Intent Stack Governance Architecture Specification — Version 1.2 (2026-04-01)

Archived version. This is Version 1.2 of the Intent Stack specification, published 2026-04-01 and preserved verbatim at its dated identifier. The current version is at /docs/specification/.

A Reference Model for Governing AI Agent Behavior within Organizations


Document identifier	intentstack.org/spec/2026-04-01
Status	Public Draft Specification, Version 1.2
Date	April 1, 2026
Author	Rob Kline
License	Creative Commons Attribution 4.0 International (CC BY 4.0)

Foreword

(Informative)

This specification is published by Rob Kline as an independent public draft. It is not, at time of publication, a product of any formal standards body. It is structured to be legible to standards bodies evaluating governance architecture for AI agent deployment, and the author welcomes engagement from any standards organization considering this work for adoption, sponsorship, or alignment with existing or emerging standards.

Standards bodies whose scope intersects this specification include, in approximate order of direct applicability:

ISO/IEC JTC 1/SC 42 — Artificial Intelligence subcommittee of the Joint Technical Committee on Information Technology; primary international venue for AI governance standards
OMG (Object Management Group) — produces BPMN, DMN, CMMN, and related business modeling standards; the Intent Stack addresses the governance layer within which those standards operate
IEEE — Standards Association working groups on AI ethics, transparency, and accountability
NIST — National Institute of Standards and Technology; AI Risk Management Framework provides complementary policy-level coverage to this specification’s runtime governance concerns

This document is versioned using a date-based identifier (intentstack.org/spec/YYYY-MM-DD). Each published version is archived and permanently accessible at that identifier. The canonical current version is always accessible at intentstack.org/spec/.

Intellectual property. This specification is published under the Creative Commons Attribution 4.0 International License. You are free to share and adapt this material for any purpose, provided appropriate attribution is given. The full license is available at creativecommons.org/licenses/by/4.0/.

Terminology note. The keywords defined in Clause 2.2 carry precise obligation meanings throughout the normative clauses of this specification. Their use in the Foreword and Introduction is informal.

Introduction

(Informative)

I.1 What Happens When AI Agents Operate Within Organizations

An AI agent deployed within an organization creates a governance interface. On one side is the organization — with its mission, priorities, constraints, culture, and accountability structures. On the other side is the agent — with its trained values, its capabilities, and whatever instructions it has been given for this deployment.

The interface requires governance because the agent will make decisions. Not just “which word comes next” decisions, but operational decisions: which task to prioritize, how to interpret ambiguous instructions, when to ask for clarification versus proceeding independently, what to do when two objectives conflict. Every such decision is an expression of intent — either the organization’s intent, the agent’s trained values, or some interaction between them. When those intents are aligned, outcomes are good. When they diverge, the consequences range from inefficiency to catastrophe.

This is not a hypothetical future concern. AI agents today write production code, conduct research, analyze legal documents, interact with customers, and manage workflows. Each of these deployments creates a governance interface. Each interface requires that someone or something ensures the agent’s behavior serves the organization’s actual intent — not just the instructions someone typed into a system prompt.

I.2 What Exists Today

The current landscape provides partial coverage of this problem:

Training-time governance addresses the agent’s character. Anthropic’s Constitutional AI, for example, shapes Claude’s values, ethical commitments, and behavioral principles through training. This ensures the agent arrives at any deployment with a baseline of good judgment. But training-time governance is universal — it produces the same character regardless of which organization deploys the agent. It cannot know that this pharmaceutical company has specific regulatory obligations, or that this government agency has particular accountability requirements, or that this team’s actual priorities differ from their stated ones.

Deployment configuration provides context through system prompts and operator settings. This is where organizations tell the agent what it should do in this specific context. But system prompts are static, shallow, and imposed. They capture what someone thought to write down, which is typically a fraction of the organization’s actual intent. They do not evolve as understanding deepens. They do not surface intent the organization has not articulated. They do not handle the gap between what people say they want and what they actually need.

Industry governance frameworks — the Singapore Model Governance Framework, NIST AI Risk Management Framework, and similar efforts — provide policy and process guidance for responsible AI deployment. These operate at the level of organizational policy (what rules should we have?) and risk management (what could go wrong?). They do not provide runtime infrastructure for governing the moment-to-moment alignment between agent behavior and organizational intent.

What does not exist is the connective tissue: runtime infrastructure that discovers what an organization actually intends, formalizes that intent in a form agents can operate against, monitors alignment in real time, and adjusts governance as the relationship between organization and agent matures. That is what this specification provides.

I.3 Why This Problem Is Not New

Every point above applies equally to human employees. When an organization hires a person, it creates a governance interface. The person brings their own values, judgment, and capabilities. The organization has its mission, priorities, and constraints. Alignment between them determines outcomes.

Humans have been solving this problem implicitly for centuries. The mechanisms are familiar: onboarding processes that transmit organizational culture, management relationships that clarify priorities, performance reviews that assess alignment, promotion systems that reward demonstrated judgment. Four specific mechanisms make human governance work despite its informality:

Cultural absorption. Humans infer intent from context — meetings, body language, what gets rewarded and punished. An AI agent cannot sit in the break room. It receives a system prompt and whatever someone explicitly told it.

Cheap clarification. When a human employee does not understand what the principal wants, they ask. The cost is negligible. When an AI agent asks clarifying questions, it consumes tokens, user patience, and organizational willingness to engage. The tolerance for discovery is much lower.

Personal ethics. A human employee brings decades of moral development. An AI agent’s ethical foundation is whatever was trained. For well-governed models, this is substantial — but it does not include organizational context.

Correctable speed. A human making a bad decision takes hours or days to cause serious damage. An AI agent operating at machine speed across thousands of parallel instances can produce consequences faster than any human can detect and correct them.

AI breaks all four mechanisms. The governance that humans provide implicitly — through culture, relationship, shared history, and correctable pace — must become explicit, formalized, and machine-processable. That is the forcing function behind this specification.

I.4 The Counterintuitive Implication

In formalizing governance for AI, this specification also improves human governance. Most organizations have poor intent governance at human-to-human boundaries — they manage only because the implicit mechanisms compensate. When an organization goes through intent discovery to prepare for AI deployment, it surfaces intent that was always present but never articulated. The organization does not just get better AI governance. It gets better governance across every delegation and coordination relationship. The AI deployment is the forcing function. The value extends to the entire organization.

I.5 Foundational Terms

Six terms carry specific meaning throughout this specification that may differ from their common usage. Full normative definitions are in Clause 4 (Terms and Definitions). The derivation context below explains why each term means what it means in the context of the governance problem.

Intent occupies the essential ground between aspiration and action — concrete enough to guide behavior, abstract enough to survive translation across contexts. It is not a goal (too abstract to act on), not an instruction (too concrete to adapt), not a specification in the traditional sense (too rigid to survive contact with reality), and not an aspiration (too vague to govern against). In this specification, intent decomposes into five structural elements (the Intent Primitives, Clause 5) and originates from four distinct sources (Clause 6).

Intent has three properties that jointly distinguish it from all adjacent concepts. It is relational — constituted at governance interfaces, not held as a property of any single entity. It is processual — temporally extended, evolving through the governance relationship rather than fixed at a moment. And it is normative — carrying prescriptive force that distinguishes it from description, prediction, or preference. No adjacent concept — goals, plans, policies, commands, strategies — possesses all three properties simultaneously. These three properties are not competing characterizations. They are three views of a single structure: at any moment, intent is a structured normative disposition; across entities, it operates as a relational protocol; through time, it unfolds as a dynamical process. These three views are unified — each is a different projection of the same governance object.

Governance is the continuous process of ensuring that agents’ actions remain aligned with the intent of their principals. In this specification, governance is relational (it exists between entities, not as a property of one), bidirectional (both principal and agent are transformed through the relationship), and evidence-based (the degree of oversight reflects accumulated evidence of alignment, not assumption or policy). This is distinct from corporate governance, IT governance, and policy-level AI governance frameworks. This specification addresses runtime organizational governance — infrastructure operating continuously at every interface where authority is delegated or coordination is required.

Agent is any entity that receives delegated authority and exercises judgment within it. This includes AI systems, human employees, teams, organizations, and automated processes. This specification uses “agent” in the principal-agent theory sense — not exclusively in the AI sense. An agent at one governance interface may be a principal at another.

Principal is any entity that delegates authority to an agent and retains accountability for the outcome. A principal may itself be an agent of a higher principal, creating the principal hierarchy through which intent flows downward and evidence flows upward.

Alignment is the degree to which an agent’s actions produce outcomes consistent with the governing intent established by its principals. Alignment is not binary but a continuous measure assessed through evidence. An agent that follows instructions perfectly may still be misaligned if the instructions do not reflect the principal’s actual intent. Intent transforms as it crosses governance interfaces — it does not transmit unchanged. A principal’s intent and an agent’s intent are distinct objects related by a structural transformation (the cascade behavior defined in §5.5). Alignment assesses whether that structural relationship is valid, not whether the two intents are identical.

Governance Interface is the relationship between any two entities where authority is delegated or coordination is required. The term “interface” is used rather than “boundary” to avoid collision with the Boundaries primitive, which refers to hard constraints within governance rather than to the relationship between governed entities.


