Open standards for AI-assisted facilitation and deliberative democracy: portable method specs, plus the research and reference behind them — all open source.

Seminars

The OFL seminar series brings together experts in facilitation, deliberative democracy, and AI-assisted dialogue.

Browse all seminars | Recent: Jorim Theuns, Cecile Green, Andy Paice, Alice Siu, Jigsaw, Martin Carcasson

Research

Key papers and research topics supporting OFL development.

PaperDescription
WHoW FrameworkCross-domain moderation analysis (Chen et al. 2024)
Fora Corpus262 facilitated dialogues from MIT (Schroeder et al. 2024)
D-agree PlatformAutomated facilitation agent at crowd scale (Ito et al. 2022)
Facilitation in the AI EraEthnographic study of 22 expert facilitators (Jigsaw 2025)
Cueing the CrowdLLM conversational cues for brainstorming (Rayan et al. 2025)
LLM Facilitation SurveyComprehensive survey on LLM-based facilitation (Korre et al. 2025)
Generative Social ChoiceLLM-augmented democratic processes (Fish et al. 2025)
AI Moderation ChatbotsAI moderation and chatbot facilitation
ConvoKit DatasetsDatasets for facilitation research

More: LLM vs Human Facilitation | Techniques | Mini-Publics | Storytelling

Browse all research →

Knowledge Base

Core concepts and definitions shared across OFL projects.

Architecture

OFL’s core artifact is the method spec: a portable, forkable definition of a facilitation method that any capable runtime can execute (Harmonica is the reference implementation). Each method spec has two halves:

  • A protocol — the method itself, as a forkable spec: its stages, roles, facilitator prompts, and what carries between them. Fork it, adapt it, run it on any capable runtime.
  • Evals — how you know it ran well: rubrics scored against the conversation, interoperable with weval’s open eval format.

Nine methods are published in the method-specs registry; browse them under protocols.

The rest of the library is the research and reference these specs draw on — abstract patterns (facilitation methodologies described with the Why-How-Who framework) and teardowns (how many real platforms orchestrate AI facilitation agents today).

A standard owned by facilitators

Good facilitation has to be defined by the people who facilitate. OFL is built so the definition they produce, the evaluations and the knowledge, stays a shared resource they own and govern, closer to a data cooperative than to data a platform extracts. The specs also map to the field rather than walling it off: each one cross-maps to Group Works patterns and IAF competencies.

How the standard is owned and governed →

Evaluation

Every method spec is paired with evals: per-stage, per-method rubrics scored against real conversation turns, so “good facilitation” is inspectable and measurable by the method’s own logic rather than assumed. Interoperable with weval’s open format.

Browse evaluation →

Support

Donate on Giveth — Help fund research, development, and community building.

Contributing

  1. Add research summaries
  2. Propose evaluation criteria
  3. Submit pull requests with documentation