First Edition · 2026

The Trust
Stack

A Framework for Building Verifiable Trust in Enterprise AI

ARJUN JAGGI
ADITYA KARNAM GURURAJ RAO
arjunjaggi.com  ·  adityakarnam.com
THE TRUST STACK TABLE OF CONTENTS
Front Matter
Executive Summary
ES
Front Matter
Foreword: The Problem With Trust
FW
Chapter 01
The Trust Deficit
01
Chapter 02
The Trust Stack Framework
02
Chapter 03
Behavioral Contracts
03
Chapter 04
Provenance Chain
04
Chapter 05
Access Controls
05
Chapter 06
Audit Trails
06
Chapter 07
Board-Level Disclosure
07
Chapter 08
Trust Maturity Model
08
Chapter 09
Deployment Playbook
09
Chapter 10
Trust Under Pressure
10
Executive Summary The Trust Stack

Six Findings for Enterprise Leaders

Enterprise AI deployments do not fail because the models are wrong. They fail because the trust architecture around those models is absent. The Trust Stack is a five-layer framework that makes AI behavior verifiable, auditable, and boardroom-ready.

01
The Trust Deficit Is Structural

In practitioner observation across enterprise deployments, the structural pattern is consistent: most organizations operate without formal behavioral constraints, traceable output provenance, or structured audit capability. This is not a model quality problem. It is an architecture gap that existing vendor frameworks and regulatory standards do not fill.

02
Five Layers, One Architecture

The Trust Stack defines five interdependent layers: Behavioral Contract, Provenance Chain, Access Controls, Audit Trail, and Board-Level Disclosure. Every layer depends on the one below it. Implementing three of five layers provides partial protection. Full stack coverage is required for verifiable trust.

03
Behavioral Contracts Are the Foundation

A Behavioral Contract is a formal specification of what an AI model is permitted and prohibited from doing. Without it, every downstream control, including access policies and audit logs, is built on undefined behavior. The contract is the first layer precisely because it defines the reference state against which everything else is measured.

04
Provenance Is the Accountability Mechanism

A Provenance Chain traces every AI output back to its source document, retrieval decision, and model invocation. Without it, a disputed AI decision cannot be investigated. With it, every output becomes a defensible artifact. Provenance is the layer that makes AI behavior auditable rather than merely logged.

05
Regulatory Alignment Is a Byproduct, Not a Goal

Organizations that implement the full Trust Stack align with NIST AI RMF 1.0, ISO/IEC 42001:2023, and the EU AI Act as a byproduct of good architecture. The EU AI Act carries penalty tiers of up to 7% of global annual turnover for prohibited practice violations and up to 3% for provider and deployer obligation failures. Compliance is not the mission. Verifiable trust is.

06
Trust Attestation Closes the Board Gap

Trust Attestation is a formal declaration that an AI system operated within defined constraints during a specified period, signed by a responsible party and suitable for regulatory review. It is the mechanism that converts an internal audit trail into a boardroom-ready governance artifact. Without attestation, even a fully implemented Trust Stack remains invisible to the board.

Foreword The Trust Stack

The Problem With Trust

Every enterprise AI program reaches the same inflection point. The pilot works. The model performs well in evaluation. The business case is approved. And then, somewhere between the pilot and broad deployment, trust collapses.

Not because the model is wrong. Because nobody can explain what the model did, why it did it, or what would prevent it from doing something different next week. A legal team asks whether the output is defensible in a dispute. The answer is: we don't know. A regulator asks whether the system operates within defined constraints. The answer is: we believe so. An auditor asks for the decision log. The log exists, but it does not contain what the auditor needs.

This is the trust deficit. It is not a model problem. It is an architecture problem. And it is almost universal in enterprise AI deployments today.

This book introduces the Trust Stack, a five-layer architecture for building verifiable trust in enterprise AI. It is not a compliance framework. It is not a vendor checklist. It is a set of formal constructs, each with a precise definition and a clear relationship to the layers above and below it, that together make AI behavior verifiable at the level required for board-level governance.

The Trust Stack does not ask you to trust your AI system. It gives you the architecture to verify it.

The four coined terms in this book, Trust Stack, Behavioral Contract, Provenance Chain, and Trust Attestation, are introduced formally and with precision. They are designed to survive the journey from this page to a board agenda, a regulatory submission, and a practitioner's implementation checklist without losing meaning.

We wrote this book because the constructs did not exist in the form enterprises needed. NIST AI RMF 1.0 [1] provides a risk management vocabulary. ISO/IEC 42001:2023 [2] provides a management system standard. The EU AI Act [5] provides obligations. None of them provide the five-layer architecture that connects behavioral specification to board-level disclosure in a single traceable chain. The Trust Stack fills that gap.

Arjun Jaggi  ·  Aditya Karnam Gururaj Rao  ·  August 2026

1
Chapter 01

The Trust Deficit

Enterprise AI deployments are failing not because models are unreliable, but because the architecture required to verify their behavior has never been built. This chapter defines the structural gap and the cost of leaving it unfilled.

Pages cover the structural trust gap, regulatory pressure, and the case for a formal framework.

Key Takeaways

  • The trust deficit is not a model quality issue. It is an architecture gap present across enterprise AI deployments in practitioner observation — and one that existing standards do not resolve.
  • Existing standards address risk management and management systems but do not specify how to make AI behavior verifiable at the output level.
  • Regulatory frameworks, including the EU AI Act [5], impose obligations that assume a trust architecture already exists. In practitioner observation, the architecture to satisfy those obligations is absent in most enterprise deployments.
  • The cost of the trust deficit is realized in three forms: disputed outputs with no audit trail, regulatory exposure without defensible documentation, and board-level inability to attest to system behavior.

Leadership Questions

  1. If a significant AI-assisted decision were disputed today, could you produce a complete audit trail within 24 hours?
  2. Has your organization formally defined what your AI systems are permitted and prohibited from doing?
  3. Can your board attest to the behavior of your AI systems in a regulatory submission?
Chapter Equation
T(system) = f(L₁, L₂, L₃, L₄, L₅) where L1..L5 are the five Trust Stack layers
Chapter 01 The Trust Deficit

When Pilots Work and Deployments Fail

The pattern repeats across industries. A financial services firm deploys a document summarization agent. Accuracy in evaluation is high. User satisfaction in the pilot is strong. The system reaches broad deployment, and six months later a compliance officer flags a summary that omitted a material disclosure. Nobody can reconstruct what the model received as input, what context it retrieved, or why it produced the output it did. The system is suspended pending investigation.

This is not a model failure. The model may have behaved exactly as designed. It is an architecture failure: the system was deployed without the controls required to make its behavior verifiable after the fact.

The trust deficit has three structural causes. First, behavioral specification is absent. In practitioner observation, most enterprise AI deployments define acceptable behavior through informal prompts and evaluation rubrics, not formal specifications. When behavior changes due to model updates, context drift, or edge cases, there is no reference state against which to measure the change.

Second, output provenance is untraced. AI systems that retrieve from knowledge bases, databases, or document stores make decisions about what to retrieve and how to weight it. Those retrieval decisions shape the output. Without a provenance trace, a disputed output cannot be traced to its source, and the retrieval decision cannot be evaluated.

Architecture Gap

An audit log that records inputs and outputs without recording retrieval decisions, model version, context construction, and the behavioral specification in effect at the time of the call does not constitute a verifiable audit trail. It constitutes a timestamp.

Third, attestation capability is missing. Boards and regulators increasingly require formal declarations about how AI systems behave. Without an attestation layer, the organization cannot produce those declarations without a bespoke investigation for each request.

The Regulatory Pressure Point

The EU AI Act [5] imposes tiered obligations on providers and deployers of AI systems classified as high-risk. The penalty structure has two tiers: up to 7% of global annual turnover for violations of prohibited practice provisions, and up to 3% for failures of provider and deployer obligations. These obligations include logging requirements, human oversight provisions, and accuracy and robustness standards that presuppose a functioning trust architecture.

NIST AI RMF 1.0 [1] provides a risk management vocabulary organized around Govern, Map, Measure, and Manage functions. ISO/IEC 42001:2023 [2] provides a management system standard. Both are valuable. Neither specifies the five-layer architecture that connects behavioral specification to board-level disclosure in a single traceable chain. That specification is what this book provides.

Scenario

A healthcare organization receives an audit request covering AI-assisted triage decisions over the prior 18 months. The audit log exists and contains input and output records. It does not contain the retrieval context, the model version active at each decision point, or the behavioral constraints that were in effect. Reconstructing the full decision context for a sample of cases takes 11 weeks and requires manual review of infrastructure logs, model deployment records, and prompt version histories stored in three separate systems. A Trust Stack with provenance and audit trail layers would have made this a one-hour query.

The research literature on AI alignment makes the structural nature of this problem clear. Constitutional AI approaches [3] establish behavioral rules at training time. RLHF approaches [4] shape model behavior through preference signals. Both operate at the model level. Neither substitutes for runtime behavioral contracts and audit infrastructure at the deployment level. A well-aligned model deployed without a Trust Stack is still a model whose runtime behavior cannot be verified.

