RFC-000: Cybernetic Foundations of Mediated Intelligence Governance

PRIVATE AND PROPRIETARY — NOT A PUBLIC RFC. Owned by Kanjani AI Research & Causum. See NOTICE.md.

Candidate for open publication (see POSITIONING.md, Option B). Relicensing is gated on patent review.

Status: Draft — foundational Category: Grounding Model Series: Foundations Grounds: RFC-031 through RFC-038 (Mediated Intelligence Quality and Concern Evaluation), AIGP Core (RFC-009/010), and the Mars post-invocation verification specification Rests on: Cybernetics (Wiener, Ashby, Conant & Ashby, von Foerster, Beer, Powers) and measurement science (Stevens, Cronbach & Meehl, Messick)

1. Abstract

This document establishes the theoretical substrate on which the Autonomous Intelligence Governance Protocol (AIGP) and the Mars post-invocation verification specification both rest.

That substrate is cybernetics — the transdisciplinary study of communication, control, feedback, adaptation, learning, and regulation in systems — extended with measurement science to make its abstractions empirically operable.

The central claim of this RFC is not novel and is not intended to be:

Governing an autonomous intelligence is a regulation problem. Regulation problems have been studied rigorously for three-quarters of a century. AIGP and Mars are engineered instances of established cybernetic results, not a new theory of AI safety.

This positioning is deliberate. A governance framework invented for one model generation becomes obsolete with the next. Cybernetics does not. By grounding governance in established regulation theory rather than bespoke rules, the protocol inherits durability, falsifiability, and the burden of proof that mature science already carries.

Every downstream specification in the series cites this document rather than re-deriving its own epistemology. This RFC is the common root.

2. Purpose and Scope

2.1 What This Document Does

Establishes cybernetics as the grounding model for mediated intelligence governance.
Identifies the specific, citable results each governance mechanism rests on.
Shows that AIGP (pre-invocation) and Mars (post-invocation) are two halves of a single regulator.
Provides the falsifiability conditions that keep the model scientific rather than philosophical.

2.2 What This Document Does Not Do

It does not define protocol messages, schemas, or endpoints (see AIGP Core and the downstream RFCs).
It does not propose new cybernetic theory. It applies existing theory.
It does not make metaphysical claims. Every load-bearing claim traces to a named prior result.

2.3 The Discipline of This Document

To avoid the failure mode of grand unifying theories — unfalsifiable ambition — this RFC holds itself to one rule:

No claim in this document stands on its own authority. Each rests on a named cybernetic result or a named measurement-science result, cited in §16.

Where this RFC introduces engineering constructs (Domain of Concern, Mediation Vector, post-hoc loop), it shows the prior result each construct implements. The constructs are contributions; the foundations are inherited.

3. The Problem, Framed Cybernetically

An autonomous intelligence is a system that perceives, decides, and acts — increasingly without a human in the immediate loop. When such a system acts on behalf of an organization, the organization must regulate it: keep its behavior within acceptable bounds despite the system’s autonomy and the environment’s unpredictability.

This is, precisely, a regulation problem in the cybernetic sense: maintaining a system’s essential variables within viable limits in the presence of disturbance (Ashby, 1956).

Three properties make mediated intelligence a hard regulation problem:

Perception is mediated before human reasoning. The AI does not merely answer questions after a human has perceived the situation; it increasingly determines what is perceived at all. This shifts the locus of control from decision to perception formation — the concern of Perceptual Control Theory (§8) and of the Fourth Law (DECLARATION.md).
The system has high behavioral variety. An autonomous agent can act in an enormous space of ways. By the Law of Requisite Variety (§6), a regulator of such a system must itself command commensurate variety.
The observer is inside the system. The organization observing its AI is also shaped by that AI’s outputs. There is no neutral external vantage by default — the concern of second-order cybernetics (§7).

The remainder of this document builds the regulator these three properties require.

