Menü
Adversarial red-team debate protocol — when to use it, the 4-step structure, role-pair recipes, Referee scoring, termination conditions, MAS failure mitigations.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Basierend auf der SOC-Berufsklassifikation
| 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"] |
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).
| 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.
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.
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).
Each side queries the other:
Single round of clarification per question. No infinite back-and-forth (this is the step-repetition MAS failure mode the research doc cites).
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]
| 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 |
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).
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 |
Saved per the orchestrator invoking the protocol (typically output/finance/, output/mna/, output/crisis/, output/board/).
AI governance — EU AI Act Article 9, NIST AI RMF, ISO/IEC 42001 — risk classification, lifecycle gates, model cards, HITL policy, AI incident response.
Crisis response toolkit — classification, trigger thresholds, escalation tiers, war-room activation, rapid liquidity stress, supply-chain alternate-source, holding-statement, AAR.
Enterprise risk management — COSO ERM 2017 + ISO 31000 — risk taxonomy, appetite statement, top-risk register, KRIs, Bow-Tie, three-lines-of-defense.
Enterprise executive decision frameworks, memo templates, board protocol, RACI, OKR, escalation rules, dissenting-opinion format.
Deterministic financial tools — WACC, NPV, IRR, payback, EVA, real-options, Monte Carlo, sensitivity, covenant/liquidity checks. The Financial Hardcoding Directive.
M&A playbook — deal thesis, screening, financial/commercial/legal diligence, valuation methods, real-options on deal structure, integration playbook, post-deal KPIs.