{
  "schema_version": "1.0",
  "generated_from": [
    "src/data/research/agent-architecture/methods.json",
    "src/data/research/agent-architecture/summary.json",
    "src/data/research/agent-architecture/diagnostic-summary.json",
    "src/data/research/agent-architecture/pilot-b/pilot-b-methods.json",
    "src/data/research/agent-architecture/pilot-b/pilot-b-summary.json",
    "src/data/research/agent-architecture/pilot-b/pilot-b-representative-traces.json"
  ],
  "figures": [
    {
      "id": "evidence-chain",
      "title": "Evidence chain from public clues to local tests",
      "subtitle": "Evidence roles stay explicit; product material proposes hypotheses, not internal implementation claims",
      "data_source": [
        "public-source review",
        "local experiment design"
      ],
      "evidence_role": "architecture_hypothesis",
      "rows": [
        {
          "stage": "Public evidence",
          "items": [
            "Official disclosures",
            "Research papers",
            "Open-source references"
          ],
          "claim_boundary": "hypothesis_only"
        },
        {
          "stage": "Engineering clues",
          "items": [
            "Context management",
            "State / legal actions",
            "Evidence and recovery"
          ],
          "claim_boundary": "hypothesis_only"
        },
        {
          "stage": "Local abstractions",
          "items": [
            "Contracted ReAct",
            "Workflow Controlled",
            "Goal-Graph Hybrid"
          ],
          "claim_boundary": "local_design"
        },
        {
          "stage": "Controlled tests",
          "items": [
            "Pilot A: rule tasks",
            "Pilot B: open tasks",
            "Code-based acceptance"
          ],
          "claim_boundary": "local_design"
        }
      ]
    },
    {
      "id": "architecture-flows",
      "title": "Shared base and three execution-control architectures",
      "subtitle": "Shared engineering base; Pilot A and Pilot B are separate two-arm comparisons, not one three-arm experiment",
      "data_source": [
        "Pilot A methods",
        "Pilot B methods",
        "local implementation"
      ],
      "evidence_role": "local_architecture_abstraction",
      "formal_comparisons": {
        "pilot_a": {
          "architectures": [
            "open_react",
            "workflow_controlled"
          ],
          "formal_runs": 30
        },
        "pilot_b": {
          "architectures": [
            "contracted_react",
            "goal_graph_hybrid"
          ],
          "formal_runs": 24
        }
      },
      "rows": [
        {
          "architecture": "Open execution family\nPilot A: Open ReAct\nPilot B: Contracted ReAct",
          "control_steps": [
            "Full trace",
            "Model selects next action",
            "Tool result",
            "Model decides stop"
          ],
          "shared_base": [
            "DeepSeek V4 Flash",
            "Frozen task",
            "Unified tool entry",
            "Code acceptance",
            "Trace recorder"
          ]
        },
        {
          "architecture": "Workflow Controlled",
          "control_steps": [
            "Explicit state",
            "Legal-action set",
            "Typed recovery",
            "Stop on pass"
          ],
          "shared_base": [
            "DeepSeek V4 Flash",
            "Frozen task",
            "Unified tool entry",
            "Code acceptance",
            "Trace recorder"
          ]
        },
        {
          "architecture": "Goal-Graph Hybrid",
          "control_steps": [
            "Goal dependencies",
            "Legal frontier",
            "Bounded branches",
            "Merge + verify"
          ],
          "shared_base": [
            "DeepSeek V4 Flash",
            "Frozen task",
            "Unified tool entry",
            "Code acceptance",
            "Trace recorder"
          ]
        }
      ]
    },
    {
      "id": "hypothesis-matrix",
      "title": "Architecture difference, hypothesis, and test matrix",
      "subtitle": "Three task families map control mechanisms to observable outcomes before reading Pilot B results",
      "data_source": [
        "Pilot B preregistered task matrix",
        "Pilot B methods"
      ],
      "evidence_role": "experiment_design",
      "rows": [
        {
          "task_family": "Alternative path",
          "architecture_difference": "Route history vs explicit dependency state",
          "test_hypothesis": "Can execution switch after a frozen route failure?",
          "primary_observations": "success · path switch calls · tokens"
        },
        {
          "task_family": "Open evidence",