3.1 The Grounding Claim — Thesis

This RFC advances one thesis:

Any governance of autonomous intelligence that produces empirical verdicts — reproducible across observers, defensible under challenge, falsifiable by evidence — must contain the structure that cybernetics describes: feedback, a model isomorphic to the governed system, variety matched to the concern, and an externalized observer position. Without this structure, governance produces observations but not facts. Observations without ground are interpretable infinitely, reproducible by no one, and defensible against nothing.

AIGP (pre-invocation) and Mars (post-invocation) are the engineering instantiation of this structure. They are not governance tools applied to AI. They are the minimum regulator that established cybernetic results prove is necessary for empirical governance of autonomous systems.

The claim is not that cybernetics is the only possible grounding. It is that any framework producing defensible verdicts must contain — explicitly or implicitly — feedback (Wiener), a model (Conant & Ashby), bounded variety (Ashby), and observer externalization (von Foerster). If a framework lacks any of these and still produces reproducible, falsifiable verdicts, that falsifies this thesis. If every working governance framework contains these properties (perhaps under different names), that confirms it.

This is the standard the protocol holds itself to: not assertion, but structural necessity derived from prior proof.

4. Foundation I — Feedback and Control (Wiener)

Cybernetics begins with Wiener’s (1948) insight that purposive behavior in machines and organisms alike is governed by feedback: a system senses the difference between its actual and desired state and acts to reduce that difference.

A governance system without feedback is not a regulator — it is an open-loop actuator issuing commands with no knowledge of their effect. Most first-generation AI governance is open-loop: it filters inputs or outputs but never closes the loop by comparing what happened to what was intended.

What this grounds in the protocol:

The pre-invocation CHECK is the reference — the declared acceptable state.
Execution is the actuation.
The post-hoc VERIFY (RFC-032) is the feedback — the comparison of actual against reference.
Governance adaptation (Quality Moderator posture) is the corrective action.

Without RFC-032, AIGP would be open-loop. RFC-032 is what makes it a control system in Wiener’s sense.

5. Foundation II — The Good Regulator Theorem (Conant & Ashby)

Conant and Ashby (1970) proved a result that is the single most important foundation for this entire program:

Every good regulator of a system must be a model of that system.

Formally: any regulator that is maximally successful and simple must be isomorphic to the system it regulates. A regulator cannot control what it does not model. Control is modeling.

The consequence for AI governance is direct and non-negotiable:

To govern a mediated intelligence, the governor must contain a model of what that intelligence does and what “good behavior” means for it. Governance without such a model is not regulation — it is noise correlated with, but not controlling, the system.

This theorem is the justification for the entire RFC-034/035 program:

The Domain of Concern (RFC-034) is the model of the system-to-be-regulated — a bounded description of what class of mediated behavior is being governed.
The Mediation Vector (RFC-035) is that model made measurable — the typed, evidence-bound variables through which the regulator represents the system’s state.

A checklist is not a model. A policy file is not a model. The Good Regulator Theorem tells us why those approaches fail: they lack the isomorphism to the governed system that regulation provably requires. A validity-aware measurement model is the minimum structure the theorem permits.

6. Foundation III — The Law of Requisite Variety (Ashby)

Ashby’s (1956) Law of Requisite Variety states:

Only variety can absorb variety.

A regulator can maintain control only if it has at least as much variety (range of distinct responses) as the disturbances it must counter. If the environment can present more distinct situations than the regulator can distinguish and respond to, control fails.

The consequence for AI governance:

Autonomous intelligence has very high behavioral variety. Therefore a regulator of it must command commensurate variety. This has three engineering implications:

A single flat policy is provably insufficient. It has far less variety than the AI it governs. This is why governance must be structured, rich, and domain-specific — not a global rulebook.
The observation apparatus must be extensible. New behaviors require new observable dimensions. This grounds the dialect registry (RFC-038): the governance vocabulary must grow as the governed system’s variety grows.
Governance variety must be matched to the domain, not maximized globally. Requisite variety is requisite — as much as needed, no more. This grounds the Domain of Concern as a bounded model: you match variety to the concern class, avoiding both under-regulation (too little variety, control fails) and over-regulation (unbounded variety, cost without benefit).

The Good Regulator Theorem tells us the regulator must be a model. Requisite Variety tells us how much model is enough.

