debate-protocol

name	debate-protocol
description	Adversarial red-team debate protocol — when to use it, the 4-step structure, role-pair recipes, Referee scoring, termination conditions, MAS failure mitigations.
allowed-tools	["Read","Write","Glob","Grep"]

Debate Protocol

Per the research doc (Section "Adversarial Red-Teaming (Debate Protocol)"), debate-style multi-agent topologies mitigate individual-LLM bias and surface hidden assumptions. Used by capital-allocation, mna-cockpit, boardroom (when explicitly invoked), and crisis-warroom (option-ranking step).

When to Use Debate vs. Consensus

Pattern	Use when
Debate	Decision is high-stakes, irreversible, contains strong narrative pull, has structural tension between functions, has well-known cognitive biases (sunk cost, anchoring, optimism)
Hierarchical consensus (boardroom)	Decision is cross-functional but not adversarial; perspectives need to compose, not contend
Solo	Narrow, high-precision, single-domain; debate adds overhead without insight gain

Default: capital allocation, M&A, growth-vs-discipline tradeoffs → DEBATE. Strategy refresh, talent calibration, ESG plan → CONSENSUS.

4-Step Protocol

Step 1 — Specification (Orchestrator)

Define decision frame and shared data bundle:

Decision: [What is being decided]
Shared data: [What both sides have access to]
Time horizon: [N years / quarters]
Reversibility: [How much we can unwind]
Decision authority: [Who has final say after debate]

Both sides agree on the frame before opening. Disagreement on frame is itself a finding — escalate to the orchestrator to resolve.

Step 2 — Opening Briefs

Structured templates — both sides answer the same questions, from opposing positions:

Pro-Action template:

Position: [What we should do]
Strategic thesis: [Why it matters; what changes for the firm]
Base case: [Quantified expected outcome]
Upside drivers: [Specific, evidenced]
Downside acknowledgment: [Honest worst case]
Optionality preserved: [What future moves does this enable?]
Execution plan: [Phased, owners, milestones]
Kill criteria: [What stops us mid-flight]

Discipline-Challenge template:

Where the base case is too optimistic: [3–5 specific assumptions]
What discipline lens shows: [Stressed financials / risk]
Hidden costs: [TCO, integration, working capital, talent]
Guardrail check: [Pass / fail / margin on each]
Opportunity cost: [What else could we do?]
Reversibility cost: [How much to unwind]
Catastrophic-failure conditions: [Reverse stress]

Briefs are written; both sides read the other's before cross-examination (no surprise rhetoric).

Step 3 — Cross-Examination

Each side queries the other:

• Assumption: "Your case assumes X — what's the evidence?"
• Data gap: "Do we have commitments or just TAM?"
• Model risk: "Your WACC assumes current capital structure — does this transaction change it?"
• Execution dependency: "Synergy curve assumes 2 integrations in 12 months — can we?"

Single round of clarification per question. No infinite back-and-forth (this is the step-repetition MAS failure mode the research doc cites).

Step 4 — Adjudication

A Referee (CEO-aligned synthesizer; not a debater) writes:

## Points of Agreement
[Where evidence converges; one side conceded]

## Resolved Tensions
[Tension + the evidence that resolved it + the new shared position]

## Unresolved Tensions
[Tension + what data would resolve it + cost/time to obtain]

## Guardrail Status
[Each guardrail: pass / fail / margin]

## Confidence
High / Medium / Low — with the reason

## Option Set
[Concrete options with explicit conditions — never a binary]

Role-Pair Recipes (common pairings)

Topic	Advocate	Challenger	Referee
Capital allocation (growth)	cmo + cpo	cfo + chief-risk-officer	ceo
Build-vs-buy (technology)	cto	cfo	cpo (user-value lens)
Speed-vs-security	cpo + cto	ciso	ceo
Marketing spend (brand vs perf)	cmo (brand)	cmo (perf) + cfo	ceo
Talent investment vs cost	chro	cfo	ceo
Sustainable sourcing vs cost	chief-sustainability-officer	csco + cfo	ceo
M&A: strategic premium	cso + cmo	cfo + chief-risk-officer	ceo
Pricing power test	cro	cfo + cmo	ceo

Referee Scoring Rubric

Each side scored on:

Criterion	Weight
Evidence quality (data + analogous precedent)	30%
Logical coherence (no contradictions; framework consistent)	20%
Engagement with counter-evidence (acknowledged + addressed)	20%
Quantification (financial + operational)	20%
Reversibility-cost honesty	10%

The Referee scores; the better-supported case is preferred. Ties default to discipline (reversibility).

Termination Conditions (MAS Failure Mitigations)

The research doc names these failure modes; the protocol mitigates each:

Failure mode	Mitigation
Step repetition	Single clarification round; Referee enforces
Derailment	Frame agreed in Step 1; off-frame questions deferred
Missing verification	Both sides answer same questions; Referee gates on completeness
Premature termination	Adjudication explicitly checks each section above
Late termination	Cross-examination time-boxed (default: 30 min / 5 exchanges)
Specification ambiguity	Step 1 is explicit; "Decision: …" sentence required
Organizational breakdown	Role-pair recipe pre-defined; no ad-hoc participation
Inter-agent conflict (irreconcilable)	Explicit "irreconcilable" flag → escalate to CEO with both positions preserved

Output

Saved per the orchestrator invoking the protocol (typically output/finance/, output/mna/, output/crisis/, output/board/).