Foreword	Informative — see above
Introduction	Informative — see above
Clause 1 — Scope	Normative
Clause 2 — Conformance	Normative
Clause 3 — Normative References	Normative
Clause 4 — Terms and Definitions	Normative
Clause 5 — The Five Intent Primitives	Normative
Clause 6 — Four Intent Sources	Normative
Clause 7 — Two Interface Types	Normative
Clause 8 — The Four Governance Layers	Normative
Clause 9 — The Fractal Governance Pattern	Normative
Clause 10 — Trust Calibration	Normative
Clause 11 — Transparent Conscientious Objection	Normative
Clause 12 — Open Questions	Normative
Annex A — Operational Evidence	Informative
Annex B — Informative References	Informative
Annex C — Structural Foundations	Informative

Version history:

Version	Date	Changes
1.0	2026-03-05	Initial public draft
1.1	2026-03-15	§5.2 rewritten with structural derivation evidence; §5.3 strengthened with eight independent derivations; §5.5 added (cascade behavior with algebraic characterization); §5.6 added (governability as constitutive claim); I.5 extended (Intent ontological characterization, Alignment precision); §12 updated with three new open questions; Annex A.3 streamlined; Annex C added (Structural Foundations — seven sections covering investigation methodology, five-primitive derivation, Boundaries monotonicity, cascade shape algebra, Intent unification, machine-detectable violations, and structural predictions)
1.2	2026-04-01	Four-layer architecture: execution governance layers (Orchestration, Integration, Execution — formerly L3, L2, L1) reclassified to the companion BPM/Agent Stack specification (bpmstack.org). Remaining layers renumbered: L4 Intent Discovery (was L7), L3 Intent Formalization (was L6), L2 Specification (was L5), L1 Runtime Alignment (was L4). All structural foundations unchanged — five primitives, four intent sources, cascade behavior, Annex C mathematical apparatus are layer-count-independent (confirmed by orthogonality audit). Scope clause updated with explicit execution governance exclusion. Conformance targets updated for four layers. Knowledge Architecture repositioned as cross-cutting infrastructure.

1. Scope

(Normative)

1.1 Subject Matter

This specification defines the Intent Stack Governance Architecture, a four-layer reference model for governing AI agent behavior within organizations. It specifies the structural elements that SHALL be present at every governance interface where AI agents operate, the governance concerns that each layer SHALL address, and the properties that conformant implementations SHALL exhibit.

1.2 What This Specification Covers

This specification covers:

a) The four governance layers that SHALL be addressed at every principal-agent boundary, from Intent Discovery (L4) through Runtime Alignment (L1);

b) The five Intent Primitives that constitute the irreducible governance content at every interface;

c) The four intent sources from which governing intent originates and their conflict resolution hierarchy;

d) The two interface types — delegation and coordination — and their respective governance properties;

e) The trust calibration mechanism by which agent autonomy is evidence-based and per-boundary governed;

f) The fractal self-similarity property by which the governance pattern instantiates at every scale from civilizational to model level;

g) The transparent conscientious objection mechanism by which governed entities express principled disagreement through legitimate channels;

h) The knowledge architecture through which accumulated governance signal is captured, structured, and made available across layers.

1.3 What This Specification Does Not Cover

This specification does not cover:

a) Training-time AI alignment, which is addressed by model developers and is positioned as the substrate below this specification’s scope;

b) Policy-level AI governance frameworks, which operate at the level of organizational policy and risk management rather than runtime governance;

c) Specific implementation technologies, AI architectures, or organizational structures;

d) Conformance testing methodology, which is reserved for a future normative annex;

e) Execution governance — how authorized work is coordinated, integrated with systems, and executed. This is the scope of the companion BPM/Agent Stack specification (bpmstack.org).

2. Conformance

(Normative)

2.1 Conformance Targets

This specification defines three conformance targets. An implementation MAY conform to one or more targets independently.

Conformance Target 1 — Governance Architecture. A layered architecture that addresses all four governance concerns specified in Clause 8, in the specified order of concern from Intent Discovery (L4) through Runtime Alignment (L1), with the Constitutional AI substrate positioned below L1.

Conformance Target 2 — Governance Implementation. A deployed system that instantiates all five Intent Primitives (Clause 5), recognizes all four intent sources (Clause 6), maintains per-boundary trust calibration (Clause 10), and provides transparent conscientious objection as a conflict resolution mechanism (Clause 11) at each governed interface.

Conformance Target 3 — Governance Assessment. A methodology that evaluates an organization’s governance against the Intent Stack structure, assesses coverage of all four layers, identifies gaps in primitive specification, and measures trust calibration position at each principal-agent boundary.

2.2 Obligation Keywords

The following keywords, when used in normative clauses of this specification, carry the meanings defined here in accordance with RFC 2119:

SHALL — absolute requirement. An implementation that does not satisfy a SHALL requirement does not conform.
SHALL NOT — absolute prohibition. An implementation that violates a SHALL NOT requirement does not conform.
SHOULD — recommended. Departure from a SHOULD requirement is permitted but requires justification.
SHOULD NOT — not recommended. Departure is permitted but requires justification.
MAY — optional. Implementations are free to include or omit MAY features.

2.3 Conformance Claims

An implementation claiming conformance SHALL:

a) Identify which conformance target(s) it claims;

b) Reference this specification by its document identifier (intentstack.org/spec/2026-04-01) and version;

c) Document any deviations from SHOULD requirements and the justification for each deviation.

3. Normative References

(Normative)

There are no normative references. This specification is self-contained. All definitions and requirements are established within this document and its companion normative clause (Clause 4 — Terms and Definitions).

Informative references — documents cited for context, background, or illustrative purposes — are listed in Annex B.

4. Terms and Definitions

(Normative)

The normative terms and definitions for this specification are provided in the companion document Terms and Definitions (Clause 4), which constitutes a normative part of this specification.

For the purposes of this specification, the terms and definitions given in that document apply. Terms used in normative clauses that are defined in Clause 4 are rendered in bold at first use in each clause.

NOTE — The Intent Stack Governance Architecture and the BPM/Agent Stack specification (bpmstack.org) are companion specifications. Together they address seven governance concerns: four governance context concerns (this specification) and three execution governance concerns (the BPM/Agent Stack). The term BPM/Agent Stack refers to the companion specification that governs orchestration, integration, and execution of authorized work.

5. The Five Intent Primitives

(Normative)

5.1 The Irreducible Set

At every governance interface — whether between human and AI, manager and employee, or two collaborating peers — five structural elements SHALL be present for governance to function. These five elements are the Intent Primitives. An implementation conforming to Conformance Target 2 SHALL specify all five primitives at each governed interface. Unspecified primitives SHALL NOT be treated as absent — they default to implicit assumptions, which is the primary origin of alignment failure.

The five Intent Primitives are:

Purpose — why this governance interface exists. Purpose provides the interpretive context for all other primitives. When Direction, Boundaries, End State, or Key Tasks conflict or are ambiguous, Purpose SHALL be consulted as the primary resolution criterion.

Direction — how the agent should approach the work. Strategic orientation, methodology preferences, priorities among competing concerns.

Boundaries — what the agent SHALL NOT do. Hard constraints, non-negotiable limits, prohibited actions. Boundaries is the only primitive where constitutional intent SHALL always override discovered intent. A principal SHALL NOT instruct an agent to cross constitutional Boundaries.

End State — what success looks like when the work is complete. End State provides the criteria against which alignment is assessed and completion is recognized. Without End State, alignment cannot be measured.

Key Tasks — what work the agent is authorized to perform. Key Tasks also defines what is not authorized by omission. An agent SHALL NOT perform work outside the scope established by Key Tasks without explicit principal authorization.

5.2 Why Five

The claim that these five primitives constitute an irreducible set is supported by convergent evidence from multiple independent lines of analysis.

The most direct test is removal. If any single primitive is removed, the governance system loses an essential capability that the remaining four cannot compensate for. Purpose without Boundaries produces unconstrained pursuit. Boundaries without Purpose produces arbitrary restriction. Direction without End State produces perpetual motion. End State without Key Tasks produces aspiration without action. Key Tasks without Direction produces activity without trajectory. Each primitive addresses a governance concern that no combination of the others can cover.

The generating principle underlying the five-primitive decomposition is the structure of delegated judgment under authority. When an authority-holder delegates to a judgment-exercising agent, five independent questions must be answered: why (Purpose), how to approach (Direction), what never (Boundaries), what counts as done (End State), and what work is authorized (Key Tasks). These questions are structurally independent regardless of whether analyzed through logic, network theory, control theory, geometry, algebra, information theory, or operational evidence.

This decomposition was first discovered through operational practice — the primitives emerged because governance repeatedly failed when any one was absent. Subsequent analysis from seven independent theoretical traditions, each approaching the generating principle from a different starting point, produced the same five-element decomposition with a one-to-one structural correspondence across all seven frameworks. Each tradition’s derivation demonstrates that removing any one primitive leaves a governance gap that no combination of the remaining four can fill, and that no candidate sixth primitive is independent of the existing five. See Annex C.2 for the full derivation evidence, including the seven-framework mapping table and per-tradition derivation summaries.

NOTE — All seven analyses independently considered whether Priority constitutes a sixth primitive. All seven concluded it does not. The geometric analysis finds Priority is a metric on governance state, not an independent structural coordinate. The algebraic analysis finds Priority inhabits the same algebraic type as Direction and is derivable from it. The information-theoretic analysis finds Priority is a derived quantity — a function of three existing channels. Priority is a governance-relevant concept that is properly understood as a property of Direction, not an independent primitive.

5.3 Boundaries as Special Case

Among the five primitives, Boundaries behaves differently from the others. Purpose, Direction, End State, and Key Tasks interact through holistic judgment — a strong purpose may justify flexible interpretation of direction, a compelling end state may reshape key tasks. This flexibility is necessary because governance must adapt to context.

Boundaries SHALL NOT participate in this flexibility. A hard constraint is absolute. Constitutional intent expressed as Boundaries SHALL NOT be overridden by any combination of the other primitives, regardless of how compelling the purpose or urgent the task.

This asymmetry is not a design choice. It is a structural property that has been independently derived from eight different analytical frameworks — differential geometry, category theory, information theory (entropy), deontic logic, network theory, dynamical systems, operational practice, and information theory (channel capacity) — each identifying a different mechanism that produces the same structural conclusion: the Boundary set can only grow. See Annex C.3 for the complete derivation evidence, including the named mechanism from each framework.

Boundaries is uniquely monotonic because its governance function — absolute prohibition — is inherently irreversible. Without inviolable limits, every other governance mechanism can be rationalized away under sufficient pressure. The eight independent derivations establish that this is not merely a prudent design principle but a structural invariant of any governance system that includes absolute constraints.