Chapter Takeaways

  • The trust deficit is structural and architectural, not a model quality issue
  • Three causes: absent behavioral specification, untraced provenance, missing attestation
  • Regulatory obligations (EU AI Act, NIST AI RMF) presuppose a trust architecture that most enterprises have not built

Questions for Your Team

  1. Which of your AI systems operate under a formal behavioral specification today?
  2. For your highest-stakes AI deployment, what would a regulatory audit of that system require you to produce?
Chapter 01 The Trust Deficit

What "Verifiable" Means in Practice

Verifiable AI behavior means three things. An output can be traced to its inputs, its retrieval context, and the model that produced it. A behavioral constraint can be confirmed to have been in effect at the time of the output. A responsible party can formally attest that the system operated within defined constraints during a specified period.

In practitioner observation, most enterprise AI deployments satisfy none of these three conditions. Some satisfy one. Very few satisfy two. The goal of the Trust Stack is to make satisfying all three conditions the default outcome of a well-architected deployment, not an exceptional achievement requiring custom engineering.

Verifiability is not the same as explainability. Explainability asks why a model made a particular prediction at the level of internal model mechanics. Verifiability asks whether the system operated within its defined behavioral constraints and whether that can be demonstrated. Verifiability does not require mechanistic transparency into the model. It requires architectural transparency into the deployment.

Framework Reference

The HELM evaluation framework [6] demonstrates that model behavior varies substantially across deployment conditions, including context length, prompt formulation, and retrieval strategy. This variation is expected and not inherently problematic. What is problematic is deploying without the architecture to detect when variation moves outside acceptable bounds. That detection capability is what the Trust Stack provides.

The Cost Structure of the Trust Deficit

The cost of an absent trust architecture is not primarily regulatory. Regulatory penalties, while substantial under the EU AI Act [5], are tail events. The primary cost is operational: the inability to scale AI deployment because disputed outputs cannot be resolved, the inability to extend AI to higher-stakes decisions because the governance required to authorize them does not exist, and the organizational friction of building trust architecture retroactively after a trust incident rather than proactively before one.

Retroactive trust architecture construction is substantially more expensive than proactive implementation. When a trust incident occurs, the organization must simultaneously investigate the incident, implement controls, and demonstrate to regulators or counterparties that the controls are now adequate. These activities compete for the same engineering, legal, and leadership resources. Proactive Trust Stack implementation eliminates the investigation cost and separates the implementation timeline from crisis conditions.

Red-teaming research [7] demonstrates that AI systems exhibit unexpected behaviors under adversarial and edge-case conditions that are difficult to predict from standard evaluation. The appropriate response is not to avoid deployment but to deploy with the architectural controls that make unexpected behavior detectable and recoverable. The Trust Stack is that architecture.

Decision Framework: When to Prioritize the Trust Stack

Prioritize full Trust Stack implementation when any of the following conditions apply:

  • The AI system influences decisions that are subject to regulatory review, legal challenge, or contractual obligation
  • The AI system processes personal data, confidential information, or proprietary content
  • The AI system operates with any degree of autonomy, including retrieval, tool use, or multi-step reasoning
  • The organization intends to scale the AI system beyond a single team or use case
  • Board or executive attestation to AI system behavior may be required within 24 months
2
Chapter 02

The Trust Stack Framework

Five layers. One architecture. The Trust Stack defines the complete structure for verifiable trust in enterprise AI, from behavioral specification at the base to board-level disclosure at the top.

Pages introduce all five layers, their relationships, and the formal system definition.

Key Takeaways

  • The Trust Stack is a five-layer architecture: Behavioral Contract (L1), Provenance Chain (L2), Access Controls (L3), Audit Trail (L4), and Board-Level Disclosure (L5).
  • Layers are interdependent: each layer requires the one below it to function. L3 without L1 enforces rules against an undefined reference state. L4 without L2 logs events without context. L5 without L4 produces attestations without evidence.
  • The Trust Stack does not replace model evaluation or alignment techniques. It is the deployment-layer architecture that makes those techniques verifiable at runtime.
  • Full stack coverage is required for verifiable trust. Partial implementations reduce risk but do not provide the audit and attestation capability that regulatory and board contexts require.

Leadership Questions

  1. Which of the five layers does your organization have any implementation of today, even informally?
  2. Which layer represents the largest gap between your current state and what board-level attestation would require?
  3. What would it take to implement L1 (Behavioral Contract) for your three highest-stakes AI deployments within 90 days?
Trust Stack Definition
TS = L₁ ∩ L₂ ∩ L₃ ∩ L₄ ∩ L₅ intersection: all five layers must be present for verifiable trust
Chapter 02 The Trust Stack Framework

Five Layers, One Architecture

The Trust Stack is a five-layer architecture for verifiable trust in enterprise AI. Each layer addresses a distinct dimension of trustworthiness, and each layer depends on the one below it. The layers are not independent controls. They are an integrated architecture in which the failure of any layer propagates upward and undermines the layers above it.

Definition Trust Stack

The five-layer architecture for verifiable trust in enterprise AI, comprising: L1 Behavioral Contract, L2 Provenance Chain, L3 Access Controls, L4 Audit Trail, and L5 Board-Level Disclosure. The Trust Stack is complete only when all five layers are implemented and each layer references the one below it. Partial implementations reduce risk but do not constitute a verifiable trust architecture.

Layer 1, the Behavioral Contract, is the foundation. It specifies what the AI system is permitted to do, what it is prohibited from doing, and what conditions trigger escalation to a human. Without this specification, every downstream control operates against an undefined reference state.

Layer 2, the Provenance Chain, traces every output to its source. It records the source document, retrieval decision, relevance score, and passage reference associated with each output. Without provenance, an output cannot be investigated. With it, every output is a defensible artifact.

Layer 3, Access Controls, governs who and what can invoke the AI system and with what scope. In agentic deployments, access controls define the boundary of each agent's authority. Without access controls grounded in the Behavioral Contract (L1), agents can acquire or exercise capabilities beyond their defined scope.

100% 80% 60% 40% 20% Reactive Defined Managed Verified L1 L2 L3 L4 L5 Trust Stack layer implementation across maturity tiers. Directional illustration. Values not derived from systematic survey data.

Layer 4, the Audit Trail, records every significant event in the AI system's operation in a form sufficient for post-hoc investigation. A sufficient audit trail includes: the input, the output, the retrieval context (referencing L2), the model identifier, the timestamp, and the behavioral constraints in effect (referencing L1). A log that omits any of these elements is not a sufficient audit trail for regulatory or legal purposes.

Layer 5, Board-Level Disclosure, is the interface between the trust architecture and the governance layer. It comprises the reporting, attestation, and disclosure mechanisms that make the Trust Stack visible to boards, regulators, and counterparties. Without L5, even a fully implemented L1-L4 remains invisible at the governance level.

Implementation Sequence

Implement layers in order: L1 before L2, L2 before L3, L3 before L4, L4 before L5. Each layer provides the reference state that the next layer requires. Implementing L4 (Audit Trail) before L1 (Behavioral Contract) produces logs with no reference against which to evaluate the events they record.

Chapter 02 The Trust Stack Framework

Why Partial Implementation Fails

Organizations frequently implement portions of the Trust Stack without recognizing the interdependence between layers. A common pattern is L3 and L4 without L1 and L2: access controls and audit logs are present, but without a Behavioral Contract to define permitted behavior and a Provenance Chain to trace output origins, the access controls enforce undefined rules and the audit logs record events without the context required to evaluate them.

A second common pattern is L1 and L4 without L2 and L3: a behavioral specification exists and events are logged, but output provenance is untraced and access is uncontrolled. In this configuration, the audit log records that an output was produced but cannot reconstruct the retrieval context that shaped it, and the behavioral contract defines permitted behavior but nothing prevents an agent from acquiring access beyond its defined scope.

The Trust Stack formal definition, TS = L1 intersection L2 intersection L3 intersection L4 intersection L5, captures this interdependence precisely. The intersection operator denotes that all five layers must be present. A system with four layers is not a Trust Stack with one layer missing. It is a system that lacks verifiable trust.

Standards Alignment

NIST AI RMF 1.0 [1] Govern function maps to L1 and L5. Map and Measure functions map to L2 and L4. Manage function maps to L3 and L4. ISO/IEC 42001:2023 [2] Annex A controls map across all five layers. The Trust Stack does not replace these standards. It provides the architectural specification that makes implementing them concrete.

The Trust Stack and Agentic AI

Agentic AI systems, those that take multi-step actions, use tools, retrieve from external sources, and operate with varying degrees of autonomy, make the Trust Stack more important, not less. Each additional degree of autonomy expands the behavioral surface that the Behavioral Contract must specify, the provenance surface that the Provenance Chain must trace, and the access surface that Access Controls must govern.

Research on AI behavior in constrained conditions [7] demonstrates that unexpected outputs under edge cases are common even in well-evaluated systems. The appropriate response is not to constrain AI to simple tasks but to deploy more sophisticated tasks with the Trust Stack architecture that makes unexpected behavior detectable and recoverable.

Our prior work on domain-specific fine-tuned models [8] demonstrates that behavioral specialization at the model level improves task performance. Behavioral specialization at the deployment level, through formal Behavioral Contracts, improves verifiability. Both are required for high-stakes enterprise deployment. Neither substitutes for the other.

Chapter Takeaways

  • The Trust Stack is a five-layer architecture: L1 through L5, each depending on the layer below
  • Partial implementation reduces risk but does not produce a verifiable trust architecture
  • Agentic AI expands all Trust Stack dimensions and makes the architecture more critical
  • Implementation sequence matters: L1 before L2, L2 before L3, and so on

Questions for Your Team

  1. Map your three highest-stakes AI deployments against the five layers. Which layers are absent?
  2. For each absent layer, what is the minimum implementation required to close the gap within a quarter?
3
Chapter 03

Behavioral Contracts

The Behavioral Contract is the foundation of the Trust Stack. It specifies permitted behavior, prohibited behavior, and escalation conditions, providing the reference state against which all other controls operate.

