If the tool is a narrative engine, treating it like a calculator is a category error. You don't get mad at a car for not flying; you learn to drive.
Each generation of computing hardware forced us to adapt how we think about expressing and solving problems. Each generation also required a new engineering discipline.
Numeric computing: We got fast arithmetic machines and developed numerical analysis.
Symbolic computing: We got bitstring crunchers and developed software engineering.
Narrative computing: Now we have storytelling engines — and we need narrative engineering.
Narrative computing is what the machine does. An LLM takes a prompt and generates a narrative. This is the primitive operation — the transistor of narrative engineering.
Narrative engineering is how we compose those primitives into reliable systems. A single LLM call is locally coherent but unreliable, just as a single transistor amplifies but drifts. You don't fix that by building a better transistor. You design circuits: redundancy (multiple perspectives), feedback (evaluation against rubrics), iteration (generate-evaluate-revise loops), and staged composition (charter → deliberation → resolution → evaluation). The engineering makes the system trustworthy even when individual components aren't.
Software engineering grew from symbolic computing. Narrative engineering grows from narrative computing.
From Practice to Theory tells the story of how these practices grew, and the theory used to guide their development.
The essay Stories All the Way Down explains how everything is a story, how stories are a way to deal with wicked problems, and how we are constantly editing and re-editing the stories we tell ourselves about the world.
Cyberneutics is a methodology for working with AI systems as collaborative sense-making partners rather than oracles that deliver answers. It provides rigorous, traceable techniques for making decisions under genuine uncertainty — the kind where you can't enumerate outcomes or assign probabilities, and the act of deciding changes what you're deciding about.
It builds upon the runtime environment of MOOLLM (Don Hopkins) and aligns with the philosophy of Narrative Engineering (Alex Boland). Think of it as: MOOLLM is the platform, Cyberneutics is the practice.
"We've never had narrative engines powerful enough to make this methodology necessary before."
Large Language Models are not databases. They are not logic engines. They are narrative generators - storytelling machines operating through what we call "the pachinko of stored literature." This changes everything about how we should work with them.
- Practitioners working with AI on complex sociotechnical problems
- Researchers exploring human-AI collaboration
- Anyone frustrated with shallow AI outputs who suspects there's a better way
- Skeptics wondering if this is just prompt engineering theater → start here
- Theorists interested in narrative engineering, cybernetics, or sense-making
New to these ideas? → Start Here — 15-minute onboarding path, then your first committee run.
Or step by step:
- Read: Why Narrative Engines Change Everything
- Try: Adversarial Committees (a concrete technique you can use today)
- Run the demo (optional): Interactive demo — naive vs committee pipeline in the browser on your machine (Node 20+)
Want the theory?
- Read the Essays in order, starting with From Practice to Theory.
- For the formal pipeline algebra, see Decorated Texts.
- For curated reading paths by audience (Practitioner, Theorist, Skeptic), see the Essays directory.
Want practical techniques?
- Start with the Quick Start Guide — your first committee deliberation in 30 minutes.
- Then explore the full Artifacts directory.
Most AI interaction fails because we treat LLMs like search engines or calculators. We ask questions expecting definitive answers. But LLMs are most powerful when understood as rapid story generators that help us navigate complexity through narrative exploration.
The core pipeline: explore possible futures (scenario generation), evaluate them adversarially (committee deliberation), repeat to map the decision landscape, and trace every commitment to the arguments that produced it. Quality propagates predictably through the pipeline — you always know where the ceiling is and what to improve. See Decisions Under Uncertainty for the full argument.
The methodology documented here has produced:
- Rigorous, traceable decision-making under genuine uncertainty
- Better identification of blind spots and load-bearing assumptions
- Richer exploration of decision spaces with formal quality tracking
- Auditable artifacts with full provenance (not just chat transcripts)
These aren't incremental improvements. They represent a fundamentally different approach to human-AI collaboration.
