Research Index

AOI Runtime Governance Evidence Map

A continuously maintained index of primary research, technical architectures, standards, market developments, and adjacent work contributing to the emerging discipline of Runtime Governance.

The map catalogs supporting, overlapping, adjacent, and challenging work. Inclusion does not imply endorsement.

The Evidence Map describes the state of the literature, not the state of AO Integrity.

Schema v744 sources9 claims

Corpus facts

Derived from stored corpus records.

Total sources

Primary sources

0 external contributions recorded.

Classification metrics

Counts depend on AOI's current taxonomy.

Overlap flagged

Challenge or complicate

Claim Map

Stored claims and source relationships.

Coverage statements are rendered exactly as stored in the Schema v7 corpus.

Claim filter

Claim	Label	Type	Coverage	Sources
EL	Execution Legitimacy Whether a specific execution remains a legitimate exercise of conferred authority at the moment it occurs. The five conditions operationalize this determination.	determination	Fourteen sources. Three added 2026-08-03, all as support or complement rather than competition: aviation dispatch release as a two-party determination made against current conditions, financial settlement finality as an ex ante determinate irrevocability point, and certificate-bound admission as a conditional-execution model. Support remains strongest from outside AI. Competing treatments continue to address individual conditions rather than the integrated five.	e005 - differentiatesIntent-Governed Tool Authorization for AI Agents
RA	Runtime Admissibility The determination of whether execution should proceed right now. Not whether memory, data, context, or a consequence is admissible — those may be inputs or subordinate questions, but RA governs execution.	gate	Twenty-one sources. Six added 2026-08-03. The certificate-bound admission and execution pair (arXiv 2606.11632, 2606.20520) is the corpus's first two-paper treatment separating proposal, admission, and execution with a signed artifact carried between them, with prototype measurements on AWS and Kubernetes. Foundational support continues to come from runtime verification and monitorability literature, now joined by financial settlement-finality standards on the ex ante determinacy of the irrevocability point. No enterprise-scale operational case study found.	e001 - differentiatesRuntime Governance for AI Agents: Policies on Paths
AC	Authority Continuity The preservation of valid, correctly scoped, and traceable authority across identity, delegation, context, and execution as conditions change over time. Broader than delegation-chain correctness: includes whether authority remains legitimate through the interval between grant and effect.	continuity	Twenty sources. Four added 2026-08-03. arXiv 2607.23586 addresses grant survival under agent mutation under the name authorization continuity, with a fixed transition envelope and effect ceiling; arXiv 2606.20520 revalidates against live state at execution rather than at grant. Aviation dispatch regulation (14 CFR 121.533 and 121.663) supplies a long-standing regulatory analogue in which release authority is joint, continuously revisited between release and completion, and non-delegable at the dispatch level. Catalogued treatments continue to address subsets of the claim: agent mutation, infrastructure state, or delegation structure, rather than authority across identity, delegation, organizational context, and execution together. Delegation chains remain structurally representable without non-widening enforced by construction in any deployed standard.	e007 - complementsOpenID Foundation advances authorization for the agent era with new AuthZEN Working Group Drafts
AD	Authority Drift The divergence, over time, between authority a system retains and the authority current organizational circumstances would justify. Not privilege accumulation — a technically correct entitlement may still be illegitimate in current context.	failure	Three sources, unchanged this week. No competing treatments found, and none of the seven sources added 2026-08-03 competes with the canonical definition. arXiv 2606.20520 checks live-state drift at execution against a certified contract, but that comparison is infrastructure state against a contract rather than retained authority against current organizational circumstances, and it was not tagged AD. No standardized business-state event vocabulary exists. No unit of measure, continuous formulation, or threshold located in any source. AD remains the thinnest claim in the corpus by competing coverage and an open search lane.	e014 - supportsOkta transmits only two outbound CAEP event types as of early 2026
CD	Capability Drift Change in an autonomous system's effective ability to produce consequences without a corresponding reevaluation of the authority governing those capabilities. Arises through learned skills, tool composition, memory, model updates, reusable behaviors, delegation, or environmental change.	failure	Three sources, up from one. Two added 2026-08-03 substantively address capability evolution rather than state persistence: a state-bound authorization model that fixes an immutable effect ceiling at grant time and proves that mutation cannot amplify protected effects beyond it, where the mutations enumerated include acquiring skills and tools, revising workflows, and delegating work (arXiv 2607.23586); and a lifecycle-time upgrade governance pipeline with interface, policy, behavioral, and recovery compatibility checks, gated activation, shadow deployment, online monitoring, and rollback, measured on an embodied-agent testbed (arXiv 2604.08059). Both report formal or quantitative results; neither uses the term Capability Drift. A third source added this week, arXiv 2605.26542, is catalogued under AC and RA rather than CD: it governs authority attenuation across a composed tool chain within a session, which is adjacent compositional attenuation rather than change over time in effective ability to produce consequences. Still absent from the record: any enterprise deployment case study, any continuous measure or threshold for capability change, and any treatment tying capability change to organizational rather than technical authority.	e022 - complementsState-Aware Runtime for Long-Horizon LLM Agents: A Conceptual Framework and Research Agenda
GE	Governance Evidence Independently reviewable evidence showing what legitimacy determination was made, which authority and contextual signals informed it, and why the execution was permitted, denied, escalated, or attested. Logs may contribute evidence; event records alone do not establish governance reasoning.	evidence	Thirteen sources. Three added 2026-08-03. Signed decision and outcome records bound to a certified execution contract (arXiv 2606.20520) and evidence-digest binding at admission (arXiv 2606.11632) extend the catalogued set of cryptographically bound evidence models. The signed dispatch release under 14 CFR 121.663 is the corpus's first regulatory instance of a governance artifact that two named parties must both sign before an operation may begin. Regulatory demand remains documented with no finalized technical standard.	e002 - differentiatesDecision Evidence Maturity Model for Agentic AI: A Property-Level Method Specification
OBJ	Evaluation Object The composite execution event: principal, delegation chain, agent, action, target, context, intent, and provenance relevant to a specific execution. Treated as an adopted and refined architecture choice, not a novelty claim.	design	Nine sources. One added 2026-08-03: arXiv 2606.11632 compiles agent proposals into typed execution contracts bound to evidence digests and policy versions, adding a fifth independent 2026 work locating governance at an execution, admission, or bind boundary.	e001 - differentiatesRuntime Governance for AI Agents: Policies on Paths
PE	Policy vs Enforcement The distinction between governance that is declared and governance that is operationally established in the execution path. A supporting claim, not AOI-originated.	supporting	Well covered externally. Primary source is Cavallo's described-versus-established framing; supported by practitioner and distributed-systems material.	e019 - differentiatesProof of Execution: Runtime Verification for Governed AI Agent Actions
MKT	Market Conditions External technical, regulatory, operational, and commercial developments that increase the need for Runtime Governance but do not themselves establish its architecture.	context	Ample. NHI ratios, over-permissioning rates, IAM capability gaps, and regulatory timelines all documented from multiple sources.	e004 - differentiatesAgent Control Specification: Portable runtime governance for AI Agents

