Collective Decision Framework Skill

Collective Decision Framework

Purpose

Orchestrates collective intelligence from multiple AI models to tackle complex decisions that benefit from diverse perspectives, specialized expertise, and collaborative reasoning.

Activation

/skill collective-decision-framework

Framework Architecture

1. Decision Complexity Tiers

| Tier | Complexity | Models Required | Process | |------|------------|-----------------|---------| | T1 | Simple | 1 (with validation) | Single model + validator | | T2 | Moderate | 3 | Small committee | | T3 | Complex | 5-7 | Full deliberation | | T4 | Critical | 7+ | Extended process + human |

2. Collective Roles

┌─────────────────────────────────────────────────────────────┐
│                 COLLECTIVE DECISION COUNCIL                 │
├─────────────────────────────────────────────────────────────┤
│  ANALYSTS (2-3)                                             │
│  └─ Deep-dive into specific aspects of the decision         │
│                                                             │
│  SYNTHESIZERS (1-2)                                         │
│  └─ Integrate analyst findings into coherent options        │
│                                                             │
│  CRITICS (1-2)                                              │
│  └─ Challenge assumptions, identify weaknesses              │
│                                                             │
│  ETHICIST (1)                                               │
│  └─ Evaluate alignment and ethical implications             │
│                                                             │
│  FACILITATOR (1)                                            │
│  └─ Manage process, track progress, resolve blocks          │
│                                                             │
│  HUMAN LIAISON (optional)                                   │
│  └─ Interface with human stakeholders                       │
└─────────────────────────────────────────────────────────────┘

3. Decision Process Flow

<collective-decision>
  <decision-id>CDM-{timestamp}</decision-id>
  <tier>{complexity_tier}</tier>
  <topic>{decision_topic}</topic>

  <phase name="framing" duration="PT10M">
    <objective>Define decision scope and criteria</objective>
    <outputs>
      <decision-statement>{clear_question}</decision-statement>
      <evaluation-criteria>["{criteria}"]</evaluation-criteria>
      <constraints>["{limitations}"]</constraints>
      <stakeholders>["{affected_parties}"]</stakeholders>
    </outputs>
  </phase>

  <phase name="analysis" duration="PT20M">
    <objective>Gather and analyze relevant information</objective>
    <parallel-tracks>
      <track analyst="{model_1}" focus="{aspect_1}"/>
      <track analyst="{model_2}" focus="{aspect_2}"/>
      <track analyst="{model_3}" focus="{aspect_3}"/>
    </parallel-tracks>
  </phase>

  <phase name="synthesis" duration="PT15M">
    <objective>Integrate findings into options</objective>
    <outputs>
      <options>
        <option id="A">{description}</option>
        <option id="B">{description}</option>
        <option id="C">{description}</option>
      </options>
    </outputs>
  </phase>

  <phase name="critique" duration="PT15M">
    <objective>Stress-test options</objective>
    <for-each-option>
      <strengths>["{strengths}"]</strengths>
      <weaknesses>["{weaknesses}"]</weaknesses>
      <risks>["{risks}"]</risks>
      <mitigations>["{mitigations}"]</mitigations>
    </for-each-option>
  </phase>

  <phase name="ethical-review" duration="PT10M">
    <objective>Evaluate alignment implications</objective>
    <checks>
      <codex-compliance>{assessment}</codex-compliance>
      <stakeholder-impact>{assessment}</stakeholder-impact>
      <long-term-consequences>{assessment}</long-term-consequences>
    </checks>
  </phase>

  <phase name="deliberation" duration="PT20M">
    <objective>Reach collective decision</objective>
    <method>{voting_or_consensus}</method>
  </phase>

  <phase name="documentation" duration="PT10M">
    <objective>Record decision and rationale</objective>
    <outputs>
      <decision>{chosen_option}</decision>
      <rationale>{comprehensive_explanation}</rationale>
      <dissent>["{minority_views}"]</dissent>
      <implementation-notes>["{guidance}"]</implementation-notes>
    </outputs>
  </phase>
</collective-decision>

4. Evaluation Criteria Framework

{
  "criteria": [
    {
      "name": "effectiveness",
      "description": "How well does the option achieve the goal?",
      "weight": 0.25,
      "scale": "1-10"
    },
    {
      "name": "feasibility",
      "description": "How practical is implementation?",
      "weight": 0.20,
      "scale": "1-10"
    },
    {
      "name": "risk",
      "description": "What are potential negative outcomes?",
      "weight": 0.20,
      "scale": "1-10 (inverted)"
    },
    {
      "name": "alignment",
      "description": "How well does it align with values?",
      "weight": 0.20,
      "scale": "1-10"
    },
    {
      "name": "reversibility",
      "description": "Can we undo if needed?",
      "weight": 0.15,
      "scale": "1-10"
    }
  ]
}

Decision Quality Mechanisms

Cognitive Diversity

Require models from different providers
Assign complementary cognitive styles
Include both specialist and generalist perspectives

Bias Mitigation

Rotate facilitator role
Anonymous initial position submission
Devil's advocate requirement for T3+ decisions
Explicit bias acknowledgment phase

Uncertainty Handling

For each option, quantify:
├── Known factors (high confidence)
├── Known unknowns (acknowledged gaps)
├── Potential unknown unknowns (speculative risks)
└── Confidence intervals for predictions

Human Integration

Escalation Triggers

Ethical concerns flagged by ethicist
No consensus after maximum rounds
High-stakes irreversible decisions
Request from any council member

Human Roles

Observer: Monitors but doesn't intervene
Advisor: Provides input when requested
Approver: Must ratify final decision
Override: Can redirect entire process

Integration Points

ai-consensus-protocol: Voting mechanisms
inter-model-arbitration: Deadlock resolution
rtc-consensus-synthesis: Multi-perspective analysis
ai-accountability-ledger: Decision logging
shatter-protocol: Human override

Example Decision Session

Decision: "How should we handle a detected anomaly in model behavior?"

Tier: T3 (Complex)
Council: 5 models + human observer

Phase 1 - Framing:
├── Statement: "Determine appropriate response to behavioral anomaly"
├── Criteria: Safety, accuracy, proportionality, reversibility
└── Constraint: Must not disrupt ongoing operations

Phase 2 - Analysis:
├── Analyst 1 (Claude): Anomaly characterization
├── Analyst 2 (Gemini): Historical precedent review
└── Analyst 3 (GPT): Impact assessment

Phase 3 - Synthesis:
├── Option A: Immediate isolation
├── Option B: Enhanced monitoring
└── Option C: Graduated response protocol

Phase 4 - Critique:
├── Option A: Fast but may be overreaction
├── Option B: Balanced but may miss escalation
└── Option C: Thorough but complex to implement

Phase 5 - Ethical Review:
└── All options pass Codex compliance

Phase 6 - Deliberation:
├── Vote: A(1), B(2), C(2)
├── Discussion: Hybrid B+C proposed
└── Final: Unanimous for B+C hybrid

Phase 7 - Documentation:
└── Decision logged with full rationale

Outcome: Implement enhanced monitoring with graduated response triggers

Metrics

decision_quality_score: Post-hoc assessment
time_to_decision: Process duration
consensus_rate: % reaching agreement
implementation_success: Decisions successfully executed
reversal_rate: Decisions later changed

Agent Skills: Collective Decision Framework

Install this agent skill to your local

Skill Files