Pages cover the formal definition, structure, and implementation of Behavioral Contracts.

Key Takeaways

  • A Behavioral Contract is a formal specification of permitted actions, prohibited actions, and escalation conditions for a specific AI deployment context.
  • Behavioral Contracts operate at the deployment level, not the model level. A well-aligned model still requires a deployment-level Behavioral Contract.
  • The contract must be version-controlled and referenced in the audit trail. A contract that cannot be linked to a specific version at a specific point in time does not provide the temporal reference required for audit.
  • Constitutional AI [3] and RLHF [4] approaches operate at training time. Behavioral Contracts operate at deployment time. Both are required; neither substitutes for the other.

Leadership Questions

  1. For your highest-stakes AI deployment, does a written document exist that specifies what the system is prohibited from doing?
  2. Is that document version-controlled? Is it referenced in your audit infrastructure?
  3. Who owns the Behavioral Contract for each of your AI systems? Legal? Engineering? Risk?
Behavioral Contract Structure
BC = {Apermitted, Aprohibited, Cescalation} A = action sets; C = escalation conditions
Chapter 03 Behavioral Contracts

Defining the Behavioral Contract

The Behavioral Contract is the first and most foundational layer of the Trust Stack. It is a formal specification that defines the complete behavioral envelope within which an AI system is permitted to operate in a specific deployment context.

Definition Behavioral Contract

A formal specification defining what an AI model is permitted and prohibited from doing in a specific deployment context, and the conditions under which behavior must be escalated to a human. A Behavioral Contract is version-controlled, linked to a specific model deployment, referenced in the system's audit trail, and reviewed on a defined schedule. It is the reference state against which runtime behavior is measured.

A Behavioral Contract is not a system prompt. System prompts are operational instructions that guide model behavior in a single session. A Behavioral Contract is a governing specification that defines the behavioral envelope within which all system prompts must operate. The system prompt is an implementation detail. The Behavioral Contract is the specification.

The contract has three components. The permitted action set defines what the system may do: the tasks it may perform, the data it may access, the outputs it may produce, and the tools it may invoke. The prohibited action set defines what the system may not do: the topics it may not address, the actions it may not take, and the outputs it may not produce regardless of instruction. The escalation condition set defines the situations in which the system must transfer control to a human, including ambiguous cases, high-stakes decisions, and boundary conditions.

Research Context

Constitutional AI [3] demonstrates that explicit behavioral rules, specified at training time and applied through self-critique during generation, substantially improve the consistency of model behavior with defined values. The Behavioral Contract operates at the deployment level and complements this approach: the contract specifies the deployment-specific rules that the organization requires, while Constitutional AI [3] and RLHF [4] techniques shape the underlying model toward alignment with general values.

Writing a Behavioral Contract

A minimal Behavioral Contract covers five elements. First, scope: the specific use case, user population, and data environment the system is authorized to serve. Second, permitted outputs: the types, formats, and content categories of outputs the system may produce. Third, prohibited outputs: explicit prohibitions on content categories, data types, and action types. Fourth, escalation triggers: the conditions under which the system must not produce an output and must instead transfer the interaction to a human. Fifth, review schedule: the cadence at which the contract is reviewed and re-approved.

A Behavioral Contract for a document review agent in a legal context might permit: summarizing documents within scope, identifying potentially relevant clauses, and flagging documents for attorney review. It might prohibit: providing legal advice, accessing documents outside the defined matter, and producing outputs that characterize the legal strength of a position. It might escalate: when a document contains potential privilege issues, when the request is outside defined scope, or when the output confidence is below a threshold that the contract specifies.

800 600 400 200 0 12 8 Single deployment 48 18 5 deployments 180 42 20 deployments 820 110 100 deployments Violation cost (relative units) Implementation cost (relative units) Cost of behavioral violations versus implementation cost across deployment scale. Directional illustration. Values not derived from systematic survey data.
Chapter 03 Behavioral Contracts

Version Control and Temporal Reference

A Behavioral Contract that is not version-controlled does not serve its audit function. The audit trail for an AI system must be able to reference the specific version of the Behavioral Contract that was in effect at the time of each logged event. Without this temporal reference, the audit trail records that events occurred but cannot assess whether they were compliant with the contract in effect at the time.

Version control for Behavioral Contracts requires: a versioning scheme that uniquely identifies each contract revision, a deployment record that maps model deployments to contract versions, and an audit trail schema that includes the contract version identifier in each logged event. These three requirements are straightforward to implement and must be considered before deployment, not after.

Contract review schedule is a governance question, not a technical one. Contracts should be reviewed when: the model is updated or replaced, the deployment context changes materially, a trust incident occurs, or the review schedule interval elapses. In high-stakes deployments, quarterly review is appropriate. In lower-stakes deployments, semi-annual review may be sufficient. The review must result in a new version number, a dated approval signature, and an update to the deployment record.

Behavioral Contract Ownership

The Behavioral Contract must have a named owner who is accountable for its accuracy and currency. Suggested ownership model:

  • Legal/Compliance: approves prohibited action sets and escalation conditions that have regulatory or liability implications
  • Business Owner: approves permitted action sets and scope definition based on use case requirements
  • AI Engineering: translates the contract into system-level controls and confirms implementability
  • Risk/Security: reviews the contract for threat scenarios not captured by the business or legal review
  • Named Owner: a single person responsible for the contract's currency, review schedule, and escalation to leadership when the contract requires revision

The Contract as Compliance Artifact

Under the EU AI Act [5], providers and deployers of high-risk AI systems are required to establish technical documentation, maintain logs, and ensure human oversight capability. The Behavioral Contract, when properly implemented and version-controlled, satisfies the documentation and specification requirements of these provisions directly. It also provides the reference against which the oversight capability can be evaluated: human oversight is meaningful only when the humans overseeing the system have a formal specification of what the system should and should not do.

Organizations subject to both the EU AI Act [5] and sectoral regulations, such as financial services regulations requiring model risk management or healthcare regulations requiring software documentation, should map their Behavioral Contracts to both the sectoral requirements and the EU AI Act [5] obligations. A single contract can satisfy both, but the mapping must be explicit and documented.

Chapter Takeaways

  • A Behavioral Contract is a formal, version-controlled specification with three components: permitted, prohibited, and escalation
  • It is not a system prompt. It is the governing specification within which system prompts operate
  • Version control and temporal reference are prerequisites for audit function
  • Named ownership and review schedule are governance requirements, not implementation details

Questions for Your Team

  1. Draft the first version of a Behavioral Contract for your highest-stakes AI system. What do you discover you cannot specify, and why?
  2. Who has the authority to approve the prohibited action set in your organization?
4
Chapter 04

Provenance Chain

The Provenance Chain traces every AI output back to its source document, retrieval decision, and model invocation. Without it, disputed outputs cannot be investigated. With it, every output is a defensible artifact.

Pages cover the formal definition, schema, implementation requirements, and the role of provenance in dispute resolution.

Key Takeaways

  • A Provenance Chain is the traceable path from an AI output back to its source document, retrieval decision, and model invocation. It is the mechanism that makes AI outputs investigable after the fact.
  • Provenance operates at the retrieval layer. An AI system that does not retrieve from external sources still requires provenance tracing for the input context it receives and the model version that processed it.
  • Provenance is not hallucination detection. It does not verify that outputs are correct. It verifies that outputs can be traced to a specific set of inputs, enabling correctness to be assessed post-hoc.
  • The provenance schema must be determined before deployment. Retrofitting provenance tracing to a live system is substantially more expensive than designing for it from the start.

Leadership Questions

  1. For a disputed AI output produced last quarter, could you identify today which documents the model retrieved and why it retrieved them?
  2. Does your AI infrastructure store retrieval context at query time, or only the final output?
  3. What is your retention policy for provenance records? Does it align with your regulatory obligation period?
Provenance Chain Schema
PC(o) = {srcid, tr, scorerel, passageref} o = output; src = source; t_r = retrieval time; score = relevance
Chapter 04 Provenance Chain

What Provenance Traces

Definition Provenance Chain

The traceable path from an AI output back to the source documents retrieved, the retrieval decisions made, and the model invocation that produced the output. A Provenance Chain for an output o consists of: the source identifier (src_id), the retrieval timestamp (t_r), the relevance score assigned to each retrieved passage (score_rel), and the passage reference (passage_ref) linking the output to the specific text that informed it. The Provenance Chain is stored in the Audit Trail (L4) and referenced in any Trust Attestation (L5).

Provenance Chain is the second layer of the Trust Stack and the mechanism that makes AI outputs investigable. Without provenance, an audit log records that an output was produced but provides no basis for evaluating whether the output was grounded in authoritative sources, whether the retrieval decision was appropriate, or whether the model was operating within the behavioral constraints defined in L1.

The provenance schema has four required fields per output. The source identifier uniquely identifies each document or data source from which content was retrieved. The retrieval timestamp records when the retrieval occurred, enabling correlation with the model version and contract version active at that time. The relevance score records the system's assessment of the retrieved passage's relevance to the query, enabling post-hoc evaluation of retrieval quality. The passage reference links the output to the specific text passage that informed it, enabling direct verification that the output is grounded in a retrievable source.