7. Foundation IV — Second-Order Cybernetics (von Foerster)

First-order cybernetics studies observed systems. Second-order cybernetics (von Foerster) studies observing systems — and recognizes that the observer is part of the system being observed.

This is not a philosophical nicety. It is a structural constraint on AI governance:

The organization observing its AI is also shaped by that AI. A governance system that generates its own success criteria, grades its own outputs, or tests only what it already imagines can go wrong is a first-order system observing itself — and cannot, by construction, perceive its own blind spots.

What this grounds in the protocol:

The internal vs. external observer distinction (Fourth Law; RFC-037 observer accreditation). The internal observer — the operator inside the loop — is structurally insufficient as a sole governance mechanism because it is inside the system it evaluates.
The prohibition on self-evaluation (RFC-032, RFC-035 §19): the observer lineage must be independent from the evaluated generation lineage. This is a direct engineering consequence of second-order cybernetics.
Second-order criterion evolution (RFC-032 §9, RFC-035 gap calculus): the system observes its own observations. When outcomes recur that no criterion anticipated, the criteria themselves must evolve. This is the system becoming, in part, an observer of its own observing — the second-order move made operational.
The edge-case epistemology (EVALUATION-GAPS.md §8): once an edge case is known it is no longer an edge case. A first-order test generator cannot exceed its own epistemic boundary; only a self-modifying (second-order) loop and exogenous challenge can.

7.1 The Objectivity Problem — and Its Resolution

Second-order cybernetics carries a claim that appears to threaten the entire evidentiary program of this protocol:

Observation is observer-relative. There is no observer-independent vantage from which “what really happened” can be read off. The observer constructs what is observed.

Taken naively, this contradicts RFC-035 (validity-aware measurement), RFC-032 (verifiable verdicts), and the Mars recomputation witness — all of which appeal to admissible, reproducible, independently verifiable evidence. If all observation is observer-relative, how can any verdict be reproducible or defensible?

This apparent contradiction is, in fact, the protocol’s central resolution:

AIGP does not claim observer-independent truth. It claims observer-independent reproducibility of an observer-relative verdict.

A recomputation witness never proves “this outcome is objectively good.” It proves a strictly weaker, strictly defensible claim:

Any observer who occupies the declared observation frame — the Domain of Concern, the Mediation Vector, the admissibility rules, the calculation semantics — and applies it to this evidence, derives the same verdict.

This is not a retreat from second-order cybernetics. It is second-order cybernetics applied correctly. Von Foerster’s own prescription for observing systems is to make the act of observation explicit so that another observer can occupy the same position. That is precisely what the protocol’s declarable artifacts do:

The Domain of Concern (RFC-034) declares what is being observed and from what frame.
The Mediation Vector (RFC-035) declares how observation is operationalized — typed, evidence-bound, warranted.
The dialect (RFC-038) packages the complete observation frame as a re-occupiable, versioned, distributable artifact — a research protocol in the scientific sense.

The move is this: because observation is frame-relative, reliability cannot come from a neutral vantage (there is none). It must come from externalizing the frame so that verdicts are reproducible relative to a declared, inspectable, shareable observer position. Observer-relativity ceases to be a threat to reliability and becomes its basis.

This reframes reproducibility itself:

Naive (first-order) reproducibility	AIGP (second-order) reproducibility
The same result appears because reality is fixed and observer-independent	The same verdict appears because the observer frame is declared and re-occupiable
Appeals to an objective vantage	Appeals to a shared, explicit vantage
Breaks under observer-relativity	Is built on observer-relativity

The prohibition on self-evaluation (§7 above) follows from the same logic: an agent evaluating its own output is not occupying a declared, external frame — it is collapsing observer and observed, which second-order cybernetics identifies as the failure mode, not the ideal. Governance requires that the observing frame be occupiable by someone other than the observed system. That is what observer accreditation (RFC-037) and dialect subscription (RFC-038) operationalize.

