Cybernetics: A Quick Reference for Architects

Core Terminology Translation

Term	Plain English	Modern Equivalent
Variety	The number of possible states a system can be in	Complexity, degrees of freedom, state space cardinality
Requisite Variety	Having enough capacity to handle the complexity you face	Complexity parity, matching bandwidth
Attenuator	Something that reduces complexity	Filter, constraint, bounded context, API contract, schema validation
Amplifier	Something that increases capacity to handle complexity	Delegation, parallelism, event streaming, distributed processing
Relaxation Time	How long a system takes to stabilise after disruption	Recovery time, time to consistency, mean time to recovery
Perturbation	An external disruption or shock	Incident, load spike, upstream failure, market shift
Homeostasis	Self-regulation to maintain stability	Steady state, auto-scaling, self-healing, dynamic equilibrium
Level of Recursion	A nested layer of the system	Bounded context, domain, team topology, system tier
Model	A simplified representation used for control	Domain model, schema, API spec, mental model
Autopoiesis	A system that exists to produce itself	Self-perpetuating bureaucracy, governance theatre
POSIWID	“The Purpose Of a System Is What It Does”	Judge by outcomes, not intentions

The Four Laws (Simplified)

1. Complexity Must Be Matched

“Only variety can absorb variety” - Ashby’s Law

A controller needs at least as much capacity as the system it controls. Mismatched complexity = instability.

Test: Can your governance framework handle the actual variety of situations it encounters? Or does reality leak around the edges?

2. Recovery Speed Beats Robustness

Relaxation time must be shorter than perturbation frequency

If disruptions arrive faster than recovery, the system never stabilises. Chaos accumulates.

Test: Is your architecture backlog growing? Is technical debt accumulating? That’s the symptom of relaxation time exceeding perturbation rate.

3. No Model, No Regulation

Every regulator contains a model of what it regulates

You cannot control what you cannot represent. But the model is supposed to be simpler than reality. The question is whether you’re simplifying the right things.

Test: Can you articulate the model your governance is based on? Is it written down? Is it testable?

4. Two Options Only

Attenuate system variety OR amplify regulator variety

When complexity is mismatched, you either reduce what you’re trying to control or increase your capacity to control it. Most organisations do both badly, installing amplifiers where they need attenuators and vice versa.

Test: Where have you added technology that generates more complexity rather than absorbing it?

The Minimum Data Principle

Beer’s most radical insight: you need less data, not more.

If your model is good, most data just confirms what you already know. Only exceptions are worth processing. Everything else can be discarded at source.

Conventional Wisdom	Different Perspective
Collect everything, analyse later	Filter at source, process exceptions
Data lakes as strategic assets	Models as strategic assets
More data = better decisions	Right data = better decisions
Store then query	Alert on deviation

Implication for AI: Use AI to identify exceptions, not to process everything. The goal is smarter filtering, not bigger pipelines (not to mention throwing yet more money and resources to tech companies!).

Nested Autonomy

Systems contain subsystems contain subsystems. Each level is autonomous within constraints from the larger system that contains it.

The Full Picture (Usually Ignored)

Most business thinking treats the economy as the outermost layer. This is the fundamental error as it sits within a series of complex systems:

- Individual
-- Team
--- Department
---- Organisation
----- Sector
------ Economy (markets, production, consumption)
------- Human Society (civilisation, culture, institutions)
-------- Natural Ecosystems (forests, oceans, watersheds, atmosphere)
--------- Biosphere     ← the foundation everything rests on

Constraints come from the foundation. Each level rests on the ones beneath it (Beer would call them higher levels of recursion). The ultimate constraints are physics, chemistry, and ecology, not policy. See the interactive version.

At civilisational scale, we’re violating Ashby’s Law in both directions:

Attenuating variety we shouldn’t: destroying biodiversity/complexity that makes natural systems resilient
Generating variety nature can’t absorb: exceeding planetary boundaries (carbon, nitrogen, waste)

Climate relaxation time: decades to centuries. Perturbation rate: years. System cannot stabilise.

Key insight: Central coordination should specify patterns, not content. But “central” at the foundation means planetary boundaries.

Wrong Approach	Right Approach
“Use these specific technologies”	“Meet these interface contracts”
“This is the salary for that role”	“This is the salary budget for your team”
“Cut all inventories by 10%”	“Stay within this working capital limit”
“Environment is an externality”	“Ecological boundaries are the top-level constraint”

Pattern specification enables autonomy. Content specification destroys it. Ignoring the levels you rest on doesn’t make you autonomous; it makes you unviable.

The Bureaucracy Test

Beer distinguishes between:

Healthy nucleus: Coordinates operations for benefit of the whole
Bureaucracy: Coordinates operations for its own continuation

Symptoms of autopoiesis (self-production):

Accepts change but produces no actual alteration
Grows while value delivered shrinks
Creates work that justifies its own existence
Optimises for auditability over outcomes

Freedom Through Constraint

The paradox: freedom requires structure.

Unconstrained variety cannot stabilise, cannot learn, cannot adapt. Real freedom comes from well-designed constraints that filter complexity appropriately.

You’re free to make decisions within your bounded context because the context is bounded.

False Dichotomy	Beer’s Resolution
Freedom vs Control	Freedom through well-designed control
Centralisation vs Decentralisation	Right constraints at right level
Governance vs Autonomy	Governance enables autonomy

Application to AI Governance

AI as Variety Handler

The question isn’t “should we use AI?” It’s “which side of the equation?”

Wrong Side (Amplifying System Variety)	Right Side (Attenuating System Variety, Amplifying the Regulator)
Generating more content	Filtering to exceptions
Creating more options	Identifying what matters
Automating everything	Automating pattern recognition
Agents generating agents	Agents surfacing anomalies

The Confirmation Fatigue Problem

When AI-generated content looks plausible even when wrong, human review becomes theatre. Beer would say: you’ve lost your regulatory model. The human can no longer distinguish signal from noise.

Solution: Don’t review everything. Review exceptions. Use AI to identify what deviates from expected patterns, then apply human judgement there.

The Capability Degradation Problem

If AI does everything, humans never develop expertise. Beer would say: you’ve attenuated human variety instead of amplifying it.

Solution: Use AI to handle complexity that exceeds human capacity, not to replace human capacity. Augment, don’t substitute.

Quick Diagnostic Questions

For any system or governance framework:

Complexity match: Does your controller have enough capacity for the variety it faces?
Recovery speed: Can you stabilise faster than disruptions arrive?
Model clarity: Can you articulate the model your regulation is based on?
Amplifier placement: Is your technology generating complexity or absorbing it?
Autonomy enablement: Are you specifying patterns or content?
Bureaucracy check: Is your governance serving delivery or perpetuating itself?
Variety direction: Are you amplifying human capacity or attenuating it?