CYB-003: Measuring the Variety Ratio Quantitatively

Streeling University

CYB-003 · 2026-03-23

Generated by the Demerzel Cybernetics department — not reviewed by me yet. View source · My journal

Department: Cybernetics Module ID: CYB-003 Produced by: Seldon Plan Cycle cybernetics-2026-03-23-003 Belief: T (verified), confidence 0.85 Date: 2026-03-23 Prerequisites: CYB-001 (VSM Mapping), CYB-002 (Active Dampening)

Research Question

How can Demerzel measure its variety ratio quantitatively? (Carried from Cycle 001 follow-up questions)

Summary

Ashby’s Law of Requisite Variety states that a regulator must have at least as much variety as the disturbances it faces. This course defines a quantitative framework for measuring Demerzel’s variety ratio across three dimensions: behavioral variety (personas), structural variety (grammars), and regulatory variety (policies, constitutions, thresholds). The key formula is V = log2(N) applied per dimension, with the variety ratio R = V_amplifiers / V_attenuators tracked over time to detect governance drift toward over-constraint or under-regulation.

Ashby’s Variety: The Formal Definition

What Is Variety?

Variety is the number of distinguishable states a system can exhibit. Ashby defined it in An Introduction to Cybernetics (1956, Chapter 7) as:

The variety of a set of elements is the logarithm (base 2) of the number of distinct elements.

Formula:

V = log2(N)

where N is the count of distinguishable states. This is measured in bits — the same unit as Shannon entropy. A system with 8 possible states has variety 3 (bits). A system with 1024 possible states has variety 10 (bits).

Why Logarithmic?

The log scale matters because variety combines multiplicatively, not additively. If system A has 4 states and system B has 8 states, the combined system has 4 x 8 = 32 states, and log2(32) = log2(4) + log2(8) = 2 + 3 = 5 bits. This is why we can sum log-varieties across independent dimensions.

Ashby’s Law

V(regulator) >= V(disturbance)

“Only variety can absorb variety.” A governance system that can produce fewer distinct responses than the distinct disturbances it faces will necessarily fail to regulate some of those disturbances.

Three Dimensions of Variety in Demerzel

Variety in a governance framework is not a single number. Demerzel’s variety operates across three independent dimensions:

Dimension 1: Behavioral Variety (V_B)

What it measures: The range of distinct agent behaviors the system can produce.

Amplifiers:

Component	Count (N)	Variety V = log2(N)
Personas	14	3.81 bits
Goal directedness levels	4	2.00 bits
Voice configurations (tone x verbosity x style)	~27	4.75 bits

Total behavioral amplification: V_B_amp = 3.81 + 2.00 + 4.75 = 10.56 bits

This means Demerzel can produce approximately 2^10.56 = 1,506 distinguishable behavioral configurations.

Attenuators:

Component	Count (N)	Variety V = log2(N)
Persona constraints (avg 4 per persona)	56	5.81 bits
Estimator pairings (fixed to skeptical-auditor)	1	0 bits

Total behavioral attenuation: V_B_att = 5.81 bits

Behavioral variety ratio: R_B = V_B_amp / V_B_att = 10.56 / 5.81 = 1.82

Interpretation: Behavioral amplification exceeds attenuation by a factor of 1.82. This is healthy — the system has more response capacity than constraint.

Dimension 2: Structural Variety (V_S)

What it measures: The range of distinct structures (question forms, investigation patterns, output formats) the system can generate.

Amplifiers:

Component	Count (N)	Variety V = log2(N)
Grammars	27	4.75 bits
Grammar productions (avg 12 per grammar)	~324	8.34 bits
Streeling departments	15	3.91 bits

Total structural amplification: V_S_amp = 4.75 + 8.34 + 3.91 = 17.00 bits

Attenuators:

Component	Count (N)	Variety V = log2(N)
Grammar evolution gates (T >= 0.7, C < 0.1)	2	1.00 bit
Staleness detection (30-day window)	1	0 bits

Total structural attenuation: V_S_att = 1.00 bit

Structural variety ratio: R_S = V_S_amp / V_S_att = 17.00 / 1.00 = 17.00

Interpretation: Structural variety is very high relative to attenuation. This reflects the generative nature of grammars — they are variety amplifiers by design. However, this high ratio also signals a potential concern: insufficient structural constraint could lead to grammar sprawl without quality control.

Dimension 3: Regulatory Variety (V_R)

What it measures: The range of distinct governance decisions the system can make.