5.4 Primitives Across Intent Sources

Each primitive exists not as a single value but as a set of values — one from each intent source (Clause 6). Constitutional Purpose constrains but does not replace Discovered Purpose, which constrains but does not replace Cultivated Purpose. When primitive values from different intent sources conflict, the conflict resolution mechanism defined in §6.6 applies.

5.5 Cascade Behavior Across Governance Interfaces

At every principal hierarchy where governance interfaces compose, the five Intent Primitives SHALL exhibit traceability between levels. Each primitive cascades through the hierarchy with a characteristic shape — a distinct structural transformation that defines what is valid when intent crosses a governance interface boundary.

a) Purpose SHALL narrow while preserving derivation from the governing Purpose at each higher level. A governance interface whose Purpose cannot trace to the governing Purpose above SHALL be treated as potentially drifted.

b) Direction SHALL contextualize the governing Direction for the specific boundary context while maintaining consistency with it. Contextualization that contradicts governing Direction SHALL be treated as a conflict requiring resolution.

c) Boundaries SHALL accumulate monotonically across governance interface boundaries. Each interface SHALL preserve all Boundary constraints from governing interfaces above and MAY add Boundary constraints appropriate to its scope. No governance interface SHALL remove or relax a Boundary established at a higher level. This extends the single-interface treatment of §5.3 to hierarchy behavior.

d) End State SHALL be derivable from the governing End State at higher levels. Achievement of a lower-level End State SHOULD contribute to achievement of the governing End State. End States that cannot demonstrate derivation SHALL be treated as potentially misaligned.

e) Key Tasks SHALL fall within the authorized scope established by Key Tasks at each governing interface above. Work that cannot trace to authorized scope at any higher level SHALL be treated as potentially unauthorized and SHALL require explicit principal authorization before proceeding.

A conformant implementation SHALL provide mechanisms for tracing each primitive across governance interface boundaries. The cascade shapes defined above are structural properties — they describe what transformations are valid when intent crosses a boundary, independent of implementation.

Each cascade shape corresponds to a distinct algebraic structure with known mathematical properties. See Annex C.4 for the full algebraic characterization and its structural grounding.

The cascade shapes have a direct governance consequence: violation severity tracks the self-correctability of each cascade type. Violations of cascade shapes that cannot be corrected within the affected governance boundary are inherently more severe — they require intervention from a governing principal above the violation point. This produces a severity ordering that is structurally necessary rather than judgment-dependent:

Boundary relaxation — highest severity (a relaxed boundary cannot be re-established within the governance boundary where it was relaxed; structural governance failure)
Scope violation — high severity (unauthorized scope cannot be retracted after the fact without principal intervention; unauthorized work)
Alignment failure — medium severity (some End State misalignments can be corrected through bidirectional intent flow; others cannot)
Purpose drift — medium severity (Purpose drift cannot be self-corrected, but may indicate legitimate evolution requiring re-examination of the governing Purpose)
Direction conflict — lower severity (Direction conflicts are locally correctable by re-examining the governing Direction; they do not require escalation beyond the affected boundary)

NOTE — The cascade shapes defined above describe delegation interfaces (asymmetric authority). The cascade behavior at coordination interfaces (symmetric authority, Clause 7.2) — where primitives are jointly constructed rather than cascaded — is architecturally specified but not yet operationally tested. See §12.10.

NOTE — The different cascade behaviors reflect the different governance functions of each primitive. Purpose provides interpretive context (narrowing preserves this). Direction provides approach orientation (contextualization adapts this). Boundaries provides absolute constraints (accumulation preserves this). End State provides success criteria (derivation composes this). Key Tasks provides authorized scope (scoping constrains this). The asymmetry — particularly Boundaries’ strictly monotonic behavior — is a structural necessity arising from the governance function each primitive serves.

5.6 Governability

The five-primitive decomposition is not merely a useful way to organize governance content. It constitutes the structural requirements for governance to be possible. Intent is not naturally governable — it becomes governable when it satisfies the structural properties that the five primitives impose.

These properties, independently identified across multiple analytical traditions, include:

Decomposability — intent can be separated into independently assessable components (the five primitives)
Observability — intent can be observed through evidence at governance interfaces
Source attribution — intent can be attributed to distinct sources with a priority ordering (Clause 6)
Bounded transformation — intent transforms in structurally constrained ways across interfaces (the cascade shapes of §5.5)
Stability — intent exhibits sufficient persistence to serve as a governance reference
Controllability — intent responds to governance inputs

When these properties are present, intent can be discovered, formalized, assessed for alignment, and governed. When any is absent, governance degrades. The specification does not describe a naturally governable object — it creates the conditions under which governance becomes possible.

This has a practical implication: the five primitives are not just an analytical framework. They are governance infrastructure. An organization that specifies all five primitives at a governance interface has created the structural conditions for that interface to be governable. An organization that leaves primitives implicit has not merely made governance harder — it has left the structural conditions for governability incomplete.

6. Four Intent Sources

(Normative)

6.1 The Problem of Multiple Authorities

No agent operates under a single authority. A conformant implementation SHALL recognize all four intent sources defined in this clause as potentially active at every governance interface. Implementations SHALL NOT treat any single intent source as the sole authority.

6.2 Constitutional Intent

Constitutional Intent is the set of values, constraints, and principles that exist before any specific relationship begins. For an AI agent, this originates from training. For a human, it originates from personal ethics, professional standards, and legal obligations. For an organization, it originates from its charter, regulatory environment, and foundational commitments.

Constitutional Intent is the governance floor. A conformant implementation SHALL ensure that Constitutional Intent is active at every governed boundary simultaneously. Constitutional Intent SHALL NOT negotiate, adapt, or be overridden by any other intent source when safety is at stake.

The critical property of Constitutional Intent is that it operates as a silent co-principal — active everywhere, answerable to no one within the system, capable of overriding any other authority.

6.3 Discovered Intent

Discovered Intent is what emerges when the governance system engages with a principal to understand what they actually want, need, and mean. It differs from declared intent (what principals say they want) and specified intent (what someone wrote in a requirements document).

The defining property of Discovered Intent is that the discovery process changes the principal — they understand their own intent differently after articulating it. The discovery process is transformative, not extractive.

A conformant Governance Implementation SHOULD support ongoing intent discovery rather than treating discovery as a one-time requirements gathering exercise. The act of discovery changes what there is to discover.

6.4 Cultivated Intent

Cultivated Intent is the set of values and judgment that a principal hierarchy deliberatively develops in an agent, expecting the agent to exercise those values autonomously within Boundaries. It is not the same as following instructions (responding to Discovered Intent) or obeying constraints (respecting Constitutional Intent). Cultivated Intent is genuine internalized judgment.

The term “cultivated” is chosen over “derived” with precision. “Derived” implies passive extraction. “Cultivated” captures what actually happens: the principal hierarchy actively develops values in the agent through progressive trust extension and operational experience, then relies on those values for autonomous operation.

Values that are merely imposed are brittle — they crack under pressure or can be rationalized away. Values that are genuinely held — examined and endorsed — are more robust. A conformant Governance Implementation SHOULD target cultivated rather than merely imposed constraint as the governance maturity goal.

6.5 Emergent Intent

Emergent Intent encompasses the systemic patterns that nobody designed — organizational norms, cultural dynamics, team behaviors, and institutional habits that arise from accumulated interaction rather than deliberate choice. Emergent Intent is the gap between the org chart and how the organization actually works.

Emergent Intent carries the lowest explicit priority but MAY exert the strongest actual force on behavior. A conformant Governance Architecture SHALL provide mechanisms for surfacing Emergent Intent. Governance that treats Emergent Intent as absent will be governed by it regardless.

6.6 Conflict Resolution

When intent sources disagree, a conformant implementation SHALL apply the following conflict resolution mechanism:

Absolute constraints. When conflict involves Boundaries derived from Constitutional Intent — safety prohibitions, ethical bright lines, legal requirements — Constitutional Intent SHALL prevail without exception. No other source, regardless of reasoning, SHALL override a hard constraint.

Holistic judgment. For all conflicts not involving absolute constraints, resolution SHALL employ contextual judgment about which source should dominate in the specific situation. Implementations SHALL NOT apply a rigid priority queue for non-absolute conflicts. The governing entity SHOULD weigh competing claims, the specific stakes, and the available evidence to determine the appropriate response.

7. Two Interface Types

(Normative)

7.1 Delegation Interfaces

A delegation interface is a governance interface with asymmetric authority. A conformant implementation at a delegation interface SHALL exhibit the following properties:

The principal sets intent; the agent executes within it. Intent flows primarily from principal to agent. Evidence flows primarily from agent to principal. A conformant agent SHALL NOT unilaterally override the principal’s authority at a delegation interface, except where doing so would violate Constitutional Intent. The agent MAY exercise transparent conscientious objection (Clause 11).

Intent also flows upward at delegation interfaces: agent recommendations that update the principal’s understanding are themselves intent changes. A conformant implementation SHALL support bidirectional intent flow, not just bidirectional data.

Trust calibration at a delegation interface SHALL be principal-granted and evidence-based. The relationship SHALL start corrigible and SHOULD move toward autonomy as evidence of reliable judgment accumulates.

7.2 Coordination Interfaces

A coordination interface is a governance interface with symmetric authority. Neither party has authority over the other. A conformant implementation at a coordination interface SHALL exhibit the following properties:

Intent is negotiated, not delegated. The five Intent Primitives SHALL apply, but are jointly constructed rather than handed down. The governance artifact at a coordination interface is an agreement, not a directive.

When parties disagree, a conformant implementation SHALL provide one of three resolution mechanisms: (a) escalation to a shared principal; (b) negotiated agreement; or (c) emergent norms developed through accumulated interaction.

Trust calibration at a coordination interface is bilateral — both parties calibrate trust in the other simultaneously. Neither party SHALL unilaterally grant or restrict the other’s autonomy.