Showing 21 of 44 sources.

Clear filters

e007

OpenID Foundation advances authorization for the agent era with new AuthZEN Working Group Drafts

Permalink

Source: OpenID Foundation
Source type: standard
Primary or secondary status: Primary
Domain: Identity
Published: 2026
Claims: AC, EL
Conditions: 3.2, 3.4
Stance: supports
Relation: complements
Strength: supporting
Overlap flag: absorption

AuthZEN Final in January 2026. New AARP draft addresses what happens when policy cannot authorize yet because a prerequisite — approval, consent, delegated authority, attestation, risk assessment, or justification — must first be satisfied. Supports the claim that the gap is upstream of the decision point; also shows the gap being actively closed.

Original link

e008

Prompt injection still drives most agentic AI security failures in production

Permalink

Source: Help Net Security / OWASP
Source type: press
Primary or secondary status: Secondary
Domain: Runtime
Published: 2026-06-11
Claims: EL
Conditions: 3.5
Stance: supports
Relation: supports
Strength: supporting
Overlap flag: none

OWASP-sourced. Prompt injection lands on roughly one in three deployed agents; 88 percent of enterprises deploying agents reported at least one agent-linked security incident. Direct empirical support for Causal Integrity as a live failure mode rather than a hypothetical.

Original link

e009

awesome-ai-agent-attacks — sourced timeline of real AI agent security incidents 2024-2026

Permalink

Source: Community timeline
Source type: incident
Primary or secondary status: Primary
Domain: Runtime
Published: 2026
Claims: EL, AC
Conditions: 3.5, 3.2
Stance: supports
Relation: supports
Strength: supporting
Overlap flag: none

Dated and sourced incident corpus. Includes the Copilot email-summarization case (hidden instructions ingested during summarization, exfiltration from OneDrive/SharePoint/Teams via a trusted Microsoft domain), CVE-2025-6514 MCP RCE, CVE-2025-59536 hooks injection, postmark-mcp supply chain, and GTG-1002. The Copilot case is a close real-world analogue of Appendix C.1.