          "architecture_difference": "Free evidence search vs tracked evidence state",
          "test_hypothesis": "Can key facts be retained without false evidence?",
          "primary_observations": "fact recall · false evidence repeated actions"
        },
        {
          "task_family": "Parallel decomposition",
          "architecture_difference": "Serial planning vs bounded independent branches",
          "test_hypothesis": "Can parallel work shorten the critical path without merge loss?",
          "primary_observations": "critical path · omissions success · total cost"
        }
      ]
    },
    {
      "id": "pilot-a-formal",
      "title": "Pilot A formal results",
      "subtitle": "DeepSeek V4 Flash · 5 scenarios × 3 repeats × 2 architectures · n=30 formal runs",
      "data_source": [
        "Pilot A reviewed public runs",
        "Pilot A summary"
      ],
      "evidence_role": "formal_experiment",
      "rows": [
        {
          "architecture": "react",
          "eligible_n": 15,
          "rate": 0.0,
          "successes": 0,
          "wilson_ci95": [
            0.0,
            0.20388330103584862
          ]
        },
        {
          "architecture": "workflow_controlled",
          "eligible_n": 15,
          "rate": 1.0,
          "successes": 15,
          "wilson_ci95": [
            0.7961166989641515,
            1.0
          ]
        },
        {
          "scenario": "clean",
          "label": "Clean",
          "react_tool_calls": [
            16,
            16,
            16
          ],
          "workflow_controlled_tool_calls": [
            7,
            7,
            7
          ]
        },
        {
          "scenario": "precondition",
          "label": "Precondition",
          "react_tool_calls": [
            16,
            16,
            16
          ],
          "workflow_controlled_tool_calls": [
            9,
            9,
            9
          ]
        },
        {
          "scenario": "transient_5xx",
          "label": "Transient 5xx",
          "react_tool_calls": [
            16,
            16,
            16
          ],
          "workflow_controlled_tool_calls": [
            8,
            8,
            8
          ]
        },
        {
          "scenario": "rate_limit",
          "label": "Rate limit",
          "react_tool_calls": [
            16,
            16,
            16
          ],
          "workflow_controlled_tool_calls": [
            9,
            9,
            9
          ]
        },
        {
          "scenario": "post_success",
          "label": "Post-success",
          "react_tool_calls": [
            16,
            16,
            16
          ],
          "workflow_controlled_tool_calls": [
            7,
            7,
            7
          ]
        }
      ]
    },
    {
      "id": "pilot-a-diagnostic",
      "title": "Pilot A post-experiment routing diagnostic",
      "subtitle": "One routing clarification · n=15 diagnostic runs · excluded from the formal Pilot A comparison",
      "data_source": [
        "Pilot A post-hoc diagnostic summary"
      ],
      "evidence_role": "post_hoc_diagnostic",
      "rows": [
        {
          "measure": "stage_reached",
          "stage": "Read inputs",
          "count": 15,
          "total": 15
        },
        {
          "measure": "stage_reached",
          "stage": "Write outputs",
          "count": 9,
          "total": 15
        },
        {
          "measure": "stage_reached",
          "stage": "Finish",
          "count": 9,
          "total": 15
        },
        {
          "measure": "stage_reached",
          "stage": "Verified",
          "count": 9,
          "total": 15
        },
        {
          "measure": "verified_by_scenario",
          "scenario": "clean",
          "count": 3,
          "total": 3
        },
        {
          "measure": "verified_by_scenario",
          "scenario": "precondition",
          "count": 0,
          "total": 3
        },
        {
          "measure": "verified_by_scenario",
          "scenario": "transient_5xx",
          "count": 3,
          "total": 3
        },
        {
          "measure": "verified_by_scenario",
          "scenario": "rate_limit",
          "count": 0,
          "total": 3
        },
        {
          "measure": "verified_by_scenario",
          "scenario": "post_success",
          "count": 3,
          "total": 3
        }
      ]
    },
    {
      "id": "pilot-b-results",
      "title": "Pilot B success, quality, and execution cost",
      "subtitle": "3 task families × 4 frozen instances × 2 architectures · n=24 paired formal runs",
      "data_source": [
        "Pilot B reviewed public summary"
      ],
      "evidence_role": "formal_experiment",
      "rows": [
        {