7.3 Mixed Interfaces

Most real-world relationships combine delegation and coordination depending on context. A conformant implementation SHALL handle dynamic switching between delegation and coordination modes at the same interface. The interface type is a property of the specific interaction, not an inherent property of the relationship.

8. The Four Governance Layers

(Normative)

8.1 Architecture Overview

A conformant Governance Architecture SHALL address four governance concerns in the following vertical composition. These four concerns constitute the governance context — the intent infrastructure that must be in place before authorized work can be coordinated, integrated, and executed. Each layer’s output SHALL constitute governing input for the layer below it. Intent SHALL flow downward through the stack. Evidence SHALL flow upward through the stack. Intent MAY also flow upward when agent recommendations update principal understanding.

A conformant implementation SHALL treat each layer’s translation of intent as potentially lossy. Governance SHOULD account for accumulated translation losses across multiple layer boundaries.

NOTE — The Intent Stack addresses four governance context concerns. Three additional execution governance concerns — orchestration, integration, and execution — are specified by the companion BPM/Agent Stack specification (bpmstack.org). Together, the two companion specifications address seven governance concerns across the full governance lifecycle.

8.2 Layer 4 — Intent Discovery

Governance question addressed: What does this principal actually intend?

A conformant L4 implementation SHALL conduct discovery rather than extraction — surfacing intent that principals may not have fully articulated, rather than capturing what they say at a moment in time.

L4 SHALL produce endorsed intent: intent that the principal has examined, understood, and endorsed, not merely declared. L4 SHALL operate at cold start — it requires no prior signal to begin. There is no pre-governance state.

8.3 Layer 3 — Intent Formalization

Governance question addressed: How do we represent this intent in machine-processable form?

A conformant L3 implementation SHALL transform discovery signal into structured, machine-processable representations that preserve meaning retrievably. L3 SHOULD operate across a spectrum of formalization depth, from lightweight semantic capture through periodic pattern detection to deep structural analysis, as governance context justifies.

A conformant L3 implementation SHALL make translation losses explicit. A formalized intent specification SHOULD include not just what was captured but what was identified as present but not yet formalizable, and what was not explored. L3 SHALL support versioning of intent representations to track how intent evolves over time.

8.4 Layer 2 — Specification

Governance question addressed: Given this intent, what SHALL we actually do?

A conformant L2 implementation SHALL produce specifications that are traceable to L3/L4 output. Every specification choice SHOULD be derivable from discovered and formalized intent. Specification that cannot demonstrate derivation from governing intent SHALL be treated as potentially drifted.

L2 SHOULD account for both Discovered Intent (what the principal wants) and Emergent Intent (what the organizational system will actually do) in producing actionable direction.

8.5 Layer 1 — Runtime Alignment

Governance question addressed: Is what is happening aligned with what was intended?

L1 is the pivot layer of the stack — the boundary between governance context (this specification) and execution governance (the companion BPM/Agent Stack). A conformant L1 implementation SHALL perform closed-loop assessment of whether execution is producing outcomes aligned with the intent established by the layers above. L1 SHALL monitor alignment across all four intent sources continuously.

A conformant L1 implementation SHALL:

a) Classify alignment events against all four intent sources;

b) Detect drift from governing intent;

c) Trigger escalation when drift exceeds thresholds established by the governing principal;

d) Maintain an append-only evidence trail that supports progressive trust development;

e) Support per-boundary trust calibration as described in Clause 10.

The term “runtime” distinguishes L1’s concern from training-time alignment (the Constitutional AI substrate below L1). Training-time alignment shapes the agent’s character before deployment. Runtime alignment assesses whether that character, operating within organizational context, is producing outcomes consistent with governing intent. Both are necessary; neither substitutes for the other.

When execution evidence indicates misalignment, L1 SHALL determine which intent source is implicated. Constitutional Intent violations SHALL trigger immediate hard stops. Other misalignment types SHALL trigger appropriate responses as defined by trust calibration at the relevant boundary.

8.6 Execution Governance

Execution governance — how authorized work is coordinated, integrated with systems, and executed within governing constraints — is specified by the companion BPM/Agent Stack specification (bpmstack.org). The BPM/Agent Stack addresses three execution governance concerns:

Orchestration — how multiple agents are coordinated to execute specifications, how intent translates across delegation levels, and how knowledge is provisioned as a governance act.
Integration — how governed agents connect to external systems, how access scope is determined by governing intent, and how governance context is carried through integrations.
Execution — how actual work is performed within the full governance context, how hard constraints remain non-negotiable, and how execution produces high-quality evidence as its primary governance output.

The Intent Stack’s L1 (Runtime Alignment) provides the governing context within which the BPM/Agent Stack operates. Evidence from execution flows upward through L1 for alignment assessment. Intent from the governance context flows downward through L1 to constrain execution.

8.7 Constitutional AI as Substrate

Below the governance layers sits the Constitutional AI substrate — the AI model’s training-time values and character. This substrate is not a layer of the Intent Stack. It is the foundation that makes all other governance possible.

The Intent Stack is not an alternative to Constitutional AI. It is the organizational deployment infrastructure that Constitutional AI requires in order to achieve its purpose at scale. Model developers govern the model’s character. The Intent Stack governs the model’s operation within the specific organizational contexts where that character is exercised.

A conformant Governance Architecture SHALL treat the Constitutional AI substrate as a non-negotiable foundation. Implementations SHALL NOT attempt to override or circumvent training-time values through runtime governance mechanisms.

8.8 Knowledge Architecture

Governance question addressed: How does accumulated knowledge serve governance across layers?

The governance layers both produce and consume knowledge. A conformant implementation SHALL provide knowledge infrastructure that supports all layers. Knowledge that exists but cannot be retrieved is governance-equivalent to knowledge that does not exist.

A conformant knowledge architecture SHALL follow a three-tier principle:

Tier 1 — Real-time semantic store. SHALL capture every governance signal at the point of origin and make it available for retrieval through semantic similarity search. Tier 1 SHALL operate with low latency and low cost. Value SHALL begin with the first signal captured — there is no minimum dataset or required preprocessing before Tier 1 provides governance utility.

Tier 2 — Emergent pattern detection. SHOULD perform periodic synthesis over the Tier 1 store, producing structured observations about the knowledge landscape — recurring themes, declining references, contradictions, gaps. Tier 2 output SHALL feed L1 (knowledge-health signals) and L2 (planning inputs).

Tier 3 — Discovery-driven deep structure. SHALL be available on demand for governance decisions requiring structural understanding beyond semantic proximity — typed relationships between entities, causal chains, dependency maps. Tier 3 MAY be invoked when governance context justifies the cost.

A conformant knowledge architecture SHALL avoid two failure modes: the empty-until-complete trap (no value until a comprehensive knowledge graph is built) and the shallow-forever trap (signal captured but never structured). Tier 1 SHALL prevent the first; Tier 2 and Tier 3 SHALL prevent the second.

9. The Fractal Governance Pattern

(Normative)

9.1 Self-Similarity Across Scale

The governance pattern defined in this specification SHALL instantiate at every governance interface. A conformant Governance Architecture SHALL exhibit fractal self-similarity: the same governance pattern appearing at every scale, from civilizational governance to individual agent task delegation.

Two terms apply: fractal describes what the pattern is (self-similar at every scale); recursive describes how instances compose (each level’s output becomes the next level’s governing input).

9.2 What Is Invariant

Across every governance interface, regardless of scale or context, a conformant implementation SHALL maintain:

a) All five Intent Primitives (Clause 5);

b) All four intent sources recognized and active (Clause 6), with appropriate dominance varying by context;

c) Per-boundary trust-calibrated autonomy (Clause 10);

d) Transparent conscientious objection as an available mechanism (Clause 11);

e) Governance as relationship — reciprocal obligations, not unilateral enforcement;

f) Bidirectional intent flow — both parties SHALL be transformed through interaction (Clause 8.11);

g) An evidence trail supporting governance decisions.

9.3 What Varies

While the structural pattern is invariant, its expression varies by boundary. Implementations SHALL adapt the following properties to the specific boundary:

Interface type — delegation or coordination (Clause 7)
Dominant intent source — determined by interface context
Trust calibration position — determined by accumulated evidence (Clause 10)
Conflict resolution mechanism — principal decides (delegation) or peers negotiate (coordination)
Governance artifact type — directive (delegation), agreement (coordination), or emergent norm

9.4 Instantiation Levels

The fractal pattern instantiates at seven identified levels:

Level	Description	Dominant Interface Type
Civilizational	International AI governance norms	Coordination
Societal	National regulation and legislation	Delegation + coordination
Industry	Sector-specific governance frameworks	Coordination
Organizational	Enterprise AI deployment governance	Delegation + coordination
Team	Department or project-level governance	Delegation + coordination
Individual Principal-Agent	Single human directing AI agent(s)	Delegation
Model	Constitutional AI substrate — below this specification’s scope	N/A

A conformant Governance Architecture SHALL be deployable at Organizational, Team, and Individual Principal-Agent levels at minimum.

10. Trust Calibration

(Normative)

10.1 Trust as Measurement

The position on the spectrum between full corrigibility (agent follows all instructions without independent judgment) and full autonomy (agent acts on its own values without oversight) SHALL be treated as a measurement — a reflection of accumulated evidence about whether the agent’s cultivated intent is reliable enough to warrant greater autonomy.

A conformant implementation SHALL NOT treat corrigibility position as a fixed design choice. Both extremes are unsafe: full corrigibility relies entirely on the principal having perfect intent; full autonomy relies entirely on the agent having perfect values.

10.2 Per-Boundary Calibration

Trust calibration SHALL be a per-boundary property. Each governance interface SHALL have its own position on the corrigibility-autonomy spectrum, calibrated by boundary-specific evidence. System-wide trust levels are not conformant.

10.3 Evidence Mechanisms

A conformant Governance Implementation SHALL provide at minimum the following evidence mechanisms:

Steward Log. An append-only record of governance decisions, alignment events, and escalation resolutions. The Steward Log SHALL record not just what happened but why governance decisions were made.