7.2 Technology as an Observing System — The Transdisciplinary Ground

Cybernetics provides a transdisciplinary grounding model through which diverse academic disciplines can be understood as systems of communication, control, feedback, adaptation, learning, and regulation. This is not a metaphor. It is a structural claim: that the formal properties of regulation (feedback, variety, modeling, recursion) are invariant across domains — biological, mechanical, social, and now artificial.

Demetis (2015) demonstrates this structural invariance by showing that technology itself — not merely the social system that uses technology — is an observing system in the second-order cybernetic sense. Drawing on Luhmann’s functional differentiation, Demetis argues that technology:

Determines what is observed and what is not. Technology does not passively transmit; it selects, filters, amplifies, and occludes. The AI model that retrieves some documents and not others is not a neutral tool — it is an observer that constructs the observed reality for its human recipient.
Operates with its own binary code. Just as Luhmann’s economy operates on payment/non-payment and science on true/false, technology as system operates on a code that determines its own closure and self-reference. For AI governance, this means the governed system has its own operational logic that is not reducible to the intentions of its deployer.
Has non-causal effects across system boundaries. Technology’s interference in other societal subsystems (economy, law, science) cannot be traced to specific causes — it is structural coupling, not linear causation. This is precisely why rule-based governance fails: rules assume causal chains that do not exist at the systemic level. Only a model of the system (Good Regulator Theorem, §5) can regulate structural coupling.

The consequence for AIGP is direct: a governed AI is not a tool being wielded. It is an observing system that constructs observations, operates with autonomous logic, and structurally couples to other systems in ways that resist causal attribution. Governance that treats AI as a tool (input/output filtering) is first-order and structurally inadequate. Governance that treats AI as an observing system — with its own operational closure, binary code, and structural effects — is second-order and appropriate to the phenomenon.

7.3 Pragmatic Interaction as Governance Substrate — The Axiomatized Observer

The observer problem in §7.1 — that observation is frame-relative and that the observer is inside the system — has a further dimension when the governed system communicates. Autonomous agents do not merely act; they interact. They exchange messages, negotiate turn-taking, accumulate relational history, and punctuate the stream of interaction differently from their interlocutors.

Laouris (2026) formalizes this by restating the Watzlawick–Bateson pragmatics as a level-separated, testable axiom system for second-order cybernetics. Five core axioms and five interactional extensions provide the minimum formal structure for understanding how communication produces — and is produced by — observer-relative interpretation:

Axiom	AIGP implication
A1: One cannot not communicate	An agent that does not send a RECORD is still communicating to the regulator — its silence is an observable absence. Governance MUST interpret non-reporting as a signal, not as the absence of one.
A2: One cannot not metacommunicate	Every agent message carries both content and a relational dimension. The TRACE is not only factual (tokens, duration) but relational: it asserts the agent’s compliance posture to the governance authority.
A4: Punctuation is observer-dependent	The same invocation sequence can be segmented differently by the agent (who sees “I completed a task”) and by the regulator (who sees “tool X was called without authorization at time T”). The Domain of Concern (RFC-034) is the mechanism for declaring whose punctuation governs.
A6–A8: Turn-taking, repair, common ground	Agent-to-governance communication is a conversation with structure: the agent sends RECORD, the regulator may respond with scope change, the agent acknowledges. Repair (circuit breaker recovery) and common ground (policy version agreement) are formally required.
Interpretive posterior entropy	Laouris proposes this as an observable: how much interpretive uncertainty remains after a message is received. For governance, this maps directly to the precision of a verdict: a Mediation Vector with high posterior entropy produces unreliable scores. This grounds RFC-035’s validity requirement quantitatively.

The contribution to AIGP is this: the governance protocol is not merely sending data from agent to authority. It is a pragmatic interaction with all the properties Watzlawick identified — unavoidability, metacommunication, code choice, observer-dependent segmentation, and relational dynamics. A governance protocol that ignores these properties will suffer from the same interactional pathologies (blame loops, escalation, relational locking) that Watzlawick documented in human dyads.