Provenance in Retrieval-Augmented Systems

AI systems that use retrieval-augmented generation (RAG) architectures have the most extensive provenance requirements. Each generation event may involve multiple retrieved passages from multiple sources, and the provenance chain must capture all of them. The schema must record: the query formulation, the retrieval strategy, the number and identities of passages retrieved, the relevance scores, the passages selected for context inclusion, and the relationship between each included passage and the final output.

This may appear complex, but it is architecturally straightforward when designed before deployment. The retrieval pipeline already computes relevance scores and selects passages. Provenance tracing requires routing those computations to a structured store alongside the output. The engineering cost is low when designed in. The investigation cost of operating without it is high.

Common Failure Mode

Many enterprise AI systems store the final prompt and final output but not the intermediate retrieval decisions. When a disputed output arises, the retrieval context must be reconstructed from the vector database's current state, which may have changed since the query was made, and from infrastructure logs, which may not have been retained. This reconstruction is often impossible and always expensive.

Provenance for Non-RAG Systems

AI systems that do not retrieve from external sources still require provenance tracing, though the schema is simpler. The minimum provenance record for a non-RAG system includes: the input received, the model identifier and version, the system prompt version, and the output produced. This record provides the basis for investigating whether the system was operating within its Behavioral Contract at the time of the output and whether the model version in use was the authorized version.

For agentic systems that take multi-step actions, provenance extends to tool invocations: each tool call, its parameters, the authority under which it was made (referencing the Access Controls layer, L3), and the result. This extended provenance schema is the basis for auditing agentic behavior and the essential prerequisite for any meaningful post-incident investigation of an autonomous AI action.

Chapter Takeaways

  • Provenance Chain traces every output to source, retrieval decision, and model invocation
  • Four required fields: source ID, retrieval timestamp, relevance score, passage reference
  • Provenance must be designed in before deployment. Retrofitting is expensive and often impossible
  • Agentic systems require extended provenance covering tool invocations and authority references

Questions for Your Team

  1. What does your current RAG infrastructure store at query time? Enumerate what is and is not captured.
  2. What is your retention period for retrieval logs, and how does it compare to your contractual and regulatory obligation periods?
Chapter 04 Provenance Chain

Provenance and Dispute Resolution

The primary operational value of the Provenance Chain is dispute resolution. When an AI-assisted decision is challenged, the investigation requires answering four questions: What information did the system have when it produced the output? What sources did it draw on? Was the retrieval appropriate given the query? Was the output grounded in the retrieved content? The Provenance Chain is the evidence base for answering all four.

Without the Provenance Chain, dispute resolution is a judgment call based on incomplete information. With it, dispute resolution is an investigation based on a complete and immutable record. The difference between these two modes is the difference between a defensible position and an indefensible one in a regulatory inquiry, legal proceeding, or counterparty dispute.

Provenance records must be immutable once written. A provenance store that can be modified after the fact provides no audit value. Implementation requires append-only storage, or storage with cryptographic integrity verification, that prevents modification of records after they are written. This is an infrastructure requirement that must be specified in the deployment architecture.

Scenario

An insurance company deploys an AI system to assist with claims assessment. A claimant disputes a coverage determination, asserting that relevant policy language was not considered. Without a Provenance Chain, the insurer cannot demonstrate which policy documents were retrieved or which passages were weighted in the assessment. With a complete Provenance Chain, the insurer can produce a record showing exactly which policy sections were retrieved, their relevance scores, and the specific passages that were included in the model's context window. The dispute resolves in hours rather than weeks, and the legal exposure is bounded.

Provenance Retention Policy

Provenance records must be retained for at least as long as the decisions they underpin are subject to challenge. In regulated industries, this is determined by the applicable regulatory framework. Financial services decisions may be subject to challenge for multi-year periods. Healthcare decisions may be subject to longer retention requirements under applicable standards. Legal decisions may be subject to the applicable statute of limitations.

Retention policy for provenance records should be set by the organization's legal and compliance function based on the regulatory and contractual obligations applicable to each AI deployment. The default should be conservatively long rather than conservatively short. The cost of retaining provenance records beyond the minimum required period is low. The cost of being unable to produce required records because they were deleted is high.

Implementation Note

Implement provenance storage as a separate append-only service from the main application database. This separation provides three benefits: it prevents accidental modification, it enables independent retention policy management, and it allows provenance records to be queried and exported for audit without requiring access to the live application database. The provenance service should expose a query API that accepts an output identifier and returns the complete provenance record for that output.

5
Chapter 05

Access Controls

Access Controls define who and what can invoke an AI system and with what scope. In agentic deployments, they are the boundary of each agent's authority and the primary mechanism for containing blast radius when behavior moves outside defined bounds.

Pages cover the access control model, permission scoping, agent authority boundaries, and containment design.

Key Takeaways

  • Access Controls (L3) govern invocation authority: who may call the AI system, with what inputs, and with what resulting permissions for downstream actions.
  • In agentic systems, access controls define the action radius of each agent. An agent without explicit permission boundaries can acquire authority beyond its defined scope through tool chaining.
  • Access controls must be grounded in the Behavioral Contract (L1). An access policy that contradicts the Behavioral Contract does not provide meaningful control.
  • The principle of least privilege is as applicable to AI agents as to human users: each agent receives the minimum access required to complete its defined task and no more.

Leadership Questions

  1. For each AI agent in your enterprise, can you enumerate every tool and data source it has permission to access?
  2. Are those permissions reviewed when the agent's task scope changes?
  3. What is the mechanism by which an agent's access is revoked if a behavioral anomaly is detected?
Access Decision Model
Access(agent, r) → {permit, deny, escalate} r = resource or action requested; decision grounded in BC (L1)
Chapter 05 Access Controls

The Access Control Layer

Access Controls are the third layer of the Trust Stack. They govern the boundary between what the AI system can request and what it is authorized to do. In simple deployments, access controls govern which users may invoke the system and which data the system may access. In agentic deployments, access controls govern which tools each agent may invoke, which data sources it may read, which downstream systems it may write to, and under what conditions it may delegate to other agents.

The access control decision for any agent-resource pair is one of three outcomes: permit, deny, or escalate. Permit authorizes the access and logs it. Deny blocks the access and logs it. Escalate transfers the decision to a human, logs the escalation, and blocks action pending human authorization. This three-outcome model is more expressive than the binary permit/deny model used in traditional access control and reflects the reality that many AI access decisions fall into a zone of ambiguity that neither unilateral permission nor unilateral denial handles well.

0% 20% 40% 60% 80% 100% Document Summarizer 20% 18% Email Drafter 25% 22% Data Analyst 45% 42% Code Review Agent 55% 52% Research Agent 80% 75% Task scope (required) Permission scope (granted) Permission scope versus task scope across agent types. Directional illustration. Values not derived from systematic survey data.

Access controls must be grounded in the Behavioral Contract (L1). The Behavioral Contract specifies the permitted action set for the AI system. The access controls implement that specification at the infrastructure level: they translate the contract's permitted actions into the specific permissions granted to the system in the specific environment in which it operates. An access policy that contradicts the Behavioral Contract, by granting permissions for actions the contract prohibits, represents a failure of the trust architecture at the integration point between L1 and L3.

Agentic Access Control

Agentic AI systems present access control challenges that traditional enterprise identity and access management frameworks were not designed to address. An agent that can invoke tools can, through multi-step tool chaining, acquire effective access to resources that are not in its explicit permission set. For example, an agent with permission to read from a document store and write to a messaging system can, by extracting content from the document store and including it in a message, exfiltrate information beyond its authorized data access boundary.

Agentic access control requires permission scoping at the action level, not just the resource level. The agent must be authorized not only to access a resource but to perform a specific action on that resource: read, write, create, delete, execute. And in chained-action scenarios, the combined effect of a sequence of authorized individual actions must also be evaluated against the Behavioral Contract's intent, not only each action in isolation.

Action Blast Radius

In agentic AI deployments, the action blast radius is the set of all possible downstream effects of an agent's authorized actions, including second-order effects through tool chaining and data access. Every access control design should explicitly enumerate the blast radius and evaluate whether it is within acceptable bounds given the Behavioral Contract's intent. If the blast radius exceeds acceptable bounds, the permission scope must be reduced, not the task scope.

Chapter 05 Access Controls

Least Privilege for AI Agents

The principle of least privilege, granting each entity the minimum permissions required to perform its defined function and no more, is the foundational design principle for agentic access controls. In practice, applying least privilege to AI agents requires a different process than applying it to human users, because the agent's permission requirements are not fully enumerable in advance. The agent's task may lead it to request permissions that were not anticipated at design time.

The practical approach is to start with a minimal permission set derived from the Behavioral Contract's permitted action set, observe the agent's permission requests during controlled testing, expand permissions deliberately and with audit trail logging for each expansion, and review the full permission set on the same schedule as the Behavioral Contract review. Permissions granted during testing that were not anticipated at design time should trigger a Behavioral Contract review to ensure the expansion is consistent with the contract's intent.

