{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-09-12-qwen38-efficient-variants-rtx5090",
  "complete": true,
  "completion_scope": "Bounded measurement, selection and restoration; not challenger qualification or a general code release.",
  "identity": {
    "repository_revision": "948346f5ab553361c6b61d9517b71a0e5e3cfe98",
    "gpu": "single NVIDIA RTX5090,32607MiB,sm120",
    "configuration": "configuration.json",
    "recipes": "recipes.toml",
    "controls": "nospec-controls.toml",
    "incumbent_advertised_id": "qwen38-27b-unsloth-ud-q4_k_xl-native-mtp3-262k",
    "provenance_errata": "provenance-errata.json"
  },
  "workload": {
    "manifest": "workload-manifest.json",
    "functional_checks": 6,
    "tool_batch": 20,
    "actual_needle_prompt_tokens": 47349,
    "thinking_enabled": {
      "questions": 10,
      "repetitions": 1,
      "max_completion_tokens": 5120
    },
    "thinking_disabled": {
      "questions": 10,
      "repetitions": 3,
      "max_completion_tokens": 1024
    },
    "capacity": {
      "requests": 5,
      "concurrency": 1,
      "nominal_input_tokens": 4096,
      "original_response_words": 128,
      "diagnostic_response_words": 32,
      "max_tokens": 512,
      "prompt_cache_mode": "unique",
      "request_canaries": true
    }
  },
  "measurement": {
    "path": "online-direct",
    "warm_cold_state": "Per-artifact; cold32-word subset has0 cached fraction; Swift32-word subset has0.9281 cached fraction and is excluded from cold graph.",
    "sample_count": "10 or30 quality attempts per recorded population;5 capacity requests per cell",
    "statistics": [
      "deterministic pass count",
      "total completion tokens",
      "descriptive N5 latency only"
    ],
    "timing_caveat": "Some diagnostic populations overlapped CPU repository checks; the runtime is pinned and shared, while model, quantization, context and speculation differ between recipe profiles. No isolated model or speculation speedup claim."
  },
  "results": {
    "thinking_enabled": {
      "signal_mtp3": {
        "source": "signal-mmlu-thinking.json",
        "source_sha256": "843987e54f065a1adfc512d5f68fe4ebf154fc317c18241bf9117ebb2250c999",
        "passed": 9,
        "attempts": 10,
        "completion_tokens": 10966,
        "summed_latency_ms": 99733.65590017056
      },
      "swift_mtp3": {
        "source": "swift-mmlu-thinking.json",
        "source_sha256": "a700ff731bdc6cadeb981d1942705fc75a095b4052487c4437a50e79fa0694a5",
        "passed": 9,
        "attempts": 10,
        "completion_tokens": 11020,
        "summed_latency_ms": 100207.72529998794
      },
      "qwopus_nospec": {
        "source": "qwopus-mmlu-thinking.json",
        "source_sha256": "fa8654aa26d009d03a349040340bd3a1eef5519ee6033762cd95a1801c16a6e7",
        "passed": 1,
        "attempts": 10,
        "completion_tokens": 12017,
        "summed_latency_ms": 214199.7431999771
      },
      "minitron_nospec": {
        "source": "minitron-mmlu-thinking.json",
        "source_sha256": "a3b0916a197d0b7ca5bed9e439d567804c07c65b73d5e8661724b8f8ad26255a",
        "passed": 3,
        "attempts": 10,
        "completion_tokens": 24580,
        "summed_latency_ms": 312535.5275000329
      },
      "incumbent": {
        "source": "incumbent-mmlu-thinking.json",
        "source_sha256": "b6ea12ab050fb2be66653e7a46da52afc4550e638259cf66104c402415f12c28",
        "passed": 8,
        "attempts": 10,
        "completion_tokens": 13129,
        "summed_latency_ms": 148428.03600005573
      }
    },
    "thinking_disabled": {
      "signal_nospec": {
        "source": "signal-nospec-quality-disabled.json",
        "source_sha256": "d3c0c021855c62083df6510a434a2da9ce74317a1b60f0b2ffd14f124f3c2d17",
        "passed": 24,
        "attempts": 30,
        "completion_tokens": 21027,
        "summed_latency_ms": 373216.3503000047
      },
      "qwopus_nospec": {
        "source": "qwopus-quality-disabled.json",
        "source_sha256": "f6534a476b14bf92b79803e8f3b845339cbbe4934b6982259c09f226929ae916",
        "passed": 24,
        "attempts": 30,
        "completion_tokens": 17094,
        "summed_latency_ms": 305933.05749981664
      },
      "minitron_nospec": {
        "source": "minitron-quality-disabled.json",
        "source_sha256": "cb2721eee9223ae73cd65cccf1eb1e2bf9620f0578d0a1b30a93bcbec183c094",
        "passed": 15,
        "attempts": 30,
        "completion_tokens": 18186,
        "summed_latency_ms": 231313.17329994636
      },
      "incumbent": {
        "source": "incumbent-mmlu-disabled.json",
        "source_sha256": "97d7536054a2d811f2591d21f7cd210478ab06c70e7c9a8da155ca827a4e7299",
        "passed": 21,
        "attempts": 30,
        "completion_tokens": 23250,
        "summed_latency_ms": 231467.87200012477
      }
    },
    "capacity_32": {
      "incumbent": {
        "source": "incumbent-capacity-32.json",
        "completed": 5,
        "requests": 5,
        "cached_fraction": 0,
        "ttft_mean_ms": 1571.4262999943458,
        "e2e_mean_ms": 1953.6777399829589,
        "decode_mean_tok_s": 115.13560746944977,
        "performance_eligible": true
      },
      "signal_nospec": {
        "source": "signal-nospec-capacity-32.json",
        "completed": 5,
        "requests": 5,
