research-debate

name	research-debate
description	Adversarial bull/bear review of a draft Atlas pick. Forks a bull researcher and bear researcher (configurable rounds), then a Research Manager synthesizes the debate into a structured 5-tier rating (Buy/Overweight/Hold/Underweight/Sell). Use when a pick needs stress-testing before CC commits capital — `python main.py picks "<query>" --debate N` triggers it.
triggers	["debate","bull bear","adversarial review","stress test","devil's advocate","counterargument","second opinion","peer review","research debate"]
tier	research
dependencies	["risk-debate","realised-return-reflection"]
origin	TauricResearch/TradingAgents (patterns)

Research Debate — Bull vs Bear Adversarial Review

Purpose

Atlas's default stock_picker.py is single-pass — one Claude call drafts a thesis from assembled context. That's fast and usually right, but it skips the "what could go wrong" pressure test that real desks run before committing capital.

This skill adds a structured bull/bear debate on top of the draft thesis, then has a Research Manager synthesize the debate into a 5-tier rating. The pattern is lifted from TauricResearch/TradingAgents (tradingagents/agents/researchers/) and reimplemented in pure Python — no LangGraph dependency.

When to Trigger

• User says "/debate " or "stress test the pick" or "give me a bear case"
• Before committing real capital on a pick with conviction ≥ 7
• When a thesis depends on a single catalyst (consensus + risk both matter)
• When CC explicitly says "second opinion" or "peer review"

How It Works

Draft thesis (from StockPickerAgent)
            │
            ▼
    Bull researcher  ◄──┐
            │           │ N rounds
            ▼           │
    Bear researcher  ◄──┘
            │
            ▼
    Research Manager (deep_think tier)
            │
            ▼
    Rating + synthesis + lesson

Each round = one bull turn + one bear turn. Default max_rounds=1 (2 exchanges). Higher counts produce diminishing returns and cost more.

Output Format

Structured triple:

• Rating — one of Buy / Overweight / Hold / Underweight / Sell
• Synthesis — 4-8 sentence prose explaining which side won and how the rating resolves
• Lesson — 1-2 sentences on what part of the draft thesis needed revision (gets fed into the next reflection)

The Pick dataclass picks up four new fields when debate runs: pm_rating, bull_bear_synthesis, debate_lesson, plus the risk-debate output if run_risk_debate is chained.

Routing

CLI: python main.py picks "<query>" --debate N Internal: StockPickerAgent.pick(debate_rounds=N, reflect=True) Direct: from research.debate import run_bull_bear_debate

Cost Note

Bull + bear use the quick_think tier (Claude Haiku 4.5 default). Research Manager synthesizes via deep_think (Claude Opus 4.7 default). Per the model tier doctrine in brain/MODEL_CONFIG.md, this saves ~30% versus running all calls on the deep-think tier.

Internal References

• [[bull_researcher]] — bull persona prompt
• [[bear_researcher]] — bear persona prompt
• [[research_manager]] — synthesis prompt with 5-tier rating
• [[debate]] — Python orchestrator
• [[researcher-bull]] — agent file
• [[researcher-bear]] — agent file