Original link

e010

Authorization Propagation in Multi-Agent AI Systems: Identity Governance as Infrastructure

Permalink

Source: arXiv 2605.05440 (Tallam)
Source type: research
Primary or secondary status: Primary
Domain: Identity
Published: 2026-05
Claims: AC
Conditions: 3.2
Stance: supports
Relation: supports
Strength: supporting
Overlap flag: none

Examines how authorization propagates across multi-agent systems and frames identity governance as infrastructure rather than policy. Consistent with the delegation-propagation reading in Appendix A. Cited in the thesis reference list. Bears on AC and condition 3.2.

Original link

e011

AI Identity: Standards, Gaps, and Directions

Permalink

Source: arXiv 2604.23280
Source type: research
Primary or secondary status: Primary
Domain: Identity
Published: 2026-04-28
Claims: AC
Conditions: 3.2
Stance: supports
Relation: supports
Strength: supporting
Overlap flag: none

Finds OAuth 2.0 handles one-hop delegation well but lacks multi-hop chaining, cross-domain asynchronous flows, and any mapping between OAuth scopes and agent capabilities; cross-organizational log correlation unsolved. Independent corroboration of the Appendix A reading of RFC 8693 and RFC 9396.

Original link

e013

EU AI Act Article 12 logging obligations reach full enforcement August 2, 2026

Permalink

Source: Salt Security / EU AI Act
Source type: regulatory
Primary or secondary status: Secondary
Domain: Regulatory
Published: 2026
Claims: GE, MKT
Conditions: EV, MKT
Stance: supports
Relation: supports
Strength: supporting
Overlap flag: none

Article 12 mandates automatic logging of events relevant to traceability; logs must be tamper-evident, retained six months minimum. No finalized technical standard yet — prEN 18229-1 and ISO/IEC DIS 24970 still in draft. Demand-side support for the decision receipt and for stating an assurance level explicitly.

Original link

e014

Okta transmits only two outbound CAEP event types as of early 2026

Permalink

Source: Andrew Doering
Source type: practitioner
Primary or secondary status: Secondary
Domain: Identity
Published: 2026
Claims: AD
Conditions: 3.3
Stance: supports
Relation: supports
Strength: emerging
Overlap flag: none

Okta does not emit risk-level-change, device-compliance-change, or assurance-level-change outbound. Sharpens the Appendix A argument that expressibility is not standardization — the deployed vocabulary is narrower than the specified one, well before reaching matter closure or contract expiry.

Original link

e015

Compositional agent harm — individually authorized sub-actions yielding harmful sequences

Permalink

Source: arXiv 2604.23338
Source type: research
Primary or secondary status: Primary
Domain: Runtime
Published: 2026-04
Claims: EL, AC
Conditions: 3.4, 3.2
Stance: supports
Relation: supports
Strength: supporting
Overlap flag: none

Systematic survey establishing that multi-step chains compose individually benign invocations into harmful sequences, and that the security property of a composition cannot be derived from its steps. Stateless per-request engines cannot enforce properties depending on execution history. Direct support for Appendix C.4.

Original link

e016

GitGuardian / Veza — non-human identity sprawl and over-permissioning at scale

Permalink

Source: Security Boulevard / GitGuardian / Veza
Source type: analyst
Primary or secondary status: Secondary
Domain: Market
Published: 2026-07
Claims: MKT, AC
Conditions: MKT, 3.1
Stance: supports
Relation: supports
Strength: emerging
Overlap flag: none

Reports non-human identity to human ratios between 45:1 and 144:1 depending on environment, 73 percent of NHI-held secrets carrying excessive permissions, over 5.5 percent of AWS machine identities holding full administrative privilege, and 92 percent of organizations stating current IAM tooling cannot manage AI agent identities. Aggregated vendor and analyst reporting rather than peer-reviewed measurement. Bears on MKT and, indirectly, condition 3.1.

Original link

e017

The Non-Human Identity Governance Vacuum

Permalink

Source: Cloud Security Alliance
Source type: industry_standard
Primary or secondary status: Primary
Domain: Identity
Published: 2026
Claims: MKT, AC
Conditions: MKT, 3.1
Stance: supports
Relation: supports
Strength: supporting
Overlap flag: none

CSA whitepaper framing NHI and agentic AI governance as a vacuum. Institutional corroboration that the gap is recognized, from a body that predates the thesis.

