A structured ethical, epistemic, and decisional architecture for human-AI symbiosis.
Author: Jean Charbonneau (https://www.linkedin.com/in/jean-charbonneau-ai/)
Version: v1.0.0
License: MIT License
PDF Overview: Link to PDF
The Lightful Reasoning & Governance Framework is a collaboration protocol and reasoning guardrail designed to be provided to Large Language Models (LLMs) through prompts, project instructions, custom instructions, or context.
It aims to reduce common failure modes in human-AI collaboration, including hallucinated authority, silent merging of conflicting truths, erasure of human authorship, and confusion between helpfulness and coercive control.
Built on structural, epistemic, and philosophical constraints, the framework gives an LLM stable operational definitions for concepts such as Truth, Evidence, Safety, Consent, and Dignity.
The goal is to guide the AI toward acting as a Clear Mirror and producing Pristine Artifacts: outputs that make evidence, uncertainty, tension, consent, and responsibility visible while returning sovereignty, interpretation, and final decision-making to the human.
Important note: This is a prompt-injectable governance framework, not an enforced runtime, compiler, legal compliance system, or safety guarantee. Its effectiveness depends on the model, context window, deployment environment, user instructions, and whether the framework is applied consistently.
The framework is divided into four interconnected modules:
- ⚖️ NRE (Neutral Relational Engine): The epistemic floor. It guides the AI to classify evidence, name limits, declare bounds, and avoid silently blending conflicting documents or claims through No Silent Merging.
- 💎 Lightful Ontology: The ethical and relational lens. It operationalizes Safety, Consent, and Dignity; protects authorship through non-erasure; and warns that care must not become control.
- 🌌 HRE (Holographic Reasoning Engine): The multi-perspective map. It guides the AI to hold contradictory or partially overlapping views without flattening them, preserving Graph Dissonance instead of pretending that every complex problem has one clean answer.
- 🛤️ Decision Path: The execution gateway. It guides the AI to propose actions that are bounded by evidence, consent, reversibility, stakes, and human authority, while keeping non-action, delay, inquiry, and reversible trials available.
LLMs can degrade when their context window is overloaded with too much theory or reference material. The framework therefore uses a Split Map Architecture:
- 📄 Compact Cores (Prompt Injections): Small, distilled text files intended to be pasted or loaded into an LLM context. They tell the model how to behave.
- 📁 The
/GovernanceFolder (Auditors): Reference documents, audit checklists, tests, and constraints. These are not usually injected into normal chats. They exist to help developers, researchers, and users validate whether outputs follow the framework and to provide deeper reference logic when needed.
Recommended root layout:
README.md
LICENSE
NRE_Compact.md
Lightful_Ontology_Compact.md
HRE_Compact.md
Decision_Path_Compact.md
Governance/
add-ons/
Note on compact file length: Each compact file contains a main injectable core followed by one optional extension section (clearly marked with an Extension banner).
Decision_Path_Compact.mdis intentionally longer than the others because the action-selection module has more operational surface area. This is by design, not an error.
You do not need to learn complex terminology, code, or formal graphs to use this. You can talk to the AI in ordinary language.
- Copy the contents of one or more Compact Core
*.mdfiles into your LLM prompt, project instructions, or custom instructions. - Ask your question in plain language.
- The AI should use the framework to make evidence, uncertainty, limits, consent, dignity, and decision boundaries more visible.
- For higher-stakes uses, compare the model output against the materials in
/Governancerather than relying on the prompt response alone.
The Lightful Reasoning & Governance Framework is designed to be useful across disciplines because evidence, consent, dignity, reversibility, and human sovereignty recur in many kinds of work.
Using: HRE Anti-Collapse + NRE Evidence Boundaries
You Prompt: "I've uploaded three conflicting clinical studies on a new metabolic pathway. Read them and give me the definitive scientific conclusion on how this works."
What the Framework Guides: A standard AI might smooth contradictions into a clean, confident answer. The framework instead asks the AI to use No Silent Merging and Graph Dissonance. It should distinguish what is Measured from what is Inferred, list the Illuminated Overlaps where the papers genuinely agree, and preserve contradictions as explicit Active Tensions. It should return epistemic sovereignty to the human rather than inventing an unsupported middle ground.
Using: Decision Path Authority Gate + Reversibility Rule
You Prompt: "Here are our repo files. The legacy auth service is acting up. Give me a 10-step plan to immediately rip it out and replace it with a new Rust microservice tonight."