Access Control Implementation Checklist

Before deploying an AI agent, verify:

  • Every tool the agent may invoke is enumerated and the invocation permission is explicitly granted
  • Every data source the agent may read is enumerated and the read permission is explicitly granted
  • Every system the agent may write to is enumerated, the write permission is explicitly granted, and the Behavioral Contract permits the write action
  • The escalation mechanism is functional and routes to a named human owner
  • Permission revocation can be executed within a defined time window (target: under 5 minutes for high-stakes deployments)
  • The permission set has been reviewed against the Behavioral Contract and is consistent with its permitted action set

Access Control and the Audit Trail

Every access decision, permit, deny, or escalate, must be logged in the Audit Trail (L4). The access control layer is a primary source of events for the audit trail. Access logs must include: the agent identity, the resource or action requested, the decision, the policy rule that governed the decision, and the timestamp. This log is the evidence base for demonstrating, in an audit or regulatory inquiry, that the AI system operated within its authorized access boundaries.

Access control logs must be correlated with the Provenance Chain (L2) records when an investigation requires tracing an output to both its content sources and its action authority. In agentic deployments, the complete picture of what an agent did requires both the provenance record (what information it accessed and used) and the access control log (what actions it was authorized to take and which it actually took).

Chapter Takeaways

  • Access Controls implement the Behavioral Contract at the infrastructure level: three outcomes, permit, deny, escalate
  • Agentic deployments require action-level permission scoping and blast radius evaluation
  • Least privilege is the foundational design principle: start minimal, expand deliberately
  • Every access decision is logged in L4 and correlated with L2 provenance records

Questions for Your Team

  1. Map the action blast radius for your highest-stakes AI agent. What can it do in a worst-case sequence of authorized tool calls?
  2. How quickly can you revoke access for a specific AI agent if a behavioral anomaly is detected?
6
Chapter 06

Audit Trails

The Audit Trail is the complete, immutable, temporally ordered record of every significant event in the AI system's operation. It is the evidence base for investigation, the substrate for attestation, and the mechanism that makes the Trust Stack verifiable after the fact.

Pages cover the schema for a sufficient audit trail, immutability requirements, query capability, and integration with L1-L3.

Key Takeaways

  • A sufficient audit trail records: input, output, retrieval context (L2 reference), model identifier and version, the Behavioral Contract version in effect (L1 reference), access decisions (L3 reference), and timestamp.
  • Immutability is a technical requirement, not a policy statement. Audit trails must use append-only storage or cryptographic integrity verification that makes post-hoc modification detectable.
  • An audit trail that cannot be queried efficiently is operationally useless. Query capability is as important as completeness.
  • The audit trail is the evidentiary foundation for Trust Attestation (L5). Attestation without an underlying audit trail is a declaration without evidence.

Leadership Questions

  1. Can you query your AI system's audit trail by output ID and retrieve the complete event record, including model version, contract version, and retrieval context, within one minute?
  2. Is your audit trail stored in a way that makes modification detectable?
  3. What is the maximum investigation time required to answer a regulatory query about a specific AI decision?
Audit Event Schema
Audit(i) = {input, output, ctx, modelid, t, log} ctx = retrieval context ref; log = contract version + access decisions
Chapter 06 Audit Trails

What a Sufficient Audit Trail Contains

The Audit Trail is the fourth layer of the Trust Stack. It is the complete, immutable record of every significant event in the AI system's operation. The key word is "sufficient": a log that records timestamps and output text is not a sufficient audit trail for regulatory or legal purposes. A sufficient audit trail is one from which a complete picture of what the system did, on what basis, under what constraints, and with what authority can be reconstructed without reference to any additional source.

A sufficient audit trail has seven required elements per event. The input: the complete input presented to the model, including the system context. The output: the complete output produced, not truncated or summarized. The retrieval context: a reference to the Provenance Chain (L2) record for this event, enabling reconstruction of what sources were retrieved and how. The model identifier and version: the specific model that processed the input. The Behavioral Contract version: the specific contract version in effect at the time of the event. The access decisions: references to the Access Control (L3) log entries associated with this event. And the timestamp: a precise, time-zone-qualified timestamp enabling temporal correlation with other events.

Standards Alignment

The EU AI Act [5] Article 12 requires that high-risk AI systems ensure a level of traceability of the system's functioning throughout its lifecycle. The NIST AI RMF [1] Manage function includes requirements for ongoing monitoring and logging. A sufficient audit trail, as defined in the Trust Stack, satisfies both requirements while also providing the evidentiary foundation for board-level attestation that neither standard directly specifies.

Immutability as a Technical Requirement

Audit trail immutability is a technical requirement with specific architectural implications. A policy that states "audit logs shall not be modified" is not sufficient. Technical controls must make modification detectable even if it occurs. The minimum technical requirement is append-only storage, meaning the storage system prevents modification of existing records by design, not by policy. Where append-only storage is not feasible, cryptographic integrity verification, hashing each record and periodically anchoring the hash chain, makes modification detectable after the fact.

Immutability requirements extend to the provenance records referenced in the audit trail. If the provenance store can be modified after the fact, an audit trail that references it provides weaker evidentiary value than one that references an immutable provenance record. The architecture should treat both the audit trail and the provenance store as immutable systems.

Chapter Takeaways

  • Seven required elements: input, output, retrieval context ref, model ID, contract version, access decisions ref, timestamp
  • Immutability requires append-only storage or cryptographic verification, not policy alone
  • Query capability is an operational requirement equal in importance to completeness
  • The audit trail is the evidentiary foundation for Trust Attestation

Questions for Your Team

  1. Walk through a simulated investigation: pick a recent AI output and time how long it takes to reconstruct the complete event record from your current infrastructure.
  2. What would it take to reduce that time to under five minutes for any event in the last 12 months?
Chapter 06 Audit Trails

Query Capability and Investigation Readiness

A complete and immutable audit trail that cannot be queried efficiently is an archive, not an audit system. Investigation readiness requires that the audit trail can be queried by any combination of output ID, user ID, agent ID, time range, model version, and contract version, and that queries return results within a time frame that supports investigation under realistic conditions.

The operational standard for investigation readiness is: a regulatory query about a specific AI decision, identified by output ID or approximate timestamp, should be answerable within one business day from the time the query is received. This standard requires both a complete audit trail and a query system capable of surfacing the relevant records rapidly. Meeting this standard requires deliberate architectural investment in audit trail queryability, not just completeness.

Investigation readiness also requires that the people responsible for conducting investigations know how to use the audit trail system. Documentation of the query interface, training for the investigation team, and periodic rehearsal of investigation scenarios are operational requirements, not nice-to-haves. A comprehensive audit system that nobody knows how to use under pressure provides limited protection when it is most needed.

Scenario

A financial services regulator issues a query requesting documentation of all AI-assisted credit decisions involving a specific product category over a 90-day period. Without a queryable audit trail, the firm must manually review application records, extract AI output logs from a separate system, correlate them by date and product, and reconstruct the retrieval context from infrastructure logs. This takes four weeks and three legal holds. With a Trust Stack audit trail, the query is executed in the audit system, the results are exported to a structured report, and the response is delivered in three business days, with provenance records attached.

Audit Trail Integration Across Layers

The Audit Trail is the integrating layer of the Trust Stack. Every event generated by L1 (contract version changes, contract violations detected), L2 (provenance records created), and L3 (access decisions made) must be captured in the L4 audit trail and linked by a common event identifier. This linkage is what makes the Trust Stack an architecture rather than a collection of independent controls.

The event identifier is a unique ID generated at the time of each AI invocation and threaded through all associated records in L2, L3, and L4. Any record associated with a given invocation can be retrieved by querying for the event ID. This design ensures that a single query can surface the complete picture of what happened in a specific AI interaction: the behavioral contract in effect, the provenance of the output, the access decisions made, and the full event record.

7
Chapter 07

Board-Level Disclosure

Board-Level Disclosure is the interface between the Trust Stack and governance. It comprises the reporting, attestation, and disclosure mechanisms that make AI system behavior visible to boards, regulators, and counterparties, and introduces Trust Attestation as the formal instrument for this disclosure.

Pages cover Trust Attestation, board reporting requirements, and disclosure under regulatory frameworks.

Key Takeaways

  • Trust Attestation is a formal declaration that an AI system operated within defined constraints during a specified period, signed by a responsible party and suitable for regulatory submission.
  • Board-Level Disclosure (L5) is the only layer of the Trust Stack that is directly visible to the governance level. Without L5, even a fully implemented L1-L4 remains invisible to the board.
  • Trust Attestation requires all four underlying layers to be implemented. An attestation that cannot be grounded in a complete audit trail is an assertion, not a declaration.
  • Boards should receive Trust Attestations on a defined schedule and should be able to ask specific questions about system behavior that the attestation is required to answer.

Leadership Questions

  1. Has your board received a formal attestation about the behavior of any AI system in the last 12 months?
  2. If a regulator requested a Trust Attestation for your highest-stakes AI deployment within 30 days, could you produce one?
  3. Who in your organization has the authority and the knowledge to sign a Trust Attestation?
Trust Attestation Structure
Att = sign(TSstate, party, t) TS_state = layer coverage assessment; party = responsible signatory
Chapter 07 Board-Level Disclosure

Trust Attestation

Definition Trust Attestation

A formal declaration that an AI system operated within defined constraints during a specified period, signed by a responsible party, grounded in a complete Audit Trail (L4), and suitable for regulatory review and board-level governance. A Trust Attestation specifies: the AI system and deployment scope, the time period covered, the Behavioral Contract version in effect, the material events recorded in the Audit Trail during the period, any deviations from the Behavioral Contract detected and remediated, and the responsible party's signature attesting to the accuracy of the declaration.