Original link

e021

Described vs. Established Governance in Agentic AI: Closing the Gap Between Policy and Enforcement

Permalink

Source: SSRN (Cavallo)
Source type: research
Primary or secondary status: Primary
Domain: Policy
Published: 2026
Claims: PE, GE
Conditions: EV
Stance: supports
Relation: supports
Strength: supporting
Overlap flag: none

Frames the gap between governance as described in policy and governance as operationally established in the execution path. Bears on Section 3.8's distinction between security correctness and governance legitimacy, and on the Section 4.4 observation that controls are often described as audited at assurance levels they do not meet.

Original link

e022

State-Aware Runtime for Long-Horizon LLM Agents: A Conceptual Framework and Research Agenda

Permalink

Source: Cambridge Open Engage
Source type: research
Primary or secondary status: Primary
Domain: Memory/State
Published: 2026
Claims: AD, CD, RA
Conditions: 3.3
Stance: supports
Relation: complements
Strength: emerging
Overlap flag: none

Conceptual framework and research agenda for state-aware runtime across long-horizon agent operation. Bears on AD and on the Section 6.1 assumption that business systems can be consulted at execution latency. CD tag is provisional: the source addresses state persistence, and qualifies for CD only where it substantively treats capability evolution.

Original link

e023

Agency doctrine: actual authority, apparent authority, and ratification

Permalink

Source: Oxford Business Law Blog
Source type: legal_doctrine
Primary or secondary status: Primary
Domain: Legal
Published: 2025-04
Claims: EL, AC
Conditions: 3.2, 3.4
Stance: supports
Relation: complements
Strength: foundational
Overlap flag: none

Governance analogue and legal foundation rather than a doctrinal conclusion about software agents. Actual authority, apparent authority, and ratification establish that possession of apparent permission does not prove actual authority; that third parties may act on authority that is operationally credible but institutionally defective; and that later acceptance of an act differs from prior legitimacy. Apparent authority resembles the structure of Appendix C.1, though prompt injection and delegated software action are not doctrinally identical to it absent jurisdiction-specific authority. Ratification is the closest legal analogue to Section 5.2 post-hoc attestation, including its all-or-nothing character.

Original link

e024

Mission command doctrine: commander's intent and disciplined initiative

Permalink

Source: US Air Force AFDP 1-1
Source type: doctrine
Primary or secondary status: Primary
Domain: Ops Doctrine
Published: 2023-08-14
Claims: EL, AC
Conditions: 3.4
Stance: supports
Relation: supports
Strength: foundational
Overlap flag: none

Mission command combines centralized intent with decentralized execution. Commander's intent is doctrinally defined as a concise expression of purpose, method, and end state, including constraints, limitations, assumptions and resources, and is stated as a necessary condition without which mission command cannot be exercised. Subordinates exercise disciplined initiative within it. Establishes an institutional precedent for delegated autonomy bounded by represented intent; does not address machine-readable intent binding. Bears on condition 3.4 and Section 6.2.

Original link

e025

Runtime verification and monitorability of safety properties

Permalink

Source: Springer / RV literature
Source type: research
Primary or secondary status: Primary
Domain: Verification
Published: 2023
Claims: RA, GE
Conditions: EV, OBJ
Stance: supports
Relation: supports
Strength: foundational
Overlap flag: none

Monitorability is defined as the ability to return a positive or negative verdict after a finite execution prefix, with the field concentrating on safety properties because violations are always detectable in finite time. This supplies formal grounding for Section 2.2's claim that the evaluable tuple is only complete at execution, and a rigorous way to state which legitimacy conditions are monitorable at all. Also gives Section 4.3's determinism requirement an existing literature to stand on.

Original link

e026

Verified runtime validation for partially observable hybrid systems

Permalink

Source: arXiv 1811.06502
Source type: research
Primary or secondary status: Primary
Domain: Verification
Published: 2018
Claims: RA, AD
Conditions: 3.3, EV
Stance: supports
Relation: supports
Strength: foundational
Overlap flag: none

Synthesizes runtime monitors from provably safe hybrid system models, robust to partial observability arising from sensor uncertainty and partial controllability arising from actuator disturbance. Establishes what guarantee survives when a monitor cannot observe full system state. The partial-observability treatment corresponds to the Section 6.1 assumption that contextual validity may proceed on cached state, and to the signal-freshness requirement in Section 4.2.

Original link

e027

Saga pattern, compensating transactions, and durable execution

Permalink

Source: Azure Architecture Center / long-running transaction literature
Source type: architecture
Primary or secondary status: Primary
Domain: Distributed Systems
Published: ongoing
Claims: RA, PE, OBJ
Conditions: OBJ
Stance: supports
Relation: adjacent
Strength: foundational
Overlap flag: none

Long-running transactions abandon locks on non-local resources and use compensating transactions that semantically undo rather than restore prior state, typically coordinated by a durable state machine. Establishes an existing engineering distinction between operations that can be compensated and those that cannot. Bears on the reversibility-based allocation in Section 5.3.