Essays - Theoretical foundations and synthesis
- Why narrative engines change everything
- From practice to theory: how we got here
- Introduction to Sense-Making Methodology
- The Stochastic Imps of Happenstance
- Cybernetics and the observer problem
- The synthesis: why sense-making is inherently cybernetic
- Societies of Thought: From Neural Evidence to Methodological Action
- Decisions Under Uncertainty
Artifacts - Practical techniques and protocols
- Adversarial committees with fixed character rosters
- Robert's Rules as forcing functions
- Independent evaluation protocols
- Hiring Decision Example
- Worked examples and transcripts
demo/ - Interactive local app (Next.js): committee vs naive comparison; see demo/README.md for quick start, architecture, and how local vs. API modes work
Palgebra - Formal algebra for LLM pipelines
- Reference Card — start here: syntax, operators, morphism types, propagation rules, composition laws
- Decorated Texts — full essay developing the formalism from first principles (soft types, enrichment vs. transformation, confidence propagation, human gates as collapse operators)
- Committee as Palgebra — worked example: the adversarial committee pipeline formalized as resource equations
- Duality and Composition — the fan/funnel duality: scenario generation as the committee's categorical dual, their composition as a decision monad, and iteration for mapping decision landscapes
- Adapts Fong and Spivak's resource-theoretic framework (Seven Sketches in Compositionality, Ch. 2) to LLM pipelines. Three isomorphic representations: resource equations, string diagrams, and YAML-decorated artifact files
The essays describe why narrative engines need narrative engineering. The artifacts provide how — concrete techniques like adversarial committees and evaluation rubrics. Palgebra provides what, precisely — a formal language for specifying pipelines, their types, their quality propagation, and their composition laws. An adversarial committee is a transformation morphism; a rubric evaluation is an enrichment morphism; a human review gate is a collapse operator. The formalism makes these relationships explicit and composable.
Applications - Domain analyses applying the framework to real-world phenomena
- Narrative Immune Systems — information warfare, journalism, and the trust commons, analyzed through the immune analogy
References - Background reading
- Dervin's Sense-Making Methodology
- Second-Order Cybernetics (von Foerster, Bateson)
- Deleuzian philosophy (becoming, difference, repetition)
- Narrative Engineering (Alex Boland)
Agent Skills — Slash commands available when working with an AI agent on this repo
| Command | What it does | When to use |
|---|---|---|
/committee [topic] |
Runs an adversarial committee deliberation using the roster in agent/roster.md |
Complex decisions, competing values, "what are we missing?" problems |
/review |
Independent evaluation of a committee transcript against five rubrics | After any /committee run, or on a pasted transcript — completes the feedback loop |
/string-diagram |
Converts resource equations to Mermaid diagrams | Visualizing pipelines, formalizing workflows, editing equation sets |
/handoff |
Generates a session handoff for successor agents | End of work sessions, before breaks, after major milestones |
These skills are the methodology made executable: /committee operationalizes the adversarial committee technique, /review operationalizes independent evaluation, /string-diagram operationalizes the palgebra formalism, and /handoff maintains continuity across agent sessions.
-
LLMs are storytelling machines. Everything they produce - mathematical proofs, legal analysis, code, decision trees - are narrative constructs, not mechanistic solutions.
-
Observation changes state. Every AI response is a control signal that modifies your cognitive state. You cannot ask a question without being changed by the answer.
-
Gaps produce bridges that change situations. The act of articulating a problem transforms the problem. Sense-making isn't discovery - it's production.
-
Repetition produces difference. Asking "the same" question multiple times isn't failure - it's exploration of latent space, mapping the territory of possible interpretations.
-
The analyst is an editor. Your role isn't truth-seeker but curator of which stories get "published to reality."
Cyberneutics is a neologism derived from
Cyber (κυβερνήτης, the steersman) - governance, feedback, control Neutics - evoking both hermeneutics (interpretation of texts) and neural (the substrate you're actually working with)
It captures something that "narrative computing" doesn't quite get at: that you're not just computing on narratives but doing something closer to steering through interpretation - the cybernetic loop applied to meaning-making with neural systems. There's also a nice echo of heuristics in there, which fits the System 1 / fast-pass dimension of the work.
This is early-stage documentation of an emerging methodology. The techniques described here have been refined through iterative practice and have reached stable behavioral equilibrium. The theoretical framework is being formalized.
Run and test: For how to run the methodology (skills in chat, string-diagram script), the local interactive demo in demo/, and how to test the repo (smoke test, structure checks, CI), see the repository review and run guide. For recent session context and maintainer hand-off, see the latest handoff in agent/.
Evidence base: The adversarial committee technique has empirical support from research on multi-agent reasoning — see Societies of Thought for a synthesis of findings from Google, UChicago, and the Santa Fe Institute showing that perspective-switching and conversational scaffolding improve reasoning quality. The theoretical foundations draw on established work in sense-making methodology (Dervin), second-order cybernetics (von Foerster, Bateson), and process philosophy (Deleuze). What remains to be validated is the specific combination of techniques and their calibration across problem domains.
Feedback and questions are welcome—especially if you’re running the demo or applying the methodology in your own work. This repo is shared so people can use and read the material; it is not oriented around an open-contribution workflow.
CC BY-SA 4.0 for essays, MIT for code artifacts
"In complex systems, mechanism fails to comprehend the full picture, but stories can capture enough relevant structure to be useful."