The Laouris axioms tell us what MUST be formalized for agent–regulator interaction to succeed:

The code in which governance messages are expressed (digital: ALLOW/DENY/RECORD)
The repair mechanism when communication fails (circuit breaker + re-registration)
The common ground that must be established before regulation can begin (policy version, Domain of Concern subscription)
The punctuation authority that resolves disputes (GOV_APP as the declared observation frame owner)

8. Foundation V — Perceptual Control (Powers)

Powers’ Perceptual Control Theory (1973) established that living control systems control their perceptions, not their outputs. The controlled variable is the perceptual signal, not the behavior.

Applied to mediated intelligence, this is the deepest form of the governance problem:

When AI mediates perception before human reasoning, it does not merely influence decisions — it shapes the perceptual inputs on which all downstream control depends. Whoever controls perception controls the system, upstream of any decision-level oversight.

What this grounds in the protocol:

The First Law (DECLARATION.md): intelligence that scales perception faster than reflection diverges. This is Powers’ insight at civilizational scale.
The pre-invocation locus of AIGP: governance must act at the point of perception formation (before the model runs), not only at the point of decision or output. Governing outputs alone leaves the perceptual layer — the actual locus of control — ungoverned.
Input attestation and context governance (AIGP stages 6–7, RFC-011 ACP): governing what enters the AI’s context window is governing perception. PCT is why this matters as much as governing what the AI outputs.

9. Foundation VI — Recursive Viability (Beer’s VSM)

Stafford Beer’s Viable System Model established that viable systems are recursive: each viable unit contains and is contained by viable units of the same structural form, each with its own regulation, autonomy, and accountability to the level above.

What this grounds in the protocol:

Delegation chains (RFC-010a, agentic governance): an orchestrator delegating to sub-agents is a recursive viable structure. Each agent is a regulated unit within a regulated whole.
Scope narrowing (B.scope = A.scope ∩ request.scope): the VSM requirement that a subunit’s autonomy is bounded by, and accountable to, its parent. A delegate cannot exceed its delegator’s authority — the recursive containment invariant.
Nested governance (Contract Governor, RFC-014): governance authorities that govern other governance authorities are VSM recursion applied to the regulator itself.
Circuit breaking and cascade halt (RFC circuit breaker): the VSM’s requirement that failure in a subunit is contained and does not destroy the viability of the whole.

10. The Impossibility Without Ground — Why Observation is Not Enough

10.1 The Structural Impossibility (Rice’s Theorem Applied)

The claim of this RFC is not merely that a grounding model is useful. It is that governance without structural grounding is mathematically impossible in the general case.

McCann (2026) demonstrates this rigorously: Rice’s theorem (1953) proves that no algorithm can decide non-trivial semantic properties of arbitrary programs — including the property “this program’s effects comply with the governance policy.” Behavioral governance — observing inputs and outputs without a structural model of the system — is undecidable. It is not that current approaches are insufficiently clever; it is that no amount of cleverness can overcome an impossibility result.

This means: any governance system that operates solely at the behavioral level (filtering inputs, classifying outputs, scoring responses) is structurally incapable of producing reliable verdicts for autonomous systems with high behavioral variety. The observations it produces look like governance data but are not — they are correlated with the system’s behavior but cannot, by construction, decide compliance.

The resolution Rice’s theorem permits is: governance must operate at a structural level — governing the architecture, the model, the frame within which the system operates — not merely observing its behavioral surface. This is precisely what the Good Regulator Theorem (§5) requires: the regulator must be a model of the system, not a filter on its outputs.

10.2 Why Cybernetics — and Not Another Meta-Theory

If structural grounding is necessary, why cybernetics specifically? The question is fair. Other transdisciplinary frameworks exist: category theory (structural mathematics), information theory (communication), complex systems science (emergence), and general systems theory (isomorphism). Each provides some of what governance requires. None provides all of it.