Original link

e029

Authenticated Delegation and Authorized AI Agents

Permalink

Source: arXiv 2501.09674
Source type: research
Primary or secondary status: Primary
Domain: Identity
Published: 2025-01
Claims: AC
Conditions: 3.2
Stance: supports
Relation: complements
Strength: supporting
Overlap flag: none

Earlier work connecting agency-law delegation concepts to authenticated agent authorization. Useful bridge between the Legal domain entries and the identity standards track, and evidence that the legal framing is already being imported by researchers.

Original link

e037

From Audit to Admissibility: Why AI Governance Must Move Into The Execution Boundary

Permalink

Source: Medium
Source type: practitioner
Primary or secondary status: Secondary
Domain: Runtime
Published: 2026
Claims: RA, OBJ, PE
Conditions: OBJ, EV
Stance: supports
Relation: supports
Strength: emerging
Overlap flag: naming_collision

Practitioner argument that model outputs are not admissible actions but proposals, and that if a proposal can directly mutate system state without boundary evaluation, the model becomes a source of authority. Uses admissibility and execution boundary framing. Bears on OBJ, RA, and PE.

Original link

e043

Responsibility for operational control and dispatch release: joint authority, continuous re-release, and non-delegable dispatch

Permalink

Source: eCFR - 14 CFR 121.533, with 121.663 (US Federal Aviation Administration)
Source type: regulatory
Primary or secondary status: Primary
Domain: Aviation
Published: 1964-12-31, as amended 1996-01-26
Claims: EL, AC, GE
Conditions: 3.1, 3.3
Stance: supports
Relation: supports
Strength: foundational
Overlap flag: none

Governance analogue from aviation regulation, not a doctrinal claim about software agents. The certificate holder is responsible for operational control; the pilot in command and the aircraft dispatcher are jointly responsible for preflight planning, delay, and dispatch release, and under 121.663 both must sign the release and may do so only if both believe the flight can be made with safety. Authority is not settled at release: the dispatcher must monitor the progress of each flight, issue information necessary for its safety, and cancel or redispatch if either the dispatcher or the pilot in command judges that the flight cannot continue to operate safely as planned or released. 121.663 further provides that a dispatcher may delegate authority to sign a release for a particular flight but may not delegate the authority to dispatch. Establishes a long-standing regulatory precedent separating conferred authority, delegable and non-delegable components, evaluation against current conditions, continuing responsibility after release, and a signed record of the determination. Bears on EL, AC, and the grant-to-effect interval; the signed release is an evidence artifact bearing on GE.

Original link

e044

Principles for financial market infrastructures - Principle 8: Settlement finality

Permalink

Source: CPSS-IOSCO / CPMI-IOSCO, PFMI (CPMI Papers No 101)
Source type: standard
Primary or secondary status: Primary
Domain: Financial Systems
Published: 2012-04-16
Claims: EL, RA
Conditions: 3.5
Stance: supports
Relation: complements
Strength: foundational
Overlap flag: none

Governance analogue from financial market infrastructure standards, not a doctrinal claim about software agents. Principle 8 requires that an FMI provide clear and certain final settlement, at a minimum by the end of the value date, and intraday or in real time where necessary or preferable. The accompanying legal-basis discussion states that there should be a clear legal basis regarding when settlement finality occurs, in order to define when key financial risks are transferred, including the point at which transactions are irrevocable, and treats settlement finality as a building block for risk management rather than a bookkeeping detail. Establishes an existing international standard requiring that the moment at which an action becomes irrevocable be defined ex ante and be legally determinate, including under participant insolvency. The standard governs the determinacy of the finality boundary; it does not specify evidence of a governance determination, and the entry is therefore not tagged GE. Bears on the reversibility-based allocation in Section 5.3 and on the treatment of the consequence-formation boundary; complements the compensating-transaction material catalogued at e027.

Original link

Methodology

How to read this map.

AO Integrity maintains the corpus and performs the classifications. Classification is interpretive. Total, primary, and externally contributed counts are corpus facts. Other metrics depend on the current taxonomy. Records can be corrected or withdrawn. Stable evidence IDs are never renumbered or reused. Withdrawn records remain addressable.

Contribution

What are we missing?

The corpus is intended to include work that supports, overlaps with, complicates, or challenges the Runtime Governance thesis. Sources that make the thesis less comfortable are especially useful.

Please include the source title, original URL, author, organization, or venue as published, publication date, and why it may be relevant. Submission does not imply guaranteed inclusion.

Submit a source