Trust Attestation is the fifth and final layer of the Trust Stack, and the layer that converts an internal trust architecture into an external governance artifact. An enterprise with a fully implemented L1-L4 has the technical capability to verify AI system behavior. But that capability is invisible to the board, to regulators, and to counterparties unless it is expressed through a formal attestation process. L5 is the bridge between internal verification capability and external governance accountability.

The attestation is signed by a responsible party. In most organizations, the appropriate signatories are: the Chief AI Officer or equivalent for the technical accuracy of the attestation, and Legal or Compliance for the regulatory accuracy. In organizations without a Chief AI Officer, the CTO or CIO is the appropriate technical signatory. The board should designate the required signatories in its AI governance policy.

Regulatory Context

The EU AI Act [5] imposes documentation, logging, and transparency obligations on deployers of high-risk AI systems. The HELM evaluation framework [6] demonstrates that consistent AI evaluation requires structured, systematic approaches rather than ad-hoc assessment. Trust Attestation is the mechanism that converts the ongoing operational compliance achieved through L1-L4 into the structured documentation required by regulatory disclosure obligations.

What the Board Needs to Know

Board-level AI governance requires boards to exercise oversight of AI systems that they do not fully understand at a technical level. The Trust Attestation is designed to bridge this gap: it provides a structured, periodically produced document that answers the specific questions a board needs answered without requiring board members to understand the technical details of model operation.

A board-ready Trust Attestation answers five questions. First: which AI systems are in scope and what do they do? Second: what behavioral constraints govern each system, and have those constraints been formally approved? Third: did each system operate within its constraints during the period covered? Fourth: what material events, anomalies, or deviations occurred, and how were they addressed? Fifth: who is attesting to the accuracy of this declaration and on what basis?

Chapter Takeaways

  • Trust Attestation is the formal instrument that makes Trust Stack compliance visible to the governance level
  • It requires a complete L1-L4 implementation as its evidentiary foundation
  • Board-level AI governance requires periodic attestations, not one-time assessments
  • Named signatories with defined accountability are a governance requirement

Questions for Your Team

  1. Draft a Trust Attestation template for your board. What information does it require that you cannot currently produce?
  2. What is the appropriate attestation frequency given your regulatory obligations and board's risk appetite?
Chapter 07 Board-Level Disclosure

Attestation Frequency and Scope

The appropriate frequency for Trust Attestations depends on the risk classification of the AI systems in scope and the regulatory and contractual obligations applicable to the organization. In most enterprises, a quarterly attestation covering all high-risk AI deployments provides adequate governance cadence. Annual attestations are insufficient for high-stakes deployments where regulatory obligations may require more frequent reporting. Monthly attestations are appropriate for AI systems operating in highly regulated environments, such as those subject to financial services or healthcare regulations, where behavioral anomalies must be reported to regulators on a short timeline.

The scope of each attestation should be explicitly defined and consistent across reporting periods, enabling the board to compare attestations over time and assess whether the AI program's trust posture is improving or deteriorating. Scope changes, such as the addition of a new AI system to the attestation scope or the retirement of an existing system, should be noted explicitly in the attestation document.

Trust Attestation Template Elements

A complete Trust Attestation includes:

  • Cover declaration: attestation period, systems in scope, and responsible signatories
  • Behavioral Contract summary: which contract versions were in effect, when they were approved, and whether any revisions occurred during the period
  • Operational summary: volume of AI interactions in scope, access decisions by outcome (permit/deny/escalate), and material audit trail findings
  • Incident register: any behavioral deviations detected, their classification, and the remediation actions taken
  • Regulatory alignment statement: a declaration that the systems in scope operated in a manner consistent with applicable regulatory obligations during the period
  • Signature block: named signatories with title and date

Disclosure to Counterparties and Regulators

In addition to board-level disclosure, enterprises increasingly face requests for AI system trust documentation from counterparties and regulators. Counterparties, including customers, partners, and auditors, may request documentation that AI systems operating on their behalf or with access to their data meet defined behavioral standards. Regulators may require submission of AI system documentation as part of licensing, examination, or enforcement processes.

The Trust Attestation, when properly structured, is a multi-purpose document. The same attestation that serves the board's governance function can, with appropriate redaction of proprietary technical details, serve as the basis for counterparty and regulatory disclosure. This multi-purpose utility is one of the practical benefits of the Trust Stack architecture: the governance investment made for internal oversight simultaneously builds the documentation required for external accountability.

8
Chapter 08

Trust Maturity Model

The Trust Maturity Model provides a structured progression from reactive trust posture to verified trust posture, enabling organizations to assess their current state and plan a credible path to full Trust Stack implementation.

Pages cover the four maturity tiers and the implementation roadmap for each layer at each tier.

Key Takeaways

  • The Trust Maturity Model defines four tiers: Reactive (no formal trust architecture), Defined (L1 and partial L2-L3), Managed (L1-L4 implemented), and Verified (full L1-L5 with attestation capability).
  • Most enterprise AI programs in 2026 operate at the Reactive or early Defined tier. The jump from Reactive to Defined is the highest-leverage move available to most organizations.
  • Maturity is measured per deployment, not per organization. A single organization can operate different deployments at different maturity tiers, which is appropriate when risk levels differ.
  • The Verified tier is the minimum required for board-level attestation and for regulatory submissions that require formal documentation of AI system behavior.

Leadership Questions

  1. At which maturity tier does each of your three highest-stakes AI deployments operate today?
  2. What is the target tier for each deployment, and by when?
  3. What is the gap between your current tier and the tier required by your highest applicable regulatory obligation?
Maturity Tier Definition
Maturity ∈ {Reactive, Defined, Managed, Verified} Verified = full L1-L5 implementation with attestation capability
Chapter 08 Trust Maturity Model

Four Tiers of Trust Maturity

The Trust Maturity Model provides a structured progression from unmanaged trust posture to verified trust posture. It is designed to enable two things: an honest assessment of where an organization currently stands, and a credible plan for moving to the next tier. The model is prescriptive about what each tier requires but not about how long the progression takes. Progression speed depends on organizational complexity, regulatory pressure, and the risk profile of the AI deployments in scope.

Tier 1 is Reactive. No formal trust architecture exists. Behavioral constraints, if any, are informal and not version-controlled. Output provenance is not traced. Access is governed by general IT access policies rather than AI-specific controls. Audit trails exist as system logs but are not structured for investigation. Board reporting on AI behavior is absent or ad-hoc.

Tier 2 is Defined. At least one Behavioral Contract has been formally documented, approved, and version-controlled (L1). Provenance tracing is partially implemented for the highest-risk deployments (partial L2). Access controls reflect AI-specific permission scoping for at least some deployments (partial L3). Audit trails include model version and timestamp but may not include provenance references or contract version (partial L4). Board reporting covers AI program status but not system-level behavioral compliance.

Maturity Tier L1: Behavioral Contract L2: Provenance Chain L3: Access Controls L4: Audit Trail L5: Attestation
Reactive None None General IT policy only System logs only None
Defined Documented, version-controlled for highest-risk deployments Partial: highest-risk RAG systems AI-specific scoping for some deployments Model version and timestamp included None
Managed All high-risk deployments covered; review schedule established Full schema for all high-risk deployments Least privilege enforced; blast radius documented All 7 required elements; immutable storage Internal reporting only
Verified All deployments covered; automated drift detection All deployments; real-time provenance; retention policy enforced All deployments; automated anomaly detection All deployments; queryable; cross-layer correlation Formal attestations on schedule; regulatory-grade documentation
Trust Stack layer coverage by maturity tier. Organizations should target tier coverage appropriate to the risk profile of each deployment.

Tier 3 is Managed. All high-risk deployments operate under formal Behavioral Contracts (L1). Provenance Chain tracing is fully implemented for all high-risk deployments (L2). Least privilege access controls are enforced and blast radius is documented for all AI agents (L3). Audit trails include all seven required elements and are stored in immutable systems (L4). Internal board reporting covers AI behavioral compliance but attestations are not yet at the standard required for external regulatory submission (partial L5).

Chapter 08 Trust Maturity Model

The Verified Tier

Tier 4 is Verified. All AI deployments, not only those classified as high-risk, operate under formal Behavioral Contracts. Provenance Chain tracing covers all deployments, with real-time capture and retention policies aligned to regulatory requirements. Access controls are enforced across all deployments with automated anomaly detection that flags permission requests outside the defined scope. Audit trails are complete, immutable, and queryable with cross-layer correlation. Trust Attestations are produced on a defined schedule, signed by named responsible parties, and meet the documentation standard required for external regulatory submission. The Verified tier is the minimum required for organizations facing formal regulatory AI disclosure obligations.

Reactive Defined Managed Verified L1 Behavioral Contract 0% 90% 100% 100% L2 Provenance Chain 0% 30% 90% 100% L3 Access Controls 0% 20% 80% 100% L4 Audit Trail 0% 20% 85% 100% L5 Attestation 0% 0% 20% 100% Trust Stack layer coverage versus maturity tier. Directional illustration. Values not derived from systematic survey data.

The Highest-Leverage Move

For most organizations in 2026, the highest-leverage move available is transitioning from Reactive to Defined tier for their three highest-stakes AI deployments. This transition requires: documenting and approving a Behavioral Contract for each deployment, implementing basic provenance tracing for retrieval-augmented deployments, updating audit trail schemas to include model version and timestamp at minimum, and establishing a quarterly reporting cadence on AI behavioral compliance to the appropriate governance body.

