research_debate

Multi-stakeholder deliberation on a non-quant question — engineering / legal / business personas in parallel, then a chair synthesises a verdict. Demonstrates that MAF is a general platform: same EventBus, control plane, ReplanAgent, DecisionMemory plumbing as the trading arenas, just a different ``target_key`` and a different outbox. Trigger via control plane: XADD maf:control:in '*' data '{ "command": "run_arena", "args": { "arena": "research_debate", "target": { "question_id": "rfc-2026-042", "title": "Migrate session storage to ScyllaDB?", "text": "<the proposal markdown>" }, "emit_action": true } }' Decisions published to ``maf:decisions:out`` as ``GenericDecision``.

manual

Phase Pipeline

analysis

parallel

engineering (1 tools)

legal (1 tools)

business (2 tools)

→

synthesis

sequential

chair

→

replan_check

sequential

replan_controller

→

emit

sequential

Ticker

Date

Configure the arena's metadata, data sources, agents and phases. Changes write to the YAML file. Restart the app to apply.

Metadata

Description One-line summary shown on the dashboard. Schedule Cron expression. Leave blank for manual-only runs. Max iterations How many replan loops are allowed before giving up. Target key Drives which outbox publishes the verdict.

Data sources

Bind data adapters to symbolic names. Agents reference these names in their sources list.

Agents

Each agent runs a ReAct loop, calls its bound sources as tools, and emits a structured signal.

Phases

Phases run in order. parallel fans out specialists, sequential runs one-by-one, debate does N-way rounds + judge.

Smart triggers

Auto-run this arena when a Redis Stream event matches a rule. Rules run through safe_eval (boolean/comparisons, abs(), payload.field). Reference →

Ticker

Start Date

End Date

Edit arena configuration (JSON). Changes write to YAML file — restart to apply.

{
  "applicable_target_types": [],
  "description": "Multi-stakeholder deliberation on a non-quant question \u2014 engineering /\nlegal / business personas in parallel, then a chair synthesises a\nverdict. Demonstrates that MAF is a general platform: same EventBus,\ncontrol plane, ReplanAgent, DecisionMemory plumbing as the trading\narenas, just a different ``target_key`` and a different outbox.\n\nTrigger via control plane:\n  XADD maf:control:in \u0027*\u0027 data \u0027{\n    \"command\": \"run_arena\",\n    \"args\": {\n      \"arena\": \"research_debate\",\n      \"target\": {\n        \"question_id\": \"rfc-2026-042\",\n        \"title\": \"Migrate session storage to ScyllaDB?\",\n        \"text\": \"\u003cthe proposal markdown\u003e\"\n      },\n      \"emit_action\": true\n    }\n  }\u0027\n\nDecisions published to ``maf:decisions:out`` as ``GenericDecision``.\n",
  "llm": {
    "deep_model": "auto",
    "deep_provider": "ollama",
    "default_provider": "ollama",
    "providers": {
      "ollama": {
        "api_key": "529224e623bb4df6bd2d35ca70830b11.1OcHilLKpOBPZbD1VizrA21P",
        "base_url": "https://ollama.com/v1",
        "max_output_tokens": 4096,
        "model": "auto",
        "temperature": 0.3
      }
    },
    "quick_model": "auto",
    "quick_provider": "ollama"
  },
  "max_iterations": 2,
  "memory": {
    "chromadb_path": "./data/chromadb",
    "embedding_model": "text-embedding-004",
    "embedding_provider": "gemini",
    "instances": [
      {
        "name": "discussion_bm25",
        "persist_path": "",
        "type": "bm25"
      }
    ]
  },
  "name": "research_debate",
  "output_stream": "maf:arena:research_debate:output",
  "phases": [
    {
      "agents": [
        {
          "applicable_target_types": [],
          "extra": {},
          "llm_tier": "quick",
          "max_react_steps": 3,
          "name": "engineering",
          "role": "specialist",
          "sources": [
            "knowledge_base"
          ],
          "system_prompt": "",
          "system_prompt_file": "src/maf/arenas/discussion/prompts/engineering_persona.txt",
          "tools": []
        },
        {
          "applicable_target_types": [],
          "extra": {},
          "llm_tier": "quick",
          "max_react_steps": 3,
          "name": "legal",
          "role": "specialist",
          "sources": [
            "knowledge_base"
          ],
          "system_prompt": "",
          "system_prompt_file": "src/maf/arenas/discussion/prompts/legal_persona.txt",
          "tools": []
        },
        {
          "applicable_target_types": [],
          "extra": {},
          "llm_tier": "quick",
          "max_react_steps": 3,
          "name": "business",
          "role": "specialist",
          "sources": [
            "knowledge_base",
            "crowd_sim"
          ],
          "system_prompt": "",
          "system_prompt_file": "src/maf/arenas/discussion/prompts/business_persona.txt",
          "tools": []
        }
      ],
      "max_rounds": 1,
      "name": "analysis",
      "pattern": "parallel",
      "signal_extract": false,
      "transition": "synthesis"
    },
    {
      "agents": [
        {
          "applicable_target_types": [],
          "extra": {},
          "llm_tier": "deep",
          "max_react_steps": 5,
          "memory": "discussion_bm25",
          "name": "chair",
          "role": "synthesis",
          "sources": [],
          "system_prompt": "",
          "system_prompt_file": "src/maf/arenas/discussion/prompts/synthesis_persona.txt",
          "tools": []
        }
      ],
      "max_rounds": 1,
      "name": "synthesis",
      "pattern": "sequential",
      "signal_extract": false,
      "transition": "replan_check"
    },
    {
      "agents": [
        {
          "applicable_target_types": [],
          "extra": {
            "available_sources": [
              "knowledge_base",
              "crowd_sim"
            ],
            "confidence_floor": 0.55,
            "loopback_phase": "analysis"
          },
          "llm_tier": "deep",
          "max_react_steps": 5,
          "name": "replan_controller",
          "role": "replan",
          "sources": [],
          "system_prompt": "",
          "tools": []
        }
      ],
      "max_rounds": 1,
      "name": "replan_check",
      "pattern": "sequential",
      "signal_extract": false,
      "transition": "emit"
    },
    {
      "agents": [],
      "max_rounds": 1,
      "name": "emit",
      "pattern": "sequential",
      "signal_extract": false,
      "transition": "END"
    }
  ],
  "sources": [
    {
      "adapter": "fomo2_knowledge",
      "applicable_target_types": [],
      "config": {
        "days_back": 365,
        "max_results": 5
      },
      "name": "knowledge_base"
    },
    {
      "adapter": "mirofish_sim",
      "applicable_target_types": [],
      "config": {
        "stale_after_seconds": 86400
      },
      "name": "crowd_sim"
    }
  ],
  "target_key": "question_id",
  "triggers": []
}

Data sources bound to this arena. Agents access these via ReAct tool calls.

Source Name	Adapter	Module	Parameters	Used By
knowledge_base	fomo2_knowledge	fomo2	—	engineering, legal, business
crowd_sim	mirofish_sim	mirofish	—	business

Decision trail — every arena run is saved for review and auditing.

Loading memory data...

research_debate

Phase Pipeline

analysis

synthesis

replan_check

emit

Agent Signals

Execution Trace

Source Monitoring

Agent Reports

Debates

Decisions

Target Context

Metadata

Data sources

Agents

Phases

Smart triggers

Performance Metrics

vs Buy & Hold

Equity Curve

Trades

Signal Timeline