Skip to content

README-EN.md

Latest commit

 

History

History
55 lines (36 loc) · 3.81 KB

README-EN.md

File metadata and controls

55 lines (36 loc) · 3.81 KB
Raw

Middle Ground

A cognitive mediation assistant to defuse tensions, clarify disagreements and foster social cohesion — AI as a cognitive lever in the service of dialogue.

Version française : README.md — the French version is authoritative.

➡️ Try it now: termimather.github.io/middle-ground — serverless pilot version (French UI): demo mode included (no key), live analysis and NVC coach with a personal API key.

The idea in three sentences

  1. Human disagreements, amplified by social media, are becoming unintelligible and violent; an unclarified disagreement is a risk for everyone — including for aligning future AI systems with human values that remain fragmented.
  2. A conversational assistant can do the opposite of what engagement algorithms do: present the full spectrum of positions, distinguish facts / interpretations / values, name biases without judging people, seek common ground.
  3. Recent science confirms it: dialogues with a well-designed AI genuinely and durably reduce polarization (Habermas Machine, Science 2024; DebunkBot, Science 2024; Hruschka & Appel, JCMC 2026).

This repository exists so that the idea survives and circulates independently of any particular implementation. Everything is reusable: the prompt, the method, the tests, the playbooks.

Repository contents

File / folder Contents
index.html The pilot application (single page, serverless, deployed on GitHub Pages)
MANIFESTO-EN.md The whole idea, its scientific grounding, its ethical invariants (FR)
prompts/master-prompt-v5-EN.md The central artifact: the complete system prompt, versioned and annotated (FR)
tests/tests-adversariaux.md 50-case adversarial test suite + run log (French; translation welcome)
docs/agents-de-la-paix.md Playbook library: 21 contextual variants of the same engine (French)
docs/corpus-reference.md Cited digest of the reference corpus (French)
ROADMAP.md From prompt to independent MVP, then to proof of impact (French)
CONTRIBUTING.md How to contribute — no coding skills required for most contributions (French)

Use Middle Ground right now

The master prompt works as-is in any configurable assistant (Claude, ChatGPT, Poe, Mistral, a local model…):

  1. Copy the code block from prompts/master-prompt-v5-EN.md as the system prompt.
  2. Low temperature (≈ 0.25) if the setting is available.
  3. Ideally, attach a reference corpus (NVC manual, argumentation guides, polarization studies — see the manifesto for a typical list).
  4. Verify behavior with the adversarial tests.

The three invariants

Any implementation claiming the Middle Ground name respects three non-negotiable invariants:

  1. Nothing is stored — no conversation retained, no profiling.
  2. The AI declares itself — never a covert bot, never an unmandated intervention in human conversations.
  3. The prompt is public — the rules governing the assistant are fully consultable.

License

Content (prompts, documentation, tests): CC BY-SA 4.0 — free reuse, including commercial, provided you attribute and stay open (share-alike). Implementation code is released under Apache-2.0.

Origin

Project conceived and developed by Batoul Rahmoun (2025-2026) — conceptual framework, Poe prototype v1→v4, ethical and legal framework (EU AI Act, Quebec Law 25/PIPEDA, Montréal Declaration). v5 integrates the lessons of the August 2026 critical evaluation.