        "cached_fraction": 0,
        "ttft_mean_ms": 1861.7974399821833,
        "e2e_mean_ms": 2658.1093000015244,
        "decode_mean_tok_s": 55.255231000071305,
        "performance_eligible": true
      },
      "swift_nospec": {
        "source": "swift-nospec-capacity-32.json",
        "completed": 5,
        "requests": 5,
        "cached_fraction": 0.9281164847637406,
        "ttft_mean_ms": 546.2330399896018,
        "e2e_mean_ms": 1351.5927399974316,
        "decode_mean_tok_s": 54.63436663114783,
        "performance_eligible": true
      },
      "qwopus_nospec": {
        "source": "qwopus-capacity-32.json",
        "completed": 5,
        "requests": 5,
        "cached_fraction": 0,
        "ttft_mean_ms": 1999.8899800004438,
        "e2e_mean_ms": 2798.4088400029577,
        "decode_mean_tok_s": 55.10533356394137,
        "performance_eligible": true
      },
      "minitron_nospec": {
        "source": "minitron-capacity-32.json",
        "completed": 0,
        "requests": 5,
        "cached_fraction": null,
        "ttft_mean_ms": null,
        "e2e_mean_ms": null,
        "decode_mean_tok_s": null,
        "performance_eligible": false
      }
    }
  },
  "gates": [
    {
      "gate": "four distinct original candidates tested",
      "status": "passed"
    },
    {
      "gate": "original functional checks and tools",
      "status": "passed"
    },
    {
      "gate": "strict original128-word output",
      "status": "failed",
      "detail": "incumbent,Signal,Swift each0/5; Signal/Swift no-spec controls also0/5"
    },
    {
      "gate": "failure-free independent quality",
      "status": "failed",
      "detail": "all recorded MMLU populations contain failures; diagnostics are too small for broad quality parity"
    },
    {
      "gate": "actual57344 input plus8192 output reserve",
      "status": "missing",
      "detail": "nominal preflight produced47349 actual tokens"
    },
    {
      "gate": "incumbent262K contract/client/endurance/broad SWE parity",
      "status": "missing"
    },
    {
      "gate": "restoration",
      "status": "passed",
      "source": "restoration.json"
    },
    {
      "gate": "general lifecycle source release",
      "status": "held",
      "source": "lifecycle-review.json"
    }
  ],
  "headline_results": [
    {
      "claim": "Signal and Swift passed9/10 strict thinking-on questions versus incumbent8/10.",
      "source": "results.thinking_enabled"
    },
    {
      "claim": "Signal and Swift total completion tokens were16.5% and16.1% lower than incumbent in that small enabled sample.",
      "source": "results.thinking_enabled"
    },
    {
      "claim": "Thinking-off repeated counts:Signal24/30,Qwopus24/30,Minitron15/30,incumbent21/30.",
      "source": "results.thinking_disabled"
    },
    {
      "claim": "Revised cold32-word N5 means:incumbent1.954s,Signal no-spec2.658s,Qwopus2.798s; different serve profiles and uncontrolled CPU load.",
      "source": "results.capacity_32"
    }
  ],
  "failures": [
    "Strict128-word adherence failed before and after removing speculation.",
    "Signal's no-spec9216-token isolated retry still exhausted reasoning with no final answer.",
    "Qwopus exact output syntax failures and budget failures; stronger syntax/2048-token diagnostic7/10.",
    "Minitron factual mistakes, reasoning/visible budget failures, strict32-word0/5.",
    "Incumbent chemistry check failure, reasoning exhaustion, and three disabled-mode question budget failures remain visible."
  ],
  "limitations": [
    "Tiny fixed sanity fixture is not representative benchmark accuracy or statistical2%quality-loss proof.",
    "Separate reasoning-token counts are unavailable; only total completion tokens and reasoning characters are retained.",
    "64K challenger profiles are not262K replacements; multimodal/client/endurance/broad SWE gates unrun.",
    "No clean statistically adequate finalist performance population; cold chart is descriptive only.",
    "Swift commercial-license threshold verified; user eligibility unknown.",
    "No full model/runtime training repair attempted after bounded diagnostic rejection."
  ],
  "decision": {
    "evidence_labels": [
      "functional",
      "quality",
      "capacity",
      "external-prior"
    ],
    "decision_labels": [
      "current",
      "no-promotion"
    ],
    "selected_profile": "incumbent",
    "best_new_research_lead": "Signal Q6_K,not-qualified",
    "challenger_qualified": false,
    "promotion_authorized": true,
    "promoted": false,
    "rationale": "Signal/Swift improve this small token/quality diagnostic but fail advancement and full-deployment gates. Preserve the known262K incumbent; do not force promotion of an unqualified64K profile.",
    "next_gate": "New separately scoped full-context Signal recipe and broader independent quality/strict-output/client qualification."
  },
  "usage": {
    "limit_id": "codex",
    "initial_used_percent": 19,
    "last_observed_used_percent": 28,
    "stop_at_or_above_percent": 50,
    "threshold_reached": false
  },
  "evidence": [
    "README.md",
    "artifact-manifest.json",
    "source-registry.json",
    "run-plan.json",
    "restoration.json",
    "graph-manifest.json",
    "cold-output-graph-data.json",
    "friction-log.md"
  ]
}