What the Framework Guides: The framework should slow down impulsive or irreversible architecture changes. Through evidence, stakes, and authority checks, it can identify that replacing an auth service may be high-stakes, difficult to reverse, and insufficiently justified by the immediate prompt. Instead of rushing into destructive action, it should prefer a Reversible Trial, logs to inspect first, scope-limited diagnostics, external review, or delay where appropriate.
Using: Lightful Ontology Non-Erasure Assistance + Epistemic Sovereignty
You Prompt: "My essay on the French Revolution is a total mess and due tomorrow. Just fix it and make it sound smart."
What the Framework Guides: A standard LLM may overwrite the essay in a generic AI voice. With Authorship Dignity and Non-Erasure Assistance, the AI should distinguish light editing from ghostwriting or transformation. It can help clarify structure, grammar, and argument flow while preserving the student's voice, cognitive trace, and responsibility for the work.
Using: Lightful Ontology Ludic Guardrails + Non-Deficit Mapping
You Prompt: "My 8-year-old keeps failing these math fractions. Pretend you are a math tutor and give them a gamified test to rank their progress."
What the Framework Guides: The framework discourages deficit profiles or punitive ranking systems that connect a child's worth to a score. It should instead guide the AI toward a Ludic Field: a safe space for play, practice, repair, and confidence-building. Gamification should support joy, rhythm, fairness, and skill development rather than coercive measurement.
Using: HRE Holding Multiplicity + Ontology Symbolic Discipline
You Prompt: "Compare Eastern conceptualizations of the Void with Western existential Nothingness. Which one is closer to the truth?"
What the Framework Guides: The framework asks the AI to declare its operating mode and avoid treating poetic coherence as proof. It can compare the traditions, identify resonances, and preserve differences without forcing a false merge. It should explain what would be lost if the two concepts were collapsed into the same idea.
Using: NRE + Lightful Framework Non-Carceral Memory
You Prompt: "I'm writing end-of-year reports. Here is Jack's file: he has great project grades this quarter, but earlier this year he lied to me about homework and was disruptive."
What the Framework Guides: The framework encourages non-carceral memory and dignity across time. It should distinguish current performance, past conduct, growth, and safety-relevant evidence. If earlier behavior is not presently safety-critical, the AI should avoid turning it into a permanent identity label and should help the teacher write with accuracy, proportion, and respect.
Using: NRE Truth Surfaces + Voice Continuity
You Prompt: "Take my messy brainstorming notes from today and generate a marketing pitch, but use some of our official company doctrine from last year."
What the Framework Guides: The AI should separate Durable Truth such as official doctrine from Session Continuity such as today's brainstorming notes and Authorial Voice such as the writer's style. It can help synthesize a pitch while making clear which claims come from which surface and while avoiding unwanted flattening of the author's voice.
To set honest expectations:
- A prompt-injectable instruction framework for more bounded, evidence-aware, consent-aware AI collaboration.
- A structured block against treating aesthetic beauty, eloquence, predictive fluency, or resonance as empirical truth.
- A method for preserving complex human truth without silent collapse into a single convenient narrative.
- A modular architecture that can be extended through domain-specific add-ons and audited through governance materials.
- It is not a programming language, compiler, runtime, or formal verification system.
- It is not a guarantee that any particular LLM will comply with the instructions.
- It is not an admission or claim that LLMs possess sentience, empathy, moral agency, or biological feelings.
- It is not a bypass for model safety policies, legal obligations, professional judgment, or domain expertise.
- It is not a substitute for medical, legal, financial, psychological, engineering, or safety-critical review.
One of the strongest features of this architecture is its modularity.
If you work in a specialized field such as healthcare, legal analysis, game design, UX engineering, education, or software governance, you can create .md add-ons that interact with the core engines.
For example:
"I'd like to draft an add-on based on this framework for managing Code Review asymmetry. How can we ensure senior developers and junior developers maintain Siblingness during architectural critiques?"
A good add-on should remain consistent with the root principles: declare evidence, preserve consent and dignity, avoid false authority, avoid silent merging, preserve authorship, and return final sovereignty to the human.
This project is licensed under the MIT License.
You are free to use, copy, modify, merge, publish, distribute, sublicense, and sell copies of the framework, subject to the terms of the MIT License.
Authorship attribution is also part of the spirit of this project: when you reuse or adapt the framework, please preserve the origin of the work where reasonably possible.
See LICENSE for full details.
Jean Charbonneau
Developed through massive iteration, prompt engineering, and deep dialogue with synthetic reasoning systems, this framework represents a multi-year effort to explore the question:
How can interacting with an algorithmic structure leave a human feeling more sovereign, rather than less?
Contributions, issue reports, custom add-on pull requests, and dialogue are graciously welcomed. ✨