality",
      "capacity",
      "external-prior"
    ],
    "decision_labels": [
      "current",
      "rejected"
    ],
    "promotion_authorized": true,
    "next_gate": null,
    "flash_next": {
      "disposition": "rejected",
      "reason": "Frozen practical latency gate failed; bounded diagnostics did not establish a useful service improvement.",
      "not_run": [
        "128K context",
        "262K context",
        "C4 capacity",
        "agentic suite",
        "SWE finalist",
        "strict repeated-word stress"
      ],
      "not_run_reason": "Stopped after failed scout; these are not passing gates."
    },
    "promotion_evidence": "promotion-final.json",
    "current": "Swift same-image host KV 8192 MiB/eight slots, 40 GiB hard RAM with no extra swap; 262144/C4 and device pool unchanged"
  }
}