Framework	Regulator theory	Observer theory	Boundary theory	Impossibility result
Cybernetics	✓ (Conant & Ashby)	✓ (von Foerster)	✓ (Ashby, Luhmann)	✓ (Requisite Variety as limit)
Category theory	—	—	✓ (morphisms)	—
Information theory	—	—	—	✓ (Shannon limits)
Complex systems	—	✓ (emergence)	—	—
General systems theory	partial	—	partial	—

Cybernetics is the only transdisciplinary framework that simultaneously provides:

A theory of the regulator — what governance must be structurally (Good Regulator Theorem). No other framework proves that regulators must model their systems.
A theory of the observer — who governs, from where, and with what limitations (second-order cybernetics). No other framework formalizes the observer’s position within the system being observed.
A theory of the boundary — what makes observation finite rather than unbounded (Requisite Variety, Luhmann’s system/environment distinction). No other framework proves that regulation fails without bounded variety.
A proven limit theorem — that regulation without model is not weak but impossible (Conant & Ashby). No other framework provides this logical necessity.
A perceptual theory — that the governed system shapes what is perceived, upstream of decision (Powers). No other framework addresses the perception-formation problem.

The claim is therefore not that cybernetics is the only possible grounding. It is stronger:

Any framework that successfully grounds AI governance must contain — explicitly or implicitly — the cybernetic structure: regulator, observer, boundary, and model. If it contains these, it IS cybernetics (perhaps under a different name). If it lacks any of them, it is structurally incomplete and will fail in the ways the theorems predict.

This is a structural uniqueness claim, not a naming claim. The substance is what matters, not the label.

10.3 Observation Without Ground — The Epistemic Status

Without a grounding model, what do AI governance systems produce? They produce observations. These observations have the appearance of empirical data: numbers, scores, classifications, pass/fail verdicts. But they lack the structural properties that make data empirical:

Property	With grounding model	Without grounding model
Reproducible	Yes — any observer who occupies the declared frame derives the same verdict	No — different observers derive different verdicts with no mechanism for resolution
Falsifiable	Yes — the model states conditions under which it is wrong (§13)	No — no declared conditions of failure exist
Bounded	Yes — the Domain of Concern limits what is relevant	No — the observation space is unbounded; any metric can be added or removed
Valid	Yes — the Mediation Vector carries a validity argument (Messick)	No — measurement is asserted, not warranted
Defensible	Yes — under challenge, the frame, model, and evidence can be inspected	No — verdicts rest on implementation choices that are not externalized

The epistemic status of ungrounded observation is: interpretation presented as fact. It is not knowledge. It is opinion wearing the costume of measurement.

This is not an academic distinction. It has direct operational consequences:

An ungrounded AI safety evaluation cannot be defended in court (no declared frame, no reproducibility).
An ungrounded compliance verdict cannot survive audit (no falsifiability, no validity argument).
An ungrounded governance system cannot detect its own blind spots (no second-order observation).

The grounding model is not an optional enhancement. It is the precondition for governance to be governance rather than commentary.

11. The Measurement Bridge

Cybernetics supplies the structure of regulation. It does not, on its own, make the regulator’s model empirical. A model of the governed system that cannot be populated from admissible evidence is a control diagram, not a control system.

The bridge from cybernetic structure to empirical operation is measurement science:

The regulator’s model (Good Regulator Theorem) is instantiated as constructs (Cronbach & Meehl) — the concern is an attribute inferred from indicators, not directly observed.
Validity of that model is an argument, not a number (Messick) — appropriate to the concern, the domain, the evidence, and the consequences of use.
Each modeled variable is measurement-typed (Stevens) — its scale determines which operations, and which gap computations, are legitimate.

This is why RFC-035 (Mediation Vector Profile) is a validity-aware measurement model: it is the point where the cybernetic regulator’s model becomes an empirical instrument. Cybernetics says the regulator must model the system; measurement science says how that model becomes defensible observation.

Cybernetics gives the regulator its necessity. Measurement science gives the regulator its evidence.