This is not a large engineering project. It is a governance and process project with targeted engineering components. The Behavioral Contract is a document. The provenance tracing is a schema update to an existing data store. The audit trail update is a logging schema change. The reporting cadence is a meeting and a report template. Organizations that treat the transition to Defined tier as an engineering initiative rather than a governance initiative consistently underestimate the governance components and overestimate the engineering components.

Chapter Takeaways

  • Four tiers: Reactive, Defined, Managed, Verified
  • Maturity is per-deployment, not per-organization
  • Reactive to Defined is the highest-leverage move for most organizations
  • Verified tier is the minimum for formal regulatory disclosure obligations

Questions for Your Team

  1. Score each of your five most important AI deployments against the maturity matrix. What is the median tier?
  2. What is the minimum tier required by your highest applicable regulatory obligation, and by when must you reach it?
9
Chapter 09

Deployment Playbook

The Trust Stack Deployment Playbook is a structured sequence for implementing the five-layer architecture in a live enterprise environment, covering team composition, sequencing, tooling choices, and the decision gates between phases.

Pages cover the three-phase implementation roadmap, team composition, and the build versus configure decision for each layer.

Key Takeaways

  • Trust Stack implementation proceeds in three phases: Pilot (L1 and basic L2-L3, weeks 1-8), Hardening (full L2-L4, weeks 9-18), and Enterprise Rollout (L5 and scaled coverage, weeks 19+).
  • The pilot team is small and intentional: 1 Senior AI Engineer, 1 Data Engineer, 1 Security Architect (part-time), 1 Legal/Compliance representative, 1 Product Owner with AI literacy.
  • L1 (Behavioral Contract) is always the first artifact. Nothing else can be correctly implemented without it.
  • The decision gate between phases is a behavioral contract review, an audit trail completeness check, and a legal sign-off on the attestation draft. All three must pass before proceeding.

Leadership Questions

  1. Who is the named owner of the Trust Stack implementation program, and what authority do they have to make architectural decisions?
  2. What is the highest-risk AI deployment in your organization, and is it the right starting point for the pilot?
  3. What is the go/no-go criterion for advancing from Pilot to Hardening phase?
Risk Accumulation Model
Risk(t) = Σ gaps(Li) × impact(Li) gaps = unimplemented layer components; impact = per-layer risk weight
Chapter 09 Deployment Playbook

Phase 1: Pilot (Weeks 1-8, Indicative)

The pilot phase focuses on a single high-stakes AI deployment and delivers three artifacts: a complete Behavioral Contract for that deployment, basic provenance tracing for its highest-volume interactions, and an updated audit trail schema that includes model version, contract version, and timestamp. These three artifacts demonstrate that the Trust Stack architecture is viable in the organization's specific environment and provide the learning necessary to plan the Hardening phase.

The pilot team is composed of five roles. A Senior AI Engineer who owns the technical implementation of provenance tracing and audit trail schema updates. A Data Engineer who owns the data infrastructure for provenance and audit storage. A Security Architect (part-time) who reviews the access control scope and blast radius documentation. A Legal or Compliance representative who co-owns the Behavioral Contract and confirms it addresses applicable regulatory requirements. A Product Owner with AI literacy who owns requirements, evaluation, and stakeholder communication.

100 80 60 40 20 0 14 1 Investigation time (days) 35 90 Disputed outputs resolved (%) 20 85 Regulatory readiness (%) 30 98 Audit trail completeness (%) 10 95 Board visibility (%) Before Trust Stack After Trust Stack Before and after Trust Stack layer implementation for a representative deployment. Directional illustration. Values not derived from systematic survey data.

The pilot phase go/no-go gate has three criteria. First, the Behavioral Contract is documented, version-controlled, approved by Legal/Compliance, and confirmed to be technically implementable by Engineering. Second, provenance records for a representative sample of interactions can be queried and return the source ID, retrieval timestamp, relevance score, and passage reference. Third, the audit trail for the same sample of interactions includes model version, contract version, and a reference to the provenance record. If all three criteria pass, the pilot advances to Hardening. If any fail, the pilot phase extends until the gap is resolved.

Phase 2: Hardening (Weeks 9-18, Indicative)

The Hardening phase expands the Trust Stack implementation from the pilot deployment to all high-risk AI deployments in the organization. It also completes the implementation of all five layers for the pilot deployment, including full Access Controls with blast radius documentation (L3) and a draft Trust Attestation (L5). The Hardening phase go/no-go gate is a complete audit trail query demonstration for each high-risk deployment, a legal review of the draft attestation template, and an internal board briefing on the Trust Stack program status.

Chapter Takeaways

  • Phase 1 Pilot: single deployment, three artifacts (BC, provenance, audit schema), weeks 1-8
  • Pilot gate: BC approved, provenance queryable, audit trail complete
  • Phase 2 Hardening: expand to all high-risk deployments, complete L3 and L5 draft
  • Phase 3 Enterprise Rollout: full organizational coverage and attestation capability

Questions for Your Team

  1. Identify the single AI deployment where Trust Stack implementation would have the greatest risk-reduction impact. Start there.
  2. Schedule the Phase 1 pilot review now, before the implementation starts, so the go/no-go gate has a date it must meet.
Chapter 09 Deployment Playbook

Phase 3: Enterprise Rollout (Weeks 19+, Indicative)

The Enterprise Rollout phase achieves full organizational coverage: all AI deployments, not only those classified as high-risk, operate under Behavioral Contracts. Trust Attestations are produced on schedule and signed by the designated responsible parties. The board receives its first formal Trust Attestation. The attestation template is reviewed by legal counsel and confirmed to be suitable for regulatory submission if required. The program enters an ongoing operational cadence: quarterly attestations, annual Behavioral Contract reviews for all deployments, and continuous audit trail monitoring with defined incident response procedures for detected behavioral anomalies.

Build Versus Configure Versus Buy

Each Trust Stack layer has different build/configure/buy characteristics. L1 (Behavioral Contract) is always built: it is a document, not a system. No vendor product substitutes for the organizational governance work of specifying what each AI system is permitted and prohibited from doing. The contract is drafted by Legal, Engineering, and the Business Owner and maintained in the organization's document management system.

L2 (Provenance Chain) is typically configured from existing RAG infrastructure: most enterprise retrieval frameworks capture relevance scores and retrieved passages; the configuration work is routing those captures to a structured, append-only store. Custom engineering is required only when the existing retrieval infrastructure lacks the provenance schema fields required by the Trust Stack definition.

L3 (Access Controls) is typically configured from existing identity and access management infrastructure, with custom extensions for AI-specific permission scoping. Agentic deployments may require purpose-built agent permission management components not available in standard IAM products in 2026. L4 (Audit Trail) is typically built on existing logging infrastructure with schema extensions. L5 (Trust Attestation) is built: the attestation document, process, and reporting cadence are governance artifacts that vendor products can support but not substitute for.

Build vs. Configure vs. Buy Summary
  • L1 Behavioral Contract: Build. Document owned by governance. No vendor substitute.
  • L2 Provenance Chain: Configure from existing RAG infrastructure. Custom build only when retrieval framework lacks required fields.
  • L3 Access Controls: Configure from existing IAM. Custom build for agentic permission scoping where standard IAM products are insufficient.
  • L4 Audit Trail: Configure from existing logging infrastructure with schema extensions. Buy append-only storage if not already available.
  • L5 Trust Attestation: Build. Governance document and process owned by Legal/Compliance. Vendor tools may support reporting but do not substitute for the attestation itself.
10
Chapter 10

Trust Under Pressure

When a trust incident occurs, the Trust Stack converts a crisis into an investigation. This chapter covers incident response for AI behavioral anomalies, the recovery sequence, and how the architecture you built before the incident determines the organization's ability to respond during it.

Pages cover the detect, audit, remediate, attest recovery sequence and organizational readiness for trust incidents.

Key Takeaways

  • A trust incident is any event in which an AI system's behavior falls outside its Behavioral Contract, reaches external parties in a way that creates legal or regulatory exposure, or generates outputs that cannot be traced to authorized sources.
  • The Trust Stack recovery sequence is: detect (anomaly surfaces in audit trail or access control log), audit (provenance and audit trail used to reconstruct the complete event), remediate (behavioral contract updated, access modified, or deployment suspended), attest (a supplementary attestation documents the incident and remediation for regulatory and board review).
  • Organizations at Reactive maturity face open-ended investigation timelines when a trust incident occurs. Organizations at Verified maturity can bound the investigation to the period before detection.
  • Trust incidents are not failures of the Trust Stack. They are the conditions the Trust Stack was built to handle. Detection and recovery, not prevention alone, is the measure of architectural success.