          "section": "family_success",
          "family": "alternative_path",
          "architecture": "contracted_react",
          "successes": 4,
          "eligible_n": 4,
          "success_rate": 1.0
        },
        {
          "section": "family_success",
          "family": "alternative_path",
          "architecture": "goal_graph_hybrid",
          "successes": 4,
          "eligible_n": 4,
          "success_rate": 1.0
        },
        {
          "section": "family_success",
          "family": "open_evidence",
          "architecture": "contracted_react",
          "successes": 4,
          "eligible_n": 4,
          "success_rate": 1.0
        },
        {
          "section": "family_success",
          "family": "open_evidence",
          "architecture": "goal_graph_hybrid",
          "successes": 4,
          "eligible_n": 4,
          "success_rate": 1.0
        },
        {
          "section": "family_success",
          "family": "parallel_decomposition",
          "architecture": "contracted_react",
          "successes": 0,
          "eligible_n": 4,
          "success_rate": 0.0
        },
        {
          "section": "family_success",
          "family": "parallel_decomposition",
          "architecture": "goal_graph_hybrid",
          "successes": 0,
          "eligible_n": 4,
          "success_rate": 0.0
        },
        {
          "section": "quality",
          "architecture": "contracted_react",
          "utility_mean": 75.0,
          "required_fact_recall_mean": 1.0,
          "false_evidence_count_mean": 0.0
        },
        {
          "section": "quality",
          "architecture": "goal_graph_hybrid",
          "utility_mean": 75.0,
          "required_fact_recall_mean": 1.0,
          "false_evidence_count_mean": 0.0
        },
        {
          "section": "cost",
          "metric": "model_calls",
          "unit": "calls",
          "contracted_react": 5.666666666666667,
          "goal_graph_hybrid": 15.583333333333334
        },
        {
          "section": "cost",
          "metric": "tool_calls",
          "unit": "calls",
          "contracted_react": 5.666666666666667,
          "goal_graph_hybrid": 4.833333333333333
        },
        {
          "section": "cost",
          "metric": "total_tokens",
          "unit": "tokens",
          "contracted_react": 3708.6666666666665,
          "goal_graph_hybrid": 12288.0
        },
        {
          "section": "cost",
          "metric": "elapsed_seconds",
          "unit": "seconds",
          "contracted_react": 5.4793617750013555,
          "goal_graph_hybrid": 17.087285999994492
        },
        {
          "section": "cost",
          "metric": "critical_path_rounds",
          "unit": "rounds",
          "contracted_react": 4.666666666666667,
          "goal_graph_hybrid": 3.6666666666666665
        }
      ]
    },
    {
      "id": "representative-divergence",
      "title": "Representative Pilot B trace and first divergence",
      "subtitle": "Open-evidence instance 01 · both runs verified · trace pair included in the reviewed public bundle",
      "data_source": [
        "Pilot B reviewed representative traces"
      ],
      "evidence_role": "representative_trace",
      "rows": [
        {
          "architecture": "contracted_react",
          "family": "open_evidence",
          "pair_id": "open_evidence-01",
          "verified_success": true,
          "selection_basis": "reviewed public representative trace bundle",
          "first_divergence_round": 6,
          "trace": [
            {
              "model_round": 1,
              "tool": "list_evidence",
              "disposition": "executed"
            },
            {
              "model_round": 2,
              "tool": "read_evidence",
              "disposition": "executed"
            },
            {
              "model_round": 3,
              "tool": "read_evidence",
              "disposition": "executed"
            },
            {
              "model_round": 4,
              "tool": "read_evidence",
              "disposition": "executed"
            },
            {
              "model_round": 5,
              "tool": "submit_result",
              "disposition": "executed"
            },
            {
              "model_round": 6,
              "tool": "finish",
              "disposition": "executed"
            }
          ]
        },
        {
          "architecture": "goal_graph_hybrid",
          "family": "open_evidence",
          "pair_id": "open_evidence-01",
          "verified_success": true,
          "selection_basis": "reviewed public representative trace bundle",
          "first_divergence_round": 6,
          "trace": [
            {
              "model_round": 1,
              "tool": "list_evidence",