12. The Composed Regulator: AIGP + Mars Across the Invocation Boundary

AIGP and Mars are not two systems. They are one regulator operating across the two halves of a single control cycle, divided at the point where the model produces its output.

graph LR
    R[Reference:<br/>declared intent + authority] --> PRE[AIGP pre-invocation<br/>CHECK / authority]
    PRE --> ACT[Execution<br/>the model acts]
    ACT --> POST[Mars post-invocation<br/>conformance verification]
    POST --> FB[Feedback:<br/>verdict + evidence]
    FB --> ADAPT[Adaptation:<br/>posture, scope, criteria]
    ADAPT --> R

Cybernetic element	Where it lives
Reference signal (desired state)	AIGP ANTICIPATE + authority context
Actuation	The governed model invocation
Sensor (measurement)	AIGP RECORD/TRACE + Mars evidence-gathering
Comparator (error detection)	AIGP VERIFY + Mars conformance verdict
Corrective action	Governance adaptation (scope, budget, circuit break)
The regulator’s model	Domain of Concern + Mediation Vector
Reproducibility anchor	D-DNA + Mars recomputation witness

AIGP regulates before and at the moment of action. Mars regulates after the action, verifying the output against a governing specification. Together they close the loop that Wiener’s feedback principle requires and that the Good Regulator Theorem says must contain a model of the governed system.

Neither half is a complete regulator alone. AIGP without post-hoc verification is open-loop at the outcome. Mars without pre-invocation authority verifies outputs whose authority was never established. Composed, they are a complete cybernetic regulator for mediated intelligence.

13. How Downstream Specifications Attach

Each specification in the series implements one or more foundations established here. This table is the dependency inversion: the RFCs depend on RFC-000, not the reverse.

Specification	Primary foundation(s)	What it implements
AIGP Core (RFC-009/010)	Feedback (§4), VSM (§9)	The control loop and recursive governance structure
RFC-011 (ACP)	Perceptual Control (§8)	Governing perception — what enters the context window
RFC-031 (Universal Humanity)	Second-order (§7), PCT (§8)	Baseline regulation across jurisdictions
RFC-032 (Post-Hoc Loop)	Feedback (§4), Second-order (§7)	Closing the loop; criterion evolution
RFC-033 (Quantitative Outcome)	Measurement bridge (§10)	Numerical scoring of outcomes
RFC-034 (Domains of Concern)	Good Regulator (§5), Requisite Variety (§6)	The regulator’s model, bounded per concern
RFC-035 (Mediation Vector)	Good Regulator (§5), Measurement bridge (§10)	The model made an empirical instrument
RFC-036 (Concern Calculation)	Feedback (§4), Measurement bridge (§10)	Error detection from measured state
RFC-037 (Observer Accreditation)	Second-order (§7)	The external observer, made qualifiable
RFC-038 (Dialect Registry)	Requisite Variety (§6)	Extensible, reproducible governance variety
Mars (post-invocation)	Feedback (§4), Good Regulator (§5)	The post-action half of the composed regulator

If a future specification cannot name the foundation it implements, that is a signal the specification may be adding variety without regulatory purpose — a violation of the requisite-variety discipline (§6.3).

14. Falsifiability and Limits

A grounding model that cannot be wrong is philosophy, not science. This section states how the model can fail, so that it remains falsifiable.

13.1 Falsification Conditions

The cybernetic grounding is falsified — for a given deployment — if any of the following hold:

The Good Regulator Theorem is violated in practice with impunity. If a governance deployment with no model of the governed concern regulates as well as one with a model, the theorem’s applicability to this domain is in question.
Requisite variety proves unnecessary. If a single flat policy regulates high-variety autonomous behavior as well as a rich domain-specific model, §6 is falsified for that domain.
The feedback loop provides no signal. If post-hoc verdicts are statistically independent of subsequent outcomes — if VERIFY tells you nothing about what happens next — the loop is not a control loop.
Second-order observation adds nothing. If self-evaluating systems detect their own failures as well as externally-observed ones, §7 is falsified.

Each of these is empirically testable from longitudinal governance data. The model earns its place by surviving these tests, not by assertion.

13.2 Acknowledged Limits

Cybernetics describes regulation structurally; it does not specify what should be valued. The choice of concern, threshold, and acceptable outcome is normative and lies outside this model (it enters through the Domain of Concern and jurisdictional layers).
The Good Regulator Theorem assumes a regulator seeking to minimize the variety of outcomes. Governance that seeks to preserve beneficial variety (e.g., protecting minority perspectives — Third Law) uses the theorem in a more nuanced form: the regulated variable is divergence from acceptable bounds, not variety as such.
Requisite variety is a necessary, not sufficient, condition. A regulator can have adequate variety and still fail if its model is wrong (poor construct validity — RFC-035 §4.2).