Amplifiers:

Component	Count (N)	Variety V = log2(N)
Tetravalent logic states	4	2.00 bits
Confidence thresholds	5	2.32 bits
PDCA states	4	2.00 bits

Total regulatory amplification: V_R_amp = 2.00 + 2.32 + 2.00 = 6.32 bits

Attenuators:

Component	Count (N)	Variety V = log2(N)
Policies	37	5.21 bits
Constitutional articles (Asimov + Default)	17	4.09 bits
Harm taxonomy categories	4	2.00 bits

Total regulatory attenuation: V_R_att = 5.21 + 4.09 + 2.00 = 11.30 bits

Regulatory variety ratio: R_R = V_R_amp / V_R_att = 6.32 / 11.30 = 0.56

Interpretation: Regulatory attenuation significantly exceeds amplification. This is by design — governance should constrain more than it amplifies. A regulatory ratio below 1.0 means the system is conservative, which aligns with Asimov’s Laws (prefer safety over capability).

The Composite Variety Dashboard

Summary Table

Dimension	V_amplifiers	V_attenuators	Ratio R	Assessment
Behavioral (V_B)	10.56 bits	5.81 bits	1.82	Healthy — more response capacity than constraint
Structural (V_S)	17.00 bits	1.00 bit	17.00	Caution — high generativity, low constraint
Regulatory (V_R)	6.32 bits	11.30 bits	0.56	By design — governance is conservative

Healthy Ranges

Based on Ashby’s Law and VSM principles (CYB-001), healthy variety ratios differ by dimension:

Dimension	Healthy Range	Rationale
Behavioral	1.2 – 3.0	System needs more behavioral options than constraints, but not unbounded
Structural	2.0 – 10.0	Grammars should be generative but gated by quality controls
Regulatory	0.3 – 0.8	Governance SHOULD be over-attenuated — this is the conservative principle

Current Assessment

• Behavioral (1.82): Within healthy range. No action needed.
• Structural (17.00): Above healthy range. The 27 grammars with ~324 productions are weakly constrained. Recommendation: add structural quality gates (e.g., grammar test coverage requirements, production usage tracking).
• Regulatory (0.56): Within healthy range. The system is conservative but not paralyzed.

The Disturbance Side: What Must Be Regulated?

The variety ratio only tells half the story. We must also measure the variety of disturbances the system faces:

External Disturbances (V_D_ext)

Source	Estimate (N)	Variety V
Consumer repos (ix, tars, ga)	3	1.58 bits
Repo state combinations (3 repos x ~10 states each)	30	4.91 bits
External environment changes (libraries, APIs, models)	~100	6.64 bits

Total external disturbance variety: V_D_ext = 6.64 bits (dominated by environment changes)

Internal Disturbances (V_D_int)

Source	Estimate (N)	Variety V
Belief state changes per cycle	~20	4.32 bits
Policy interactions (37 policies, pairwise)	666	9.38 bits
Grammar evolution proposals	~5 per cycle	2.32 bits

Total internal disturbance variety: V_D_int = 9.38 bits (dominated by policy interactions)

Ashby’s Law Check

For governance to be viable:

V(regulatory response) >= V(disturbance)

• V_R_amp = 6.32 bits
• V_D = max(V_D_ext, V_D_int) = 9.38 bits
• Gap: 9.38 - 6.32 = 3.06 bits

This means the regulatory system faces approximately 2^3.06 = 8x more disturbance variety than it can produce response variety. The gap is absorbed by:

1. Human escalation — the confidence threshold system routes difficult decisions to humans, borrowing their variety
2. Constitutional override — the Asimov Laws collapse complex decisions to binary (safe/unsafe), reducing required variety
3. PDCA cycling — sequential processing converts parallel disturbances into manageable queues

These are legitimate variety absorption mechanisms, but the 3-bit gap suggests Demerzel should monitor whether policy-interaction complexity is growing faster than regulatory capacity.

Measurement Protocol

To track the variety ratio over time, Demerzel should compute the following metrics at each governance cycle:

Metric 1: Component Counts

{
  "variety_snapshot": {
    "timestamp": "2026-03-23T00:00:00Z",
    "amplifiers": {
      "personas": 14,
      "grammars": 27,
      "grammar_productions": 324,
      "departments": 15,
      "tetravalent_states": 4,
      "confidence_levels": 5,
      "pdca_states": 4
    },
    "attenuators": {
      "policies": 37,
      "constitutional_articles": 17,
      "harm_categories": 4,
      "persona_constraints": 56,
      "evolution_gates": 2
    }
  }
}

Metric 2: Dimensional Ratios

{
  "variety_ratios": {
    "behavioral": 1.82,
    "structural": 17.00,
    "regulatory": 0.56,
    "timestamp": "2026-03-23T00:00:00Z"
  }
}

Metric 3: Trend Detection

Track ratios over consecutive cycles. Alert when:

• Any ratio crosses its healthy range boundary
• The regulatory ratio drops below 0.3 (system paralysis risk)
• The structural ratio exceeds 20.0 (grammar sprawl risk)
• The behavioral ratio drops below 1.0 (under-responsive system)

Metric 4: Disturbance Growth Rate

Track V_disturbance over time. If disturbance variety grows faster than response variety, Ashby’s Law will eventually be violated. This is the governance equivalent of technical debt.