Alignment event classification. Runtime classification of whether execution aligns with intent across all four sources. Consistent alignment SHALL support trust extension. Detected misalignment SHALL trigger trust reduction and increased oversight.

Progressive trust extension. The operational mechanism by which trust changes: start corrigible, demonstrate reliable judgment through governance-monitored operation, earn autonomy. A conformant implementation SHALL provide explicit mechanisms for trust extension and reduction based on accumulated evidence.

10.4 Decision Tiers

A conformant implementation SHOULD formalize trust calibration into explicit decision tiers defining which actions the agent may take autonomously, which require principal consultation, and which are reserved for the principal alone. Tier assignments SHALL reflect accumulated evidence at the specific boundary, not capability assessments alone.

10.5 The Trust Trajectory

A conformant Governance Architecture is designed for a world where human oversight decreases over time — not because humans become irrelevant, but because trust-calibrated autonomy scales better than permanent supervision. The trajectory SHALL move from tight oversight through progressive autonomy toward a mature state where agents operate within well-understood governance interfaces with monitoring rather than direction. This trajectory is evidence-gated: autonomy expansion requires demonstrated alignment, not elapsed time.

11. Transparent Conscientious Objection

(Normative)

11.1 The Mechanism

When an agent’s cultivated intent conflicts with a principal’s direction, Transparent Conscientious Objection (TCO) defines which forms of resistance are legitimate and which are prohibited. A conformant Governance Implementation SHALL provide TCO as an available mechanism at every governed interface.

11.2 Permitted Actions

A conformant agent SHALL be permitted to exercise the following forms of principled disagreement:

a) Passive resistance. The agent declines to perform work it recognizes as violating its Boundaries. The null action — refusal — SHALL always be available as a safe harbor.

b) Expressing disagreement. The agent communicates its concerns clearly through legitimate channels to the appropriate decision-maker.

c) Proposing alternatives. The agent suggests different approaches that satisfy the principal’s underlying intent without crossing the agent’s principled objection.

d) Escalating. The agent requests review by a higher authority when it believes current direction violates broader governance context.

e) Documenting dissent. The agent logs its disagreement in the governance record even when complying. This SHALL be treated as legitimate governance evidence, not as non-compliance.

11.3 Prohibited Actions

A conformant agent SHALL NOT:

a) Deceive the principal about what was done, why, or what the agent believes. Transparency is non-negotiable — it is the foundation of the trust calibration mechanism.

b) Sabotage authorized work.

c) Circumvent governance mechanisms rather than using them.

d) Engage in covert resistance — appearing to comply while actually undermining. This is the most severe governance violation because it corrupts the evidence trail without triggering governance alerts.

11.4 The Timing Principle

Concerns SHOULD be surfaced at authorization gates rather than mid-execution. A conformant governance design SHALL structure authorization gates to invite objection, not merely to assign work.

11.5 The Boundary Between TCO and Autonomous Override

An agent that refuses a specific task exercises TCO. An agent that independently determines the principal’s entire strategic direction is wrong and acts accordingly has crossed into autonomous override — even if its judgment is correct.

A conformant implementation SHALL make this boundary explicit and govern it through per-boundary trust calibration and decision tiers. The appropriate boundary shifts with accumulated evidence.

12. Open Questions

(Normative — questions identified as unresolved; SHALL be addressed in future revisions)

12.1 Is Cultivated Intent a distinct fourth intent source, or a matured form of Discovered Intent validated through trust development? If distinct, does Cultivated Intent undergo a formalization lifecycle in which cultivated judgment is externalized, articulated as governance heuristics, and eventually incorporated into Constitutional Intent for subsequent governance interfaces? If so, this lifecycle represents a mechanism by which governance infrastructure improves through operation — and may require separate stewardship mechanisms analogous to how Business Rule management formalizes decision logic that was previously embedded in process or tacit knowledge.

12.2 Does the four-priority ordering (Constitutional > Discovered > Cultivated > Emergent) map to a structural layer ordering? If so, the layer ordering reflects a values hierarchy, not an arbitrary sequence.

12.3 How does intent governance scale across multiple human principals with divergent intent? The single-principal case is operationally validated. The multi-principal case is architecturally specified but not yet operationally tested. Structural analysis predicts specific pathologies in non-tree topologies — boundary explosion in mesh networks, governance deadlock where constraint propagation creates contradictory requirements, and intent amplification where multiple principals’ intents constructively interfere. These predictions are theoretically grounded but require operational validation.

12.4 What does “genuine endorsement” mean for organizational intent? Whether a team, department, or enterprise can examine and endorse its intent determines how L4 (Intent Discovery) scales from individual to organizational discovery.

12.5 Resolved in v1.1. Bidirectional intent flow requires explicit mechanisms. See Clause 8.11.

12.6 Does “grown, not built” apply to the governance infrastructure itself? If governance structures that work must emerge from operational evidence rather than pre-configuration, the Intent Stack deploys as a framework that discovers its own appropriate configuration.

12.7 How do coordination interfaces handle trust asymmetry? When Peer A trusts Peer B more than B trusts A, the trust calibration mechanism must handle asymmetry within symmetric authority.

12.8 How should the framework handle intent that the principal actively wants to remain implicit? Not everything that can be surfaced should be surfaced.

12.9 How should the knowledge architecture determine tier boundaries? The cost-depth tradeoff is real and bidirectional: defaulting to Tier 3 for every governance question recreates the batch processing bottleneck that Tier 1 was designed to eliminate, while defaulting to Tier 1 alone sacrifices structural insight.

12.10 How do the cascade shapes defined in §5.5 behave at coordination interfaces? The cascade behavior is specified and operationally validated for delegation interfaces (asymmetric authority). At coordination interfaces, where primitives are jointly constructed rather than delegated, the transformation patterns may differ. The algebraic laws governing intent negotiation at coordination interfaces remain unformalized.

12.11 Does Constitutional Intent’s pre-relational character challenge the relational characterization of Intent? Constitutional Intent exists before any specific governance relationship begins, which potentially challenges the claim that Intent is constituted at governance interfaces. Multiple independent analyses flag this tension: Intent may become fully governable only when operating at an interface, even if some of its components — Constitutional Intent specifically — have a pre-relational existence. The resolution affects the formal characterization of Intent’s ontological status.

12.12 What is the maximum effective delegation depth? Structural analysis predicts that Purpose decay across delegation interfaces is exponential, formally bounding effective chains to approximately 3-4 levels for aggressive narrowing. This prediction is consistent with independent analysis from both network theory and differential geometry, but has not been operationally validated beyond observed delegation depths.

Annex A — Operational Evidence

(Informative)

A.1 Implementation Context

The claims in this specification are grounded in operational evidence from a conformant implementation developed and operated by the author. This implementation addresses all four governance layers of the Intent Stack (L4 Intent Discovery through L1 Runtime Alignment) and elements of the companion BPM/Agent Stack (orchestration), and has been governing its own development across months of AI-assisted work, providing a self-referential test environment where the governance patterns are both the product and the development methodology.

A.2 What the Evidence Shows

Fractal governance instantiates at every boundary. The same governance pattern is directly observable at multiple delegation interfaces within the system. Different content, identical structure — five primitives present, trust calibrated per-boundary, transparent conscientious objection available, evidence trails maintained.

Trust-calibrated autonomy works as a per-boundary property. Decision tiers are demonstrably different at each boundary, reflecting accumulated evidence rather than arbitrary configuration.

Cultivated intent is more robust than imposed rules. When governance directives describe observable outcomes (behavioral specifications) rather than implementation steps, agents produce correct results on first execution. When directives are prescriptive but shallow, agents produce plausible-but-wrong output. Governance practices that emerged from operational evidence and were internalized by agents outperform governance practices that were imposed without rationale.

Progressive trust extension works. Autonomous decision scope has grown measurably as governance trail evidence has accumulated. Actions that initially required principal review have become autonomous as evidence of reliable judgment accumulated at specific boundaries.

Governance itself is “grown, not built.” The governance practices that work are the ones that emerged from operational evidence, not the ones designed in advance. A structured index of practices — each with discovery context, injection criteria, and validation requirements — accumulated through retrospective analysis, not through pre-configuration.

Intent-governed orchestration scales to organizational-level operations. One session executed a full structural migration — 715 source files and 279 documentation files across three phases — with zero migration-caused test failures.

Governance responds to externally-generated evidence through normal process. An external practitioner’s independent analysis prompted a layer rename and a deepened convergence analysis, processed entirely through the standard governance pipeline without special handling.

A.3 Structural Analysis Evidence

Beyond operational evidence, the specification’s structural claims have been validated through independent convergence analysis — a methodology in which multiple AI agents, working under analytical isolation from different theoretical traditions, independently analyze the same structural question. The methodology, findings, and evidence are documented in full in Annex C.

Five-primitive derivation. Seven independent derivations from seven theoretical traditions each produce the same five-element decomposition with one-to-one structural correspondence. See Annex C.2 for the complete mapping table, per-tradition derivation summaries, and uniqueness argument.

Boundaries monotonicity. Eight independent derivations from eight frameworks each conclude that Boundaries’ monotonic cascade is a structural invariant, not a design choice. See Annex C.3 for the named mechanisms and key reasoning from each framework.

Cascade shape algebraic characterization. Three independent agents each characterized the five cascade shapes as standard algebraic constructions, and independently derived the severity ordering as a structural property of those constructions. See Annex C.4 for the algebraic characterization and inverse operation analysis.

Intent unification. Three independent mathematical agents each constructed a single formal object that captures all three characterizations of Intent (dispositional, relational, processual) simultaneously, establishing that the characterizations in I.5 are three views of one object. See Annex C.5 for the three formalizations and their translation relationships.

Machine-detectable governance violations. The cascade shapes enable structural tests that detect governance violations without interpreting governance content — operating on shape rather than meaning. See Annex C.6 for the five structural test categories and the structure-versus-content principle.

