{
  "artifact": "AgentSEO bounded SEO-agent architecture and recovery evaluation",
  "published_at": "2026-08-24",
  "contract_source": "lib/seo/agent-workflow.ts",
  "workflow_contract": {
    "decision_values": ["ready", "act_now", "monitor", "blocker"],
    "required_fields": [
      "summary",
      "decision",
      "confidence",
      "recommended_actions",
      "evidence",
      "limitations",
      "next_calls"
    ],
    "review_rule": "Route uncertain, exhausted, or consequential actions to human review; do not silently accept them."
  },
  "bounded_loop": [
    "collect structured search evidence",
    "validate the tool result and required fields",
    "produce a recommendation with evidence and limitations",
    "apply a retry budget",
    "route uncertain or exhausted runs to human review",
    "record the decision and next call"
  ],
  "evaluation": {
    "run_date": "2026-08-20",
    "method": "Deterministic synthetic control-flow simulation",
    "seeded_tasks_per_strategy": 3000,
    "total_strategy_runs": 12000,
    "assumptions": {
      "task_mix": {
        "routine": 0.55,
        "ambiguous": 0.25,
        "adversarial": 0.12,
        "tool_unstable": 0.08
      },
      "evaluator_true_positive_rate": 0.92,
      "evaluator_false_positive_rate": 0.09,
      "evaluator_cost_units": 0.25
    },
    "results": [
      {
        "strategy": "single_pass",
        "valid_accept_rate": 74.2,
        "invalid_accept_rate": 20.5,
        "human_review_rate": 0.0,
        "failure_rate": 5.3,
        "mean_cost_units": 1.0
      },
      {
        "strategy": "blind_retry",
        "valid_accept_rate": 78.5,
        "invalid_accept_rate": 21.2,
        "human_review_rate": 0.0,
        "failure_rate": 0.4,
        "mean_cost_units": 1.05
      },
      {
        "strategy": "evaluator_loop",
        "valid_accept_rate": 95.8,
        "invalid_accept_rate": 2.3,
        "human_review_rate": 0.0,
        "failure_rate": 1.9,
        "mean_cost_units": 1.72
      },
      {
        "strategy": "verified_fallback",
        "valid_accept_rate": 86.9,
        "invalid_accept_rate": 2.1,
        "human_review_rate": 11.0,
        "failure_rate": 0.0,
        "mean_cost_units": 1.66
      }
    ]
  },
  "reproducibility": {
    "script_repository_path": "scripts/run-agent-workflow-recovery-benchmark.ts",
    "seed_formula": "20260820 + task_id * 97",
    "source_summary_path": "docs/marketing/guest-posts/blogarena360/research/workflow-recovery-2026-08-20/summary.json"
  },
  "limitations": [
    "This is a synthetic control-flow evaluation, not production AgentSEO telemetry.",
    "The declared evaluator rates are assumptions, not measurements of a specific model.",
    "The result compares routing strategies under the same seeded attempts; it does not measure SEO recommendation quality.",
    "A human-review route reduces silent acceptance only when reviewers receive enough evidence and have authority to reject the action."
  ]
}