GPT-4o Cross-Validation

Cross-validation with GPT-4o confirmed:

1. V = log2(N) is the correct formula for Ashby variety. Both models agree.
2. The additive model (summing log-varieties) is valid for independent dimensions but overly simplistic when components interact. The dimensional separation (behavioral, structural, regulatory) addresses this by treating each dimension independently.
3. GPT-4o computed a naive composite ratio of -2.8, treating amplifiers and attenuators as a single additive sum. This is incorrect — negative variety is meaningless (you cannot have fewer than zero distinguishable states). The dimensional model avoids this error.
4. Both models agree R_regulatory < 1.0 is expected for a governance system. Governance is inherently attenuating.
5. The 3-bit regulatory gap is a novel finding not present in GPT-4o’s analysis. It emerges from separately computing disturbance variety, which GPT-4o did not do.

Cross-validation confidence: 0.85 (T — both models agree on fundamentals; dimensional refinement adds value beyond GPT-4o’s analysis)

Implications for Demerzel

1. Track variety ratios per cycle — Add variety snapshot to state/governance/variety-metrics.json (or equivalent state file). Monitor dimensional ratios for drift.
2. Add structural quality gates — The structural ratio (17.00) is above healthy range. Introduce grammar test coverage requirements and production usage tracking to increase attenuation without reducing generativity.
3. Monitor the 3-bit regulatory gap — Policy-interaction complexity (666 pairwise combinations from 37 policies) is the largest source of internal disturbance. As policies grow, this gap will widen quadratically. Consider policy grouping or hierarchical policy organization.
4. Human escalation is a variety bridge — The confidence threshold system (Article 6: Escalation) is Demerzel’s primary mechanism for absorbing variety that exceeds her regulatory capacity. This is a feature, not a limitation.
5. Evolve grammar Section 6 — The sci-cybernetics.ebnf grammar’s requisite variety section (lines 78-82) should be expanded with quantitative measurement productions.

Connection to CYB-001 and CYB-002

• CYB-001 identified that Ashby’s Law applies to Demerzel and listed variety amplifiers/attenuators qualitatively. CYB-003 makes this quantitative.
• CYB-001 Recommendation 5 (“Monitor variety ratio”) is now operationalized with specific formulas, healthy ranges, and a measurement protocol.
• CYB-002 addressed System 2 dampening. The deadband and hysteresis mechanisms from CYB-002 are themselves variety attenuators — they reduce the variety of signals flowing through coordination channels. CYB-003’s structural attenuation metric should include these when implemented.

Sources

• Ashby, W. R. (1956). An Introduction to Cybernetics. Chapman & Hall. (Chapter 7: Quantity of Variety; Chapter 11: Requisite Variety)
• Ashby, W. R. (1952). Design for a Brain. Chapman & Hall.
• Beer, S. (1979). The Heart of Enterprise. John Wiley. (Chapter 6: Variety Engineering)
• Beer, S. (1985). Diagnosing the System for Organizations. John Wiley.
• Shannon, C. E. (1948). “A Mathematical Theory of Communication.” Bell System Technical Journal, 27(3), 379-423.
• Schwaninger, M. (2024). “What is variety engineering and why do we need it?” Systems Research and Behavioral Science.
• Fathom (2025). Ashby Workshops — AI governance and requisite variety, Independent Verification Organizations (IVO) model.

Follow-Up Questions for Cycle 004

1. Can the 3-bit regulatory gap be closed by hierarchical policy grouping (reducing pairwise interactions from O(n^2) to O(n log n))?
2. How should grammar production usage be tracked to detect dead productions and inform structural attenuation?
3. What is the information-theoretic relationship between Demerzel’s tetravalent logic (T/F/U/C) and Shannon entropy — does U (Unknown) carry more bits than T (True)?

Cross-References

• Prerequisite: state/streeling/courses/cybernetics/en/CYB-001-vsm-ai-governance-mapping.md
• Prerequisite: state/streeling/courses/cybernetics/en/CYB-002-active-dampening-cross-repo-oscillation.md
• Grammar: grammars/sci-cybernetics.ebnf (Section 6, Requisite Variety)
• Department: state/streeling/departments/cybernetics.department.json
• Policy: policies/seldon-plan-policy.yaml