15. Relationship to the Five Laws

The Five Laws of Mediated Intelligence Systems (DECLARATION.md) are the civilizational expression of these cybernetic foundations. This RFC is the engineering expression. They are the same structure at two scales.

Law	Cybernetic foundation
I — Divergence (perception outpaces reflection)	Feedback (§4) + Perceptual Control (§8)
II — Dual Reflection (evaluate process, not only output)	Second-order cybernetics (§7)
III — Representational Closure (optimization drifts from reality)	Requisite Variety (§6) + Good Regulator (§5)
IV — External Observation (internal observer insufficient)	Second-order cybernetics (§7)
V — Organizational Intelligence (legitimacy-constrained learning)	VSM recursion (§9) + feedback (§4)

The Laws state what must hold for human civilization under mediated intelligence. This RFC states the regulation theory that makes those Laws implementable rather than aspirational.

16. Non-Goals

This RFC does not define enforcement mechanics, message formats, or APIs.
It does not claim cybernetics is the only valid grounding — only that it is a sufficient, established, and durable one.
It does not resolve normative questions (what to value); it structures how declared values are regulated.
It does not replace the downstream RFCs; it grounds them.

17. References

Foundational works underpinning this grounding model. Content drawn from these sources is paraphrased; consult originals for authoritative text.

Cybernetics

Wiener, N. (1948). Cybernetics: or Control and Communication in the Animal and the Machine. MIT Press.
Ashby, W. R. (1956). An Introduction to Cybernetics. Chapman & Hall. — Law of Requisite Variety.
Conant, R. C., & Ashby, W. R. (1970). Every good regulator of a system must be a model of that system. International Journal of Systems Science.
von Foerster, H. (1974). Cybernetics of Cybernetics. — Second-order cybernetics; the observer in the system.
Beer, S. (1972). Brain of the Firm. — The Viable System Model; recursive viability.
Powers, W. T. (1973). Behavior: The Control of Perception. — Perceptual Control Theory.
Shannon, C. E. (1948). A Mathematical Theory of Communication. Bell System Technical Journal. — Information and communication.
Demetis, D. S. (2015). Technology as an Observing System: A 2nd Order Cybernetics Approach. — Technology as a functionally differentiated system with autonomous observational capacity; the systemic character of technological interference across function-system boundaries.
Laouris, Y. (2026). Beyond Watzlawick: Axioms of Pragmatic Interaction for Second-Order Cybernetics. Kybernetes, 55(13), 69–103. — Formalization of Watzlawick–Bateson pragmatics as level-separated, testable axioms; observer-dependent punctuation; interpretive posterior entropy as an observable for code-variance; applicability to human–AI and multi-agent interaction.

Measurement Science (the empirical bridge) 10. Stevens, S. S. (1946). On the Theory of Scales of Measurement. Science. 11. Cronbach, L. J., & Meehl, P. E. (1955). Construct Validity in Psychological Tests. Psychological Bulletin. 12. Messick, S. (1989). Validity. In Educational Measurement (3rd ed.).

Communication and Interaction 13. Watzlawick, P., Beavin, J. H., & Jackson, D. D. (1967). Pragmatics of Human Communication. W. W. Norton. — The five axioms of communication; relationship vs. content; symmetry/complementarity. 14. Bateson, G. (1972). Steps to an Ecology of Mind. University of Chicago Press. — Information as “a difference that makes a difference”; schismogenesis. 15. Luhmann, N. (1995). Social Systems. Stanford University Press. — Autopoietic social systems; functional differentiation; binary coding of system operations.

Internal cross-references 16. DECLARATION.md — The Five Laws of Mediated Intelligence Systems. 17. RFC-034 (Domains of Concern), RFC-035 (Mediation Vector Profile), RFC-038 (Dialect Registry). 18. EVALUATION-GAPS.md — The governance/quality non-causality and edge-case epistemology.