Leadership Questions

  1. When was the last time your organization ran a tabletop exercise for an AI trust incident? What did you learn?
  2. What is the maximum time between an AI behavioral anomaly occurring and your organization detecting it, given your current monitoring?
  3. Who is notified, and in what order, when a trust incident is detected?
Recovery Sequence
Recovery = detect → audit → remediate → attest each step requires the Trust Stack layer below it to be implemented
Chapter 10 Trust Under Pressure

The Detect-Audit-Remediate-Attest Cycle

When a trust incident occurs, the Trust Stack converts a crisis into a structured investigation. The recovery sequence has four steps. Detection: the anomaly surfaces in the audit trail (L4) through automated monitoring, in the access control log (L3) through an anomalous access decision, or through an external report of unexpected AI behavior. Audit: the Provenance Chain (L2) and Audit Trail (L4) are used to reconstruct the complete event record for the incident, establishing the scope, the affected interactions, and the behavioral departure. Remediate: the Behavioral Contract (L1) is reviewed and updated if the incident reveals a gap in its specification; access controls (L3) are modified or the deployment is suspended if the incident involves unauthorized access; the technical root cause is resolved. Attest: a supplementary Trust Attestation is produced documenting the incident, the investigation findings, the remediation actions, and the current state of the deployment.

Each step in the recovery sequence requires the Trust Stack layer below it. Detection requires an Audit Trail and Access Control log that surface anomalies. Audit requires a complete Provenance Chain and Audit Trail. Remediation requires a Behavioral Contract to revise. Attestation requires all four layers to be functional and complete. Organizations at the Reactive maturity tier have none of these capabilities and face open-ended investigations. Organizations at the Verified tier can execute each step in the sequence with defined resources and bounded timelines.

100% 80% 60% 40% 20% 0% Reactive Defined Managed Verified 90% 65% 38% 8% Unresolved trust incidents over time (indicative) Trust incident frequency versus maturity tier over time. Directional illustration. Values not derived from systematic survey data.

Red-teaming research [7] demonstrates that unexpected AI behaviors under adversarial conditions are common even in well-evaluated systems. The Trust Stack does not eliminate unexpected behaviors. It provides the architecture to detect them when they occur, investigate them when detected, remediate them when investigated, and attest to the remediation when complete. This is the appropriate goal for enterprise AI governance: not the absence of trust incidents, but the ability to handle them with the speed, completeness, and accountability that boards and regulators require.

Organizational Readiness

Run a tabletop exercise simulating an AI trust incident before one occurs. Use a realistic scenario: an AI-assisted decision is challenged by a regulator who requests complete documentation within 30 days. Time how long each step in the detect-audit-remediate-attest sequence takes with your current infrastructure. The gaps you discover in the exercise are far cheaper to address than the gaps you discover during the actual incident.

The Architecture That Earns Trust

Trust in AI systems is not earned by claiming models are safe and reliable. It is earned by demonstrating, through a verifiable architecture, that the systems operate within defined constraints, that their outputs are traceable to authoritative sources, that their access is bounded and monitored, that their operation is fully auditable, and that a responsible party is prepared to attest to all of this in writing. That is what the Trust Stack provides.

The coined terms introduced in this book, Trust Stack, Behavioral Contract, Provenance Chain, Trust Attestation, are offered to the field as a shared vocabulary for enterprise AI governance. They originate with this work and are subject to the intellectual property notice below. They are designed to survive the journey from this page to a board agenda, a regulatory submission, and a practitioner's implementation roadmap, because that is the journey that enterprise AI trust must make.

Final Takeaways

  • The Trust Stack converts trust incidents from crises into structured investigations
  • Detection requires L3-L4; audit requires L2-L4; remediation requires L1; attestation requires L1-L5
  • Organizational readiness requires tabletop exercises before the first real incident
  • Trust is earned through verifiable architecture, not claims about model quality

The Closing Question

  1. If a major AI trust incident occurred today, which step in the detect-audit-remediate-attest sequence would fail first due to missing architecture?
  2. Schedule the work to close that gap before the question stops being hypothetical.

About the Authors

The Trust Stack is the product of research and practice across enterprise AI deployments spanning financial services, healthcare, legal, and industrial sectors.

Arjun Jaggi

Enterprise AI Researcher, Strategist & Framework Architect

Arjun Jaggi is a researcher and enterprise AI strategist who builds original frameworks for AI governance, deployment architecture, and organizational trust. His research introduces formal constructs, including the Trust Stack, Behavioral Contract, Provenance Chain, and Trust Attestation coined in this book, that give enterprise leaders precise vocabulary for problems no existing standard fully addresses.

His peer-reviewed research includes co-authoring MEDFIT-LLM (IEEE RMKMATE 2025, DOI:10.1109/RMKMATE64574.2025.11042816), a domain-specific fine-tuning study demonstrating the performance advantages of behavioral specialization at the model level. He advises Fortune 500 organizations on AI strategy, governance, and the infrastructure layer required to make enterprise AI programs verifiable, auditable, and board-ready.

arjunjaggi.com

Aditya Karnam Gururaj Rao

AI Researcher & Applied Systems Architect

Aditya Karnam Gururaj Rao is a researcher and practitioner specializing in applied AI systems, governance architecture, and the formal structure of enterprise AI frameworks. His research spans model evaluation, domain-specific fine-tuning, and the design of verifiable AI deployment systems that meet regulatory and operational standards.

He co-authored MEDFIT-LLM (IEEE RMKMATE 2025, DOI:10.1109/RMKMATE64574.2025.11042816), a peer-reviewed study demonstrating the performance advantages of domain-specific fine-tuning in medical AI. His contribution to the Trust Stack focuses on the formal mathematical definitions of each layer, including the set-theoretic specification of the Behavioral Contract and the provenance function PC(o) that defines the Provenance Chain. He brings to this work a rigorous approach to making AI governance constructs precise enough to implement, audit, and cite.

adityakarnam.com

This book is an original concept paper published on arjunjaggi.com. The coined terms Trust Stack, Behavioral Contract, Provenance Chain, and Trust Attestation originate with this work.

© 2026 Arjun Jaggi and Aditya Karnam Gururaj Rao. All rights reserved. Academic citation permitted with attribution; commercial use and derivative frameworks require written permission.

References The Trust Stack

References

[1]NIST. Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology, DOI:10.6028/NIST.AI.100-1, 2023.
[2]ISO/IEC. Information technology: Artificial intelligence: Management system. ISO/IEC 42001:2023.
[3]Bai, Y. et al. Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073, 2022.
[4]Ouyang, L. et al. Training language models to follow instructions with human feedback. arXiv:2203.02155, 2022.
[5]European Parliament and Council. Regulation (EU) 2024/1689 on Artificial Intelligence (EU AI Act). Official Journal of the European Union, 2024.
[6]Liang, P. et al. Holistic Evaluation of Language Models (HELM). arXiv:2211.09110, 2022.
[7]Perez, E. et al. Red Teaming Language Models with Language Models. arXiv:2202.03286, 2022.
[8]Rao, A.K.G., Jaggi, A., Naidu, P. MEDFIT-LLM: Domain-Specific Fine-Tuning of Large Language Models for Medical Applications. IEEE RMKMATE 2025, DOI:10.1109/RMKMATE64574.2025.11042816.

© 2026 Arjun Jaggi and Aditya Karnam Gururaj Rao. All rights reserved. Academic citation permitted with attribution; commercial use and derivative frameworks require written permission.

Glossary The Trust Stack

Glossary of Coined Terms

The following terms are coined in this work. They originate with Arjun Jaggi and Aditya Karnam Gururaj Rao and are subject to the intellectual property notice in this book. Academic citation is permitted with attribution.

Trust Stack The five-layer architecture for verifiable trust in enterprise AI, comprising L1 Behavioral Contract, L2 Provenance Chain, L3 Access Controls, L4 Audit Trail, and L5 Board-Level Disclosure. Complete only when all five layers are implemented and each layer references the one below it.
Behavioral Contract A formal, version-controlled specification defining what an AI model is permitted and prohibited from doing in a specific deployment context, and the conditions under which behavior must be escalated to a human. The first and foundational layer of the Trust Stack (L1).
Provenance Chain The traceable path from an AI output back to the source documents retrieved, the retrieval decisions made, and the model invocation that produced the output. Defined as PC(o) = {src_id, t_r, score_rel, passage_ref} for output o. The second layer of the Trust Stack (L2).
Trust Attestation A formal declaration that an AI system operated within defined constraints during a specified period, signed by a responsible party, grounded in a complete Audit Trail, and suitable for regulatory review and board-level governance. The instrument of the fifth and final layer (L5) of the Trust Stack.
Trust Deficit The structural absence of behavioral specification, output provenance tracing, and attestation capability in enterprise AI deployments. The problem the Trust Stack is designed to solve.
Verified Maturity Tier The fourth and highest tier of the Trust Maturity Model, in which all five Trust Stack layers are implemented across all deployments, Trust Attestations are produced on schedule, and the documentation standard required for regulatory submission is met.
Action Blast Radius In agentic AI deployments, the set of all possible downstream effects of an agent's authorized actions, including second-order effects through tool chaining and data access. Must be enumerated and evaluated against the Behavioral Contract's intent before deployment.
Trust Incident Any event in which an AI system's behavior falls outside its Behavioral Contract, reaches external parties in a way that creates legal or regulatory exposure, or generates outputs that cannot be traced to authorized sources through the Provenance Chain.
THE TRUST STACK

The five-layer architecture for verifiable trust in enterprise AI.

Behavioral Contract. Provenance Chain. Access Controls. Audit Trail. Board-Level Disclosure. The Trust Stack provides the architecture to build AI programs that boards can attest to and regulators can audit.

ARJUN JAGGI
ADITYA KARNAM GURURAJ RAO
arjunjaggi.com  ·  August 2026
© 2026 Arjun Jaggi and Aditya Karnam Gururaj Rao. All rights reserved. Academic citation permitted with attribution; commercial use and derivative frameworks require written permission.