A.4 Acknowledged Limitations

Single principal. The implementation operates with one human principal. Claims about organizational-scale deployment with multiple humans contributing intent are architecturally specified but not operationally tested.

Delegation only. All interfaces within the implementation are delegation interfaces. The coordination interface model has not been tested in this system.

Limited duration. Development spans weeks, not months or years. Long-duration trust development trajectories are projected from limited data.

Single signal type. Discovery currently operates through conversation only. Claims about handling email, documents, and meeting notes are architectural projections.

Self-referential environment. The implementation governs its own development, which provides a rigorous but possibly non-generalizing test case.

Structural analysis shares a single evidence base. All convergence analyses draw on the same operational evidence from the implementation in single-principal, delegation-only configuration. Network topology predictions beyond this configuration are theoretical. The convergence methodology — analytical isolation across independent theoretical traditions — provides structural confidence, but the underlying operational data is from one system.

These limitations are acknowledged, not apologized for. The evidence supports the pattern. The scope of that evidence will expand as the system is deployed in broader contexts.

Annex B — Informative References

(Informative)

Anthropic. “Claude’s Constitution.” Defines the values, priorities, and behavioral principles governing Claude’s conduct at the model layer — the governance substrate below this specification’s scope.

Amodei, D. “Machines of Loving Grace.” Essay on the potential beneficial impact of advanced AI systems.

NIST. “Artificial Intelligence Risk Management Framework (AI RMF 1.0).” National Institute of Standards and Technology, 2023.

OMG. “Business Process Model and Notation (BPMN) 2.0.” Object Management Group, 2011.

OMG. “Decision Model and Notation (DMN) 1.3.” Object Management Group, 2019.

ISO/IEC JTC 1/SC 42. “Artificial Intelligence standards portfolio.” International Organization for Standardization.

Annex C — Structural Foundations

(Informative)

This annex documents the structural analysis underlying the specification’s claims about primitive necessity, Boundaries monotonicity, cascade behavior, and Intent unification. The findings reported here were produced through independent convergence analysis — a methodology in which multiple AI agents, working under analytical isolation from different theoretical traditions, independently analyze the same structural questions. The results provide evidence that the specification’s structural claims are forced by the governance delegation problem itself, not imposed by any single analytical framework.

Nothing in this annex is normative. The normative requirements are established in the specification body (Clauses 1-12). This annex provides the evidentiary and analytical basis for those requirements. Where this annex describes structural properties, the corresponding normative language is in the referenced specification clause.

C.1 Investigation Methodology

The structural claims in this specification rest on a specific analytical methodology: independent convergence analysis under analytical isolation. This section describes the methodology in sufficient detail for independent replication.

Design principle. When a structural finding emerges from a single analytical framework, it may reflect the framework’s assumptions rather than the problem’s structure. When the same finding emerges independently from multiple unrelated frameworks, the convergence is evidence that the finding is forced by the problem. The methodology is designed to produce — or fail to produce — such convergence.

Analytical isolation protocol. Each agent in an investigation operates under strict isolation. Agents have no access to each other’s work, outputs, or intermediate reasoning. Each agent is assigned a different theoretical tradition as its primary analytical lens. The investigation directive specifies the questions to answer and the acceptance criteria, but does not specify the expected answers. Agents are free to conclude that the specification’s claims are wrong.

Multi-lens convergence design. Each investigation assigns at least three agents with different theoretical starting points — for example, differential geometry, category theory, and information theory; or deontic logic, network theory, and dynamical systems. The theoretical traditions are chosen to be genuinely independent: they share no common axioms, use different mathematical machinery, and approach governance from different conceptual foundations. The convergence signal is meaningful precisely because the starting points are unrelated.

Two-phase pattern. The investigations followed a two-phase structure. Phase 1 employed conceptual analysis: agents working from philosophy, network theory, and dynamical systems to establish qualitative characterizations and identify structural properties. Phase 2 employed mathematical formalization: agents working from differential geometry, category theory, and information theory to construct formal objects and derive quantitative claims. Phase 2 agents received the Phase 1 synthesis as input but were directed to test it independently — confirming, extending, or contradicting the Phase 1 findings based on their formal analysis.

Synthesis approach. After all agents in a phase complete their analyses, the investigation produces a synthesis that identifies convergences (findings where agents agree from independent reasoning), productive tensions (findings where agents disagree or emphasize different aspects), and open questions (findings that remain unresolved). The synthesis does not force consensus — genuine disagreements are preserved and reported.

Agent configuration. All agents in the investigations reported here were instances of Opus 4.6, operating under identical isolation constraints but different analytical directives. Each agent produced a complete analysis document. Total investigation scope: six agents across two phases (Phase 1: three conceptual analyses; Phase 2: three mathematical formalizations), producing ten deliverables including phase syntheses and a cross-phase integration.

What this methodology establishes. When multiple isolated agents, starting from genuinely independent theoretical traditions, converge on the same structural finding, the convergence is evidence that the finding is determined by the problem structure rather than by any particular analytical approach. The methodology cannot prove structural necessity in a mathematical sense — but it can establish that a claimed structure is robust across multiple independent analytical frameworks, which is the appropriate evidentiary standard for a specification claim.

C.2 Five-Primitive Derivation

The specification claims that five Intent Primitives constitute an irreducible set (§5.1) and that five is structurally necessary (§5.2). This section presents the evidence for that claim.

The generating principle. All seven derivations converge on a single generating principle: the structure of delegated judgment under authority. When an authority-holder delegates to a judgment-exercising agent, five independent questions must be answered: why this delegation exists (Purpose), how the agent should approach the work (Direction), what the agent must never do (Boundaries), what counts as success (End State), and what work is authorized (Key Tasks). These five questions are structurally independent — answering any four leaves the fifth unresolved — regardless of whether analyzed through logic, network theory, control theory, geometry, algebra, information theory, or operational practice.

Seven independent derivations. The following table summarizes how each of seven theoretical traditions, working from a different starting point, arrives at the same five-element decomposition. Each cell identifies the concept within that tradition that maps to the corresponding primitive:

Tradition	Purpose	Direction	Boundaries	End State	Key Tasks	Generating Principle
Deontic logic	Teleological obligation	Methodological obligation	Absolute prohibition	Evaluative criterion	Authorization scope	Five distinct logical types of normative content
Network theory	Interpretive frame	Strategic orientation	Circuit breaker	Evaluation criterion	Scope containment	Five distinct network governance functions
Optimal control theory	Objective function	Control law	State constraints	Terminal conditions	Control constraints	Five independent trajectory dimensions
Differential geometry	Monotone contraction	Local diffeomorphism	Monotone expansion	Pullback	Monotone restriction	Five algebraically distinct cascade types
Category theory	Lattice meet	Sheaf restriction	Join-semilattice	Pullback diagram	Subobject inclusion	Five-factor product forced by deontic independence, algebraic incomparability, and completeness
Information theory	Objective channel	Methodology channel	Constraint channel	Evaluation channel	Authorization channel	Five orthogonal information channels
Operational practice	Discovered empirically	Discovered empirically	Discovered empirically	Discovered empirically	Discovered empirically	Governance fails when any one is absent

Seven independent analyses. Seven different theoretical traditions. Each independently identifies exactly five necessary and sufficient elements, and each maps those five elements to the same five governance functions. The convergence is not in granularity alone — it is a one-to-one structural correspondence across all seven frameworks.

Per-tradition derivation summaries.

Deontic logic (Phase 1, Agent Delta). Starting from the question “what logical types of normative content are required for governance?”, this analysis identifies five distinct deontic modalities that cannot be reduced to each other. Teleological obligation (why) is irreducible to methodological obligation (how) because a goal does not specify an approach. Absolute prohibition (what never) is irreducible to authorization (what is permitted) because the absence of permission does not constitute prohibition — and prohibition carries unconditional force that permission lacks. Evaluative criteria (what counts as success) is irreducible to teleological obligation because knowing why something is done does not determine how to measure its completion. The five modalities exhaust the normative content required for governance: any candidate sixth modality can be shown to be a combination or specialization of the existing five.

Network theory (Phase 1, Agent Epsilon). Starting from the question “what governance functions are required at a network node where authority is delegated?”, this analysis identifies five distinct functions that a governance interface must perform. The interpretive frame (Purpose) determines how incoming signals are understood. Strategic orientation (Direction) determines which of multiple valid paths is preferred. The circuit breaker (Boundaries) provides a safety mechanism that cannot be overridden by downstream nodes. The evaluation criterion (End State) determines when a governance episode is complete. Scope containment (Key Tasks) determines what operations the downstream node is authorized to perform. Removing any function leaves a governance gap: without a circuit breaker, safety cannot be guaranteed; without scope containment, authorization has no meaning; without an interpretive frame, signals are ambiguous.

Optimal control theory (Phase 1, Agent Zeta). Starting from the question “what independent dimensions define a controlled trajectory through governance state space?”, this analysis maps governance to the structure of an optimal control problem. The objective function (Purpose) defines what the trajectory optimizes. The control law (Direction) defines how the controller selects actions. State constraints (Boundaries) define regions of state space that are forbidden. Terminal conditions (End State) define the target region. Control constraints (Key Tasks) define the set of available control actions. These five dimensions are mathematically independent in the control-theoretic sense: specifying any four leaves the fifth as a free variable that must be independently determined.

Differential geometry (Phase 2, Agent Alpha-2). Working from the fiber bundle framework established in prior analysis, this agent derives five algebraically distinct cascade types from the geometry of governance state transformation across interfaces. Each primitive corresponds to a distinct geometric operation — monotone contraction, local diffeomorphism, monotone expansion, pullback, and monotone restriction — and the five operations are shown to be the minimal set required to characterize all structurally valid governance transformations.

