Why-How-Who Evaluation Framework

A framework for encoding, comparing, and evaluating facilitation methodologies, developed through research at Cooperative AI (Joseph Low) in collaboration with OFL.

Overview

Traditional facilitation evaluation focuses on outcomes: Did the group reach agreement? Were participants satisfied? But this misses crucial process dimensions that determine facilitation quality.

The Why-How-Who framework provides a structured way to:

  1. Encode facilitation methodologies into comparable formats
  2. Measure how close a conversation is to known facilitation styles
  3. Generate feedback signals for training AI facilitators

The Three Dimensions

Why (Purpose & Outcomes)

What is the facilitation trying to achieve?

Outcome TypeDescriptionExample Methods
Agreement BuildingHelp group reach consensusDelphi, Consensus Workshop
Preference ElicitationSurface individual preferencesPolling, Harmonica
Error SurfacingIdentify gaps in thinkingDevil’s Advocate, Red Team
Perspective TakingExpose to other viewpointsCross-pollination, Fishbowl
SynthesisAggregate into actionable outputAffinity Mapping
IdeationGenerate new ideasBrainstorming, Six Hats
Conflict ResolutionResolve disagreementsNVC, Mediation

How (Process & Techniques)

What does the facilitator actually do?

Intervention Styles:

  • Non-directive: Facilitator only manages logistics
  • Semi-directive: Facilitator asks questions, summarizes
  • Directive: Facilitator guides thinking, challenges assumptions

Question Types:

  • Open: “What do you think about X?”
  • Closed: “Do you agree with Y?”
  • Probing: “Can you say more about that?”
  • Clarifying: “Did you mean Z?”
  • Challenging: “What if the opposite were true?”

Timing:

  • Scheduled: Interventions at predetermined points
  • Responsive: React to participant inputs
  • Threshold-based: Intervene when certain conditions met

Who (Participants & Dynamics)

Who is involved and how do they interact?

Interaction Modes:

  • One-to-one (interviewer and participant)
  • Small group (3-12 people)
  • Large group (12+ people)
  • Plenary (whole assembly)

Power Dynamics:

  • Hierarchy sensitivity
  • Anonymity support
  • Minority voice protection

Conversation Signatures

A key insight from this framework: we can compute “signatures” of conversations based on Why-How-Who dimensions, then compare them.

Computing a Signature

  1. Label dialogue acts with Why-How-Who tags
  2. Count frequencies of each tag type
  3. Create vector representing conversation characteristics

Comparing Conversations

Instead of asking “Is this good facilitation?” (subjective), ask:

  • “How similar is this to restorative justice facilitation?”
  • “How similar is this to Socratic dialogue?”

This relative comparison is more tractable and produces measurable feedback signals.

Applications

For Pattern Development

  • Encode patterns using consistent dimensions
  • Compare patterns across methodologies
  • Identify gaps in pattern library

For AI Training

  • Generate labeled datasets from conversations
  • Create feedback signals for reinforcement learning
  • Measure facilitator agent performance

For Evaluation

  • Process-based metrics (not just outcomes)
  • Comparison to reference methodologies
  • Automated quality assessment

Evaluation Metrics

Process Metrics

  • Intervention frequency and timing
  • Question type distribution
  • Speaking time balance
  • Topic coverage

Outcome Metrics

  • Participant satisfaction
  • Agreement level achieved
  • Idea quantity/quality
  • Action item completion

Signature Metrics

  • Distance to target methodology
  • Consistency within session
  • Appropriate adaptation to context

Implementation

Manual Evaluation

  1. Review conversation transcript
  2. Tag each facilitator turn with Why-How-Who labels
  3. Compute signature statistics
  4. Compare to reference patterns

Automated Evaluation

  1. Use LLM to tag dialogue acts
  2. Compute embeddings for semantic analysis
  3. Compare to known methodology signatures
  4. Generate evaluation report

References

  • Joseph Low, Cooperative AI Fellowship research (2026)
  • OFL Substack publications
  • Conversation Networks paper (referenced in AI Facilitation Sync meetings)