{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-10-26-ignis-080-preflight",
  "complete": true,
  "completion_block": null,
  "identity": {
    "phase_one_source_commit": "dc576b25083d7cfdab1554db89379feee82346b5",
    "phase_two_harness_patch_sha256": "0c9b211fd6bf8d49493a2db46f5def4ebfe9a887d2470d8a74f6f37331b46206",
    "phase_two_reviewed_source_commit": "c727ff01f34d2d7564fab63d94203fec68890baf",
    "model": "gpillon/Qwen3.8-Flash-Next-trellis-a25-ignis",
    "model_revision": "8d94db83cd8e58d6be31e796723e5e35b960638c",
    "served_model": "qwen3.8-flash-next",
    "runtime": "Ignis 0.8.0",
    "runtime_image": "sha256:0226b62047113c219124ca4a02b1e52492402b44f5a3020db91e77756914b8cc",
    "hardware": "one NVIDIA RTX PRO 6000 Blackwell Max-Q; second equal GPU protected",
    "phase_two_runtime_changed": false,
    "phase_two_harness_only_change": true
  },
  "phase_one": {
    "tool_choice": "auto",
    "thinking": "default",
    "seed": null,
    "named_checks_passed": 4,
    "named_checks_total": 5,
    "observations_passed": 8,
    "observations_total": 9,
    "shared_prefix_tool_calls_passed": 3,
    "shared_prefix_tool_calls_total": 4,
    "terminal_engine_requests": 9,
    "literal_http_statuses_retained": 0,
    "outcome": "frozen auto-mode harness expectation unmet; not a protocol violation or engine/model defect"
  },
  "phase_two": {
    "oracle_revision": "tool-call/v2-exact-weather",
    "maximum_generation_requests": 16,
    "generation_requests": 16,
    "terminal_engine_requests": 16,
    "captured_http_200_tool_requests": 12,
    "all_sixteen_literal_http_200_supported": false,
    "mechanism_cells": {
      "passed": 3,
      "total": 3,
      "seed": 17
    },
    "selection": {
      "tool_choice": "required",
      "thinking": "default",
      "seed": 17,
      "rule": "prefer required/default when it passes; otherwise required/disabled; stop if neither passes"
    },
    "full_preflight": {
      "named_checks_passed": 5,
      "named_checks_total": 5,
      "observations_passed": 9,
      "observations_total": 9,
      "terminal_engine_request_ids": [
        4,
        5,
        6,
        7,
        8,
        9,
        10,
        11,
        12
      ]
    },
    "repeat": {
      "shared_prefix_tool_calls_passed": 4,
      "shared_prefix_tool_calls_total": 4,
      "seed": 18,
      "terminal_engine_request_ids": [
        13,
        14,
        15,
        16
      ]
    },
    "policy_boundary": "Use required as a request-level control for a step where a tool call is mandatory; do not make it a global coding/chat/agent default.",
    "auto_boundary": "The seeded auto 1/1 mechanism cell neither rescored the older unseeded 3/4 batch nor established auto reliability."
  },
  "measurement": {
    "path": "direct online OpenAI-compatible functional preflight",
    "evidence_label": "functional",
    "statistics": [],
    "performance_claims": [],
    "capacity_claims": [],
    "coding_or_swe_claims": []
  },
  "resources": {
    "candidate_gpu_memory_mib": 96027,
    "protected_gpu_memory_mib": 0,
    "phase_two_candidate_cgroup_peak_bytes": 44944740352,
    "phase_two_candidate_cgroup_peak_gib": 41.86,
    "phase_two_candidate_cgroup_oom_events": 0,
    "phase_two_host_mem_available_bytes_after_probes": 61764468736,
    "boundary": "GPU/host values are retained point observations; cgroup peak is a managed lifetime high-water mark. None is a capacity result or performance comparison."
  },
  "gates": [
    {
      "gate": "exact_candidate_identity",
      "status": "passed",
      "evidence": "configuration.json and tool-choice-followup.json"
    },
    {
      "gate": "phase_one_auto_all_four",
      "status": "failed",
      "evidence": "preflight-native.json and tool-failure-analysis.md"
    },
    {
      "gate": "phase_two_required_mechanism",
      "status": "passed",
      "evidence": "tool-choice-followup.json"
    },
    {
      "gate": "phase_two_selected_full_preflight",
      "status": "passed",
      "evidence": "tool-choice-followup.json"
    },
    {
      "gate": "phase_two_selected_seed_repeat",
      "status": "passed",
      "evidence": "tool-choice-followup.json"
    },
    {
      "gate": "phase_two_request_accounting",
      "status": "passed",
      "evidence": "tool-choice-followup.json"
    },
    {
      "gate": "phase_one_operational_restoration",
      "status": "passed",
      "evidence": "restoration.json and restoration-smoke.json"
    },
    {
      "gate": "phase_two_operational_restoration",
      "status": "passed",
      "evidence": "restoration.json and restoration-smoke-phase-two.json"
    },
    {
      "gate": "coding_campaign",
      "status": "not-run",
      "evidence": "prepared-followup.json"
    }
  ],
  "failures": [
    {
      "gate": "phase_one_auto_all_four",
      "status": "failed",
      "reason": "One of four unseeded auto-mode responses returned prose with finish_reason stop and no tool_calls; this is an unmet frozen harness expectation, not an engine/model defect."
    }
  ],
  "limitations": [
    "required is qualified only as a request-level policy for mandatory-tool steps on this exact pin",
    "one seeded auto pass does not establish auto reliability or overwrite the original 3/4 result",
    "no root cause or engine/model defect is assigned to the phase-one miss",
    "literal HTTP 200 is retained for 12 captured phase-two tool requests, not all 16 observations",
    "concurrent requests at a shared seed exercise reuse/concurrency but are not independent stochastic samples",
    "GPU/host point snapshots are not continuous peak-watermark, capacity, or performance evidence; cgroup peak is a managed lifecycle high-water mark",
    "the prepared 12-cell coding campaign was not run",
    "JSON prompt adherence does not establish response_format json_object support",
    "operational restoration passed; its single smoke is not full quality or strict-JSON acceptance"
  ],
  "decision": {
    "evidence_labels": [
      "functional"
    ],
    "decision_labels": [
      "no-promotion"
    ],
    "promotion_authorized": false,
    "promoted": false,
    "recommendation": "Use tool_choice required only on workflow requests whose contract requires a tool call; preserve ordinary auto/default behavior elsewhere and retain the phase-one auto miss as diagnostic evidence.",
    "next_gate": "Independent exact-head publication review; broader coding or client qualification remains separate."
  },
  "evidence": [
    "plan.json",
    "configuration.json",
    "preflight-native.json",
    "engine-request-summary.json",
    "resource-summary.json",
    "tool-failure-analysis.md",
    "tool-usage-contract.md",
    "tool-choice-followup.json",
    "prepared-followup.json",
    "restoration.json",
    "restoration-smoke.json",
    "restoration-smoke-phase-two.json"
  ],
  "metrics_coverage": null,
  "comparison_evidence": null
}