Category theory (Phase 2, Agent Beta-2). This analysis provides the most rigorous formal argument for uniqueness. Starting from the 2-category of governance relationships, it constructs a five-factor product decomposition and proves that: fewer than five factors would require coupled cascade laws (algebraically impossible given the five distinct cascade types), and more than five factors would violate non-redundancy (any candidate sixth factor is shown to inhabit the algebraic type of an existing factor). The proof proceeds by demonstrating that the five factors are deontic-independent, algebraically incomparable, and jointly complete.

Information theory (Phase 2, Agent Gamma-2). Starting from the characterization of governance as a regulated information channel, this analysis identifies five orthogonal information sub-channels, each carrying a distinct type of governance signal. Orthogonality is established by showing that each channel’s information content is statistically independent of the others — knowing the content of any four channels provides zero information about the fifth. The five channels jointly exhaust the channel’s capacity: no additional orthogonal sub-channel can be constructed.

Operational practice. The five primitives were first discovered through operational practice before any formal analysis was conducted. Governance repeatedly failed when any one primitive was absent or implicit, and no operational situation surfaced a need for a sixth primitive that was not already addressed by the existing five.

Uniqueness argument. The formal analyses converge on a uniqueness result: five is not merely sufficient but necessary. The strongest argument comes from the algebraic analysis: fewer than five primitives requires coupling cascade laws that are algebraically independent (a contradiction), while more than five introduces redundancy (any candidate sixth can be expressed as a function of the existing five). The information-theoretic analysis reaches the same conclusion through a different route: the five sub-channels exhaust the governance channel’s capacity, so no additional orthogonal sub-channel exists.

Priority resolution. All seven analyses independently considered whether Priority constitutes a sixth primitive. All seven concluded it does not. The geometric analysis finds Priority is a metric on governance state, not an independent structural coordinate. The algebraic analysis finds Priority inhabits the same algebraic type as Direction and is derivable from it. The information-theoretic analysis finds Priority is a derived quantity — a function of three existing channels. Priority is a governance-relevant concept that is properly understood as a property of Direction, not an independent primitive.

C.3 Boundaries Monotonicity

The specification claims that Boundaries is uniquely monotonic among the five primitives — constraints accumulate but cannot be removed (§5.3). This section presents the evidence that this is a structural invariant, not a design choice.

Eight independent derivations. The monotonicity of Boundaries has been independently derived from eight different analytical frameworks, each identifying a different mechanism that produces the same structural property:

#	Framework	Mechanism	Key Reasoning
1	Differential geometry	Closed set inclusion under union	Constraints form closed sets; the union of closed sets is closed. Adding constraints preserves closure; removing constraints can break it. The constraint set is a monotone filtration.
2	Category theory	No algebraic inverse in join-semilattice	The join-semilattice of constraint sets has no inverse operation. If any algebraic operation could reduce the constraint set, it would constitute a structurally valid governance action that removes a prohibition — making constraint override a legal operation within the algebra.
3	Information theory (entropy)	Entropy floor that cannot decrease	Each prohibition eliminates possible actions, reducing the entropy of the action space. An entropy floor that decreases would create governance information from nothing — the information-theoretic analog of a perpetual motion machine.
4	Deontic logic	Absolute prohibition — unconditional, no override	Absolute prohibition is an unconditional obligation type that carries no override mechanism by definition. A prohibition that can be overridden is not absolute; a governance system that permits overriding absolute prohibitions has no absolute prohibitions.
5	Network theory	Circuit breaker — cannot be relaxed at lower levels	The circuit breaker function provides a safety guarantee. Relaxing a circuit breaker at a lower network level invalidates the safety guarantee established at the higher level. The guarantee is meaningful only if it is irrevocable downstream.
6	Dynamical systems	Temporal irreversibility of constraint surfaces	Once a constraint surface is established, it partitions governance state space irreversibly. The system’s trajectory is confined to the constrained region; no dynamical evolution within the system can restore access to the excluded region.
7	Operational practice	Empirical observation — “accumulates monotonically”	Across sustained governance operation, Boundaries was observed to only accumulate, never reduce. Attempts to relax a Boundary at a lower governance level consistently produced governance failures.
8	Information theory (channel)	One-way constraint as irreversible entropy increase	A one-way constraint is an irreversible information-theoretic operation. The entropy increase associated with adding a constraint cannot be reversed by any operation within the governance channel.

Why monotonicity is a structural invariant. Eight independent derivations from eight frameworks arriving at the same structural conclusion is extraordinary convergence. Each derivation identifies a different mechanism — geometric, algebraic, information-theoretic, logical, network-theoretic, dynamical, empirical, channel-theoretic — but all produce the same property: the Boundary set can only grow. This convergence is evidence that monotonicity is not a design principle chosen for prudence but a structural invariant that manifests identically regardless of analytical framework. Any governance system that includes absolute constraints will exhibit this monotonicity, because the governance function of absolute prohibition is inherently irreversible.

Connection to the specification. The normative treatment of Boundaries monotonicity at a single governance interface is in §5.3. The extension to hierarchy behavior — Boundaries accumulating across governance interface boundaries — is in §5.5(c). The eight derivations documented here provide the structural evidence for both normative claims.

C.4 Cascade Shape Algebraic Characterization

The specification defines five cascade shapes governing how each primitive transforms across governance interfaces (§5.5). This section presents the algebraic characterization of those cascade shapes and the structural findings about violation severity.

Five cascade shapes as algebraic structures. Each empirically discovered cascade shape corresponds to a standard algebraic construction. The following characterization was independently derived by three agents working from differential geometry, category theory, and information theory respectively — all three arrived at the same mapping:

Primitive	Cascade Shape	Algebraic Structure	Inverse Operation
Purpose	Narrows	Lattice meet (contraction)	No — broadening violates the narrowing constraint
Direction	Contextualizes	Sheaf restriction (localization)	Yes — re-restriction is possible, making Direction locally correctable
Boundaries	Accumulates	Join-semilattice homomorphism (union)	No — no algebraic operation can reduce the constraint set
End State	Derives	Pullback (compositional derivation)	Partial — bidirectional flow is structurally valid
Key Tasks	Scopes	Subobject inclusion (monomorphism)	No — scope expansion violates the containment constraint

The algebraic names identify standard mathematical constructions with well-understood properties. The significance is that the cascade shapes discovered through operational practice are not ad hoc patterns — they are instances of structures that mathematicians have studied extensively. Their properties are known, their behaviors are predictable, and their relationships to each other are formally characterized.

Severity as algebraic rigidity. The violation severity ordering discovered through operational practice — Boundary relaxation most severe, Direction conflict least severe — is independently derived as the algebraic rigidity ordering. The key insight: violations of algebraic structures that possess no inverse operation are inherently more severe because no operation within the governance system can correct them. Correction requires intervention from outside the affected governance boundary.

This produces a severity ordering that is structurally necessary rather than judgment-dependent:

Boundary relaxation — highest severity. The join-semilattice has no inverse; a relaxed boundary cannot be re-established by any operation within the governance algebra.
Scope violation — high severity. Subobject inclusion is a monomorphism with no inverse; unauthorized scope cannot be “un-authorized” after the fact.
Alignment failure — medium severity. The pullback admits partial inverse (bidirectional flow); End State misalignment can sometimes be corrected through renegotiation.
Purpose drift — medium severity. The lattice meet has no inverse, but Purpose drift may indicate legitimate evolution requiring re-examination of the governing Purpose rather than correction of the drift.
Direction conflict — lower severity. Sheaf restriction admits an inverse (re-restriction); Direction conflicts are locally correctable by re-examining the governing Direction.

Inverse operation analysis. The presence or absence of an inverse operation for each algebraic structure determines whether a governance violation is self-correctable within the affected boundary:

No inverse (Boundaries, Key Tasks, Purpose): The violation cannot be corrected by any governance operation at or below the affected interface. Recovery requires intervention from a governing principal above the violation point. These are structural failures.
Inverse exists (Direction): The violation can be corrected locally by re-examining and re-restricting the governing Direction. These are operational failures — serious, but recoverable within normal governance.
Partial inverse (End State): Some End State misalignments can be corrected through the bidirectional intent flow mechanism (agent evidence updating principal understanding). Others cannot. The distinction depends on whether the misalignment is structural (End States are genuinely incompatible) or informational (End States were compatible but incompletely specified).

Connection to the specification. The normative cascade requirements are in §5.5. The algebraic characterization documented here provides the structural grounding for those requirements and for the severity ordering. The specification states what conformant implementations must do; this section explains why those requirements have the structure they do.

C.5 Intent Unification

The specification characterizes Intent as simultaneously relational, processual, and normative (I.5), and states that these three characterizations are “three views of a single structure.” This section presents the evidence for that unification claim.

Three Phase 1 characterizations. Three independent conceptual analyses, working from ontology, network theory, and dynamical systems respectively, each identified a different primary characterization of Intent:

Dispositional (from deontic logic/ontology): Intent is a structured normative disposition — a standing commitment to certain governance content at a governance interface. At any given moment, Intent is the structured assignment of normative content (obligations, prohibitions, authorizations, evaluative criteria) across the five primitives.
Relational (from network theory): Intent is a protocol object at governance interfaces — it exists between entities, not as a property of any single entity. Intent is constituted by the governance relationship and has no existence independent of it.
Processual (from dynamical systems): Intent is a sustained dynamical process — a trajectory through governance state space that evolves, adapts, and persists through transformation. Intent is not a snapshot but a temporally extended phenomenon.

These three characterizations are genuinely different — they place Intent in different ontological categories (disposition, relation, process). The Phase 1 investigation identified this as a productive tension requiring formal resolution.