              "disposition": "executed"
            },
            {
              "model_round": 2,
              "tool": "read_evidence",
              "disposition": "executed"
            },
            {
              "model_round": 3,
              "tool": "read_evidence",
              "disposition": "executed"
            },
            {
              "model_round": 4,
              "tool": "read_evidence",
              "disposition": "executed"
            },
            {
              "model_round": 5,
              "tool": "submit_result",
              "disposition": "executed"
            },
            {
              "model_round": 6,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 7,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 8,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 9,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 10,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 11,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 12,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 13,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 14,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 15,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 16,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 17,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            },
            {
              "model_round": 18,
              "tool": "submit_result",
              "disposition": "controller_rejected"
            }
          ]
        }
      ]
    },
    {
      "id": "architecture-decision",
      "title": "Task conditions to architecture choice",
      "subtitle": "Engineering decision aid derived from both pilots; recommendations are conditional, not a universal ranking",
      "data_source": [
        "Pilot A formal result",
        "Pilot B formal result",
        "engineering interpretation"
      ],
      "evidence_role": "conditional_recommendation",
      "rows": [
        {
          "condition": "stable and code-verifiable completion rules",
          "recommended_architecture": "workflow_controlled",
          "evidence": "Pilot A",
          "guardrail": "keep outcome verifier independent from the prescribed route"
        },
        {
          "condition": "independent subgoals with verifiable merge",
          "candidate_architecture": "goal_graph_hybrid",
          "recommendation_strength": "candidate_only",
          "evidence": "Pilot B negative result",
          "pilot_b_parallel_successes_per_arm": 0,
          "pilot_b_parallel_eligible_n_per_arm": 4,
          "pilot_b_parallel_merge_omissions_per_run": 3.0,
          "guardrail": "require a domain merge acceptance gate before adoption",
          "conservative_fallback": "contracted_react"
        },
        {
          "condition": "open path without stable workflow or safe parallel merge",
          "recommended_architecture": "contracted_react",
          "evidence": "Pilot B",
          "guardrail": "clear tool contract, complete trace, hard aggregate budget"
        }
      ]
    },
    {
      "id": "product-reference-architecture",
      "title": "Five-layer product reference architecture",
      "subtitle": "A deployable structure that separates business policy, execution control, tool access, evidence, and evaluation",
      "data_source": [
        "local architecture synthesis",
        "Pilot A and Pilot B instrumentation"
      ],
      "evidence_role": "product_reference_design",
      "rows": [
        {
          "layer": "Product layer",
          "components": [
            "Task intake",
            "Business policy",
            "Human handoff"
          ],
          "feedback_to_control": false
        },
        {
          "layer": "Control layer",
          "components": [
            "Planner",
            "State / goal graph",
            "Legal actions",
            "Recovery / stop"
          ],
          "feedback_to_control": false
        },
        {
          "layer": "Tool layer",
          "components": [
            "Unified entry",
            "Schema validation",
            "Timeout / retry",
            "Permission boundary"
          ],
          "feedback_to_control": false
        },
        {
          "layer": "Evidence layer",
          "components": [
            "Execution trace",
            "Evidence record",
            "Artifact versions",
            "Cost ledger"
          ],
          "feedback_to_control": false
        },
        {
          "layer": "Evaluation layer",
          "components": [
            "Code acceptance",
            "Quality metrics",
            "Regression set",
            "Release gate"
          ],
          "feedback_to_control": true
        }
      ]
    }
  ]
}