{
  "schema": "anvil-serving.benchmark-campaign-state/v1",
  "campaign_id": "2026-09-12-qwen38-efficient-variants-rtx5090",
  "updated_at": "2026-09-12T20:12:29Z",
  "status": "complete",
  "current_stage": "publication",
  "launcher": {
    "invocation": "python -m anvil_serving.cli",
    "version": "1.0.0",
    "repository_revision": "948346f5ab553361c6b61d9517b71a0e5e3cfe98",
    "module_source": "anvil_serving/cli.py",
    "verified": true
  },
  "stages": [
    {
      "id": "research",
      "status": "complete",
      "evidence": [
        "source-registry.json"
      ]
    },
    {
      "id": "feasibility",
      "status": "complete",
      "evidence": [
        "feasibility.json"
      ]
    },
    {
      "id": "scout",
      "status": "complete",
      "evidence": [
        "summary.json"
      ],
      "failures": [
        "Original strict 128-word output fails; failures retained."
      ]
    },
    {
      "id": "finalist",
      "status": "not-run",
      "reason": "No new profile passed the replacement contract; retries are diagnostic, not finalist qualification."
    },
    {
      "id": "quality",
      "status": "complete-diagnostic",
      "evidence": [
        "summary.json"
      ],
      "limitations": [
        "Ten independent MMLU-Pro questions; not broad qualification.",
        "262K challenger context and real client integration not qualified."
      ]
    },
    {
      "id": "restoration",
      "status": "complete",
      "evidence": [
        "restoration.json",
        "restored-state.json"
      ]
    },
    {
      "id": "publication",
      "status": "complete",
      "evidence": [
        "artifact-manifest.json"
      ]
    }
  ],
  "active_assignments": [
    {
      "role": "incumbent",
      "model": "qwen38-27b-unsloth-ud-q4_k_xl-native-mtp3-262k",
      "state": "restored-healthy"
    }
  ],
  "completed_cells": [
    "four-original-profiles",
    "two-no-spec-controls",
    "incumbent-comparison",
    "bounded-failure-retries",
    "restoration"
  ],
  "next_actions": [],
  "promotion_authorized": true,
  "authority": "Operator explicitly authorized unattended sequential testing, selection and promotion of the best supported option, with up to three justified fix attempts per issue; the 50 percent Codex usage stop remains in force.",
  "candidate_containers_remaining": 0,
  "promoted": false,
  "decision": "Retain incumbent; Signal is the best new research lead, not a qualified replacement.",
  "usage": {
    "initial_used_percent": 19,
    "last_observed_used_percent": 28,
    "stop_at_used_percent": 50
  }
}