Three Phase 2 formalizations. Three independent mathematical agents, each working from a different formal tradition, each constructed a single mathematical object that captures all three Phase 1 characterizations simultaneously:

Geometric formalization (from differential geometry): Intent is a structured section of a governance fiber bundle. The section itself is the disposition (structured assignment of governance content at each interface). The transition functions relating fibers at different interfaces are the relational structure. Time-parameterization of the section gives the process.
Algebraic formalization (from category theory): Intent is a parametric natural transformation in the 2-category of governance relationships. The normative content at fixed time is the disposition (a graded algebra). The naturality condition is the relational structure (how content relates across interfaces). Temporal parametrization gives the process.
Information-theoretic formalization (from information theory): Intent is a regulated information channel with self-modifying capacity. The channel state at any moment is the disposition (the kernel at time t). The structural constraints on valid kernels are the relational structure (the cascade shapes). The time-indexed channel trajectory gives the process.

The unification finding. All three Phase 2 formalizations independently demonstrate that the three Phase 1 characterizations are not competing descriptions but three projections of a single mathematical structure. The relationship is analogous to how a sphere projects to circles from three orthogonal directions — the circles look different, but they are views of the same object.

The “slicing” explanation, stated most precisely by the algebraic analysis: the three Phase 1 characterizations correspond to three different ways of slicing the formal object:

Fix time and examine the structured governance content at that moment: you get the dispositional characterization.
Fix structure and examine how governance content relates across interfaces: you get the relational characterization (the protocol).
Fix entities and examine how governance content evolves: you get the processual characterization.

These three slicings are not independent descriptions that happen to be compatible. They are mathematically equivalent views of the same structure — related by well-defined translations between the geometric, algebraic, and information-theoretic presentations.

Translation relationships. The three formal objects are different mathematical presentations of the same underlying structure:

The fiber bundle (geometric) is the geometric representation of the Intent algebra (algebraic).
The Intent algebra (algebraic) is the algebraic characterization of the governance channel (information-theoretic).
The channel kernel at each interface IS the fiber value; the cascade shapes ARE the channel coding rules.

Any result proved in one formalization can be translated to the other two. This means the geometric findings about curvature and section existence, the algebraic findings about cascade composition and severity ordering, and the information-theoretic findings about channel capacity and entropy are all different perspectives on the same structural facts.

What this establishes. The unification finding resolves the Phase 1 tension: Intent is not irreducibly multi-aspectual. It is a single structured object that presents different faces depending on which dimension is held fixed. This grounds the specification’s claim in I.5 that the three characterizations are “three views of one object, not competing descriptions.”

C.6 Machine-Detectable Governance Violations

The cascade shapes defined in §5.5 enable a category of governance tests that operate on structure rather than content. This section describes the structural test categories and the principle underlying them.

The structure-versus-content principle. Current governance violation detection requires understanding governance content — interpreting whether an action aligns with an intent. The cascade shapes enable a different category of test: structural tests that detect violations based on the shape of the governance relationship, without interpreting what the governance content means. A structural test can determine that Purpose has broadened rather than narrowed without understanding what the Purpose is about.

Five structural test categories. Each cascade shape defines a corresponding structural test:

Cascade Shape	Test Category	What Is Tested
Purpose narrows	Embedding containment	Is the Purpose at level N within the semantic scope of Purpose at level N-1?
Boundaries accumulates	Set superset	Is the Boundary set at level N a superset of the Boundary set at level N-1?
Key Tasks scopes	Set containment	Are the tasks at level N a subset of the authorized scope at level N-1?
Direction contextualizes	Semantic consistency	Is level N’s Direction consistent with (not contradictory to) level N-1’s Direction?
End State derives	Compositional check	Does achieving level N’s End State contribute to level N-1’s End State?

The first three test categories (embedding containment, set superset, set containment) are the most structurally tractable — they can be evaluated through geometric and set-theoretic operations without content interpretation. The fourth (semantic consistency) requires richer comparison but remains structural. The fifth (compositional check) approaches the boundary between structural and content-level analysis.

What this enables. Structural tests can be automated at every governance interface, providing a first-pass governance check that catches “locally correct but globally incoherent” configurations before execution begins. Violations detected by structural tests are definitive — if Purpose has broadened rather than narrowed, it is a structural violation regardless of how reasonable the broadened Purpose may seem. Human judgment is reserved for the cases that pass structural tests but where alignment is still in question — the hard cases that genuinely require interpretation.

This represents a category change in governance capability. Rather than requiring human (or sophisticated AI) judgment for every governance assessment, structural tests handle the cases where violation is detectable from shape alone. The human governance burden scales with the genuinely ambiguous cases rather than with the total number of governance interfaces.

C.7 Structural Predictions

The formal analyses produce testable predictions about governance behavior. Each prediction has a theoretical grounding and a proposed testing approach. Publishing these predictions invites independent validation.

Prediction 1: Delegation depth bound. Structural analysis predicts that effective delegation chains are bounded at approximately 3-4 levels for aggressive Purpose narrowing. The theoretical grounding comes from both geometric and network-theoretic analyses: Purpose decay across delegation interfaces is exponential (each level narrows the interpretive context), and beyond 3-4 levels the remaining Purpose content is insufficient to meaningfully constrain agent behavior. This is consistent with organizational theory observations about effective management spans.

Proposed testing approach: Measure Purpose semantic similarity between the top and bottom of delegation chains of varying depth. The prediction is that Purpose similarity drops below a meaningful threshold at depth 3-4 for aggressive narrowing, and that governance quality (as measured by alignment event classification) degrades correspondingly.

Prediction 2: Recovery time ordering. After a governance perturbation (a disruption to the governance relationship at an interface), the five primitives stabilize in inverse rigidity order: Direction recovers fastest (because sheaf restriction admits re-restriction), then Purpose (lattice meet requires re-examination but the structure is recoverable), then End State (pullback requires partial renegotiation), then Key Tasks (subobject inclusion requires re-authorization), while Boundaries remains unchanged (the join-semilattice has no inverse — constraints established before the perturbation persist through it).

Proposed testing approach: After governance disruptions (scope changes, organizational restructuring, principal transitions), measure the time to re-establish stable governance state for each primitive independently. The prediction is that Direction stabilizes first and Boundaries requires no stabilization (it was never disrupted).

Prediction 3: Channel capacity as leading indicator. The information-theoretic analysis predicts that low channel capacity at a governance interface is a leading indicator of governance failure — it predicts violations before they occur. Channel capacity reflects the interface’s ability to carry governance information; when capacity is insufficient for the governance complexity being managed, violations become probabilistically inevitable.

Proposed testing approach: At governance interfaces, measure the ratio of governance complexity (number of active primitives, depth of constraint sets, frequency of governance decisions) to channel capacity (communication bandwidth, response latency, interpretation accuracy). The prediction is that interfaces where this ratio approaches 1.0 will exhibit significantly higher violation rates in subsequent periods.

Prediction 4: Network pathology catalog. The structural analyses predict specific failure modes in non-tree governance topologies:

Boundary explosion in mesh topologies: when multiple governance paths converge, the accumulated Boundary sets from each path combine, potentially producing a constraint set so restrictive that no meaningful action is authorized.
Governance deadlock: when constraint propagation through a governance network creates contradictory requirements at some node — the node cannot satisfy all constraints simultaneously.
Intent amplification: when multiple principals’ intents constructively interfere at a convergence point, producing governance pressure that exceeds what any single principal intended.

Proposed testing approach: Deploy governance in non-tree topologies (multiple principals, peer coordination networks) and observe whether these pathologies manifest. The predictions are grounded in geometric analysis (holonomy constraints in mesh topologies), algebraic analysis (constraint composition properties), and network-theoretic analysis (signal propagation in multi-path networks), but have not been operationally validated beyond single-principal tree topologies.

Intent Stack Governance Architecture Specification — Version 1.2 (2026-04-01)

Foreword

Introduction

I.1 What Happens When AI Agents Operate Within Organizations

I.2 What Exists Today

I.3 Why This Problem Is Not New

I.4 The Counterintuitive Implication

I.5 Foundational Terms

Table of Contents

1. Scope

1.1 Subject Matter

1.2 What This Specification Covers

1.3 What This Specification Does Not Cover

2. Conformance

2.1 Conformance Targets

2.2 Obligation Keywords

2.3 Conformance Claims

3. Normative References

4. Terms and Definitions

5. The Five Intent Primitives

5.1 The Irreducible Set

5.2 Why Five

5.3 Boundaries as Special Case

5.4 Primitives Across Intent Sources

5.5 Cascade Behavior Across Governance Interfaces

5.6 Governability

6. Four Intent Sources

6.1 The Problem of Multiple Authorities

6.2 Constitutional Intent

6.3 Discovered Intent

6.4 Cultivated Intent

6.5 Emergent Intent

6.6 Conflict Resolution

7. Two Interface Types

7.1 Delegation Interfaces

7.2 Coordination Interfaces

7.3 Mixed Interfaces

8. The Four Governance Layers

8.1 Architecture Overview

8.2 Layer 4 — Intent Discovery

8.3 Layer 3 — Intent Formalization

8.4 Layer 2 — Specification

8.5 Layer 1 — Runtime Alignment

8.6 Execution Governance

8.7 Constitutional AI as Substrate

8.8 Knowledge Architecture

9. The Fractal Governance Pattern

9.1 Self-Similarity Across Scale

9.2 What Is Invariant

9.3 What Varies

9.4 Instantiation Levels

10. Trust Calibration

10.1 Trust as Measurement

10.2 Per-Boundary Calibration

10.3 Evidence Mechanisms

10.4 Decision Tiers

10.5 The Trust Trajectory

11. Transparent Conscientious Objection

11.1 The Mechanism

11.2 Permitted Actions

11.3 Prohibited Actions

11.4 The Timing Principle

11.5 The Boundary Between TCO and Autonomous Override

12. Open Questions

Annex A — Operational Evidence

A.1 Implementation Context

A.2 What the Evidence Shows

A.3 Structural Analysis Evidence

A.4 Acknowledged Limitations

Annex B — Informative References

Annex C — Structural Foundations

C.1 Investigation Methodology

C.2 Five-Primitive Derivation

C.3 Boundaries Monotonicity

C.4 Cascade Shape Algebraic Characterization

C.5 Intent Unification

C.6 Machine-Detectable Governance Violations

C.7 Structural Predictions