Network Working Group A. Ehstand Internet-Draft Independent Researcher Intended status: Informational 26 September 2026 Expires: 30 March 2027 Terminology for Human Oversight Acts in Automated and Agentic Systems draft-ehstand-oversight-acts-00 Abstract Records produced by automated and agentic systems often represent human oversight as a single, undifferentiated event, such as an approval flag or a confirmation. Such a record does not say whether the person was shown an output, determined whether a named property of it holds, chose a course of action, or permitted an action, and consumers of the record may treat a confirmation click as if it were a check. This document defines terms for four kinds of human oversight act (observation, check, decision, and release) and for related concepts: the oversight act and the overseer, the named property, standing authority, the oversight record, the undifferentiated approval, the check step, error detectability, the fail-open check step, and the check test. It states what a record of each kind of act is, and is not, evidence of, and relates the terms to existing vocabularies. The document defines terminology only. It specifies no protocol, data format, procedure, or measurement method. Discussion Venues This note is to be removed before publishing as an RFC. This is an individual submission. It is intended for discussion on the mailing list of the DISPATCH working group (dispatch@ietf.org), which provides a venue for proposals of new work in the Security area, among others. Related work is discussed on the WIMSE (wimse@ietf.org) and OAuth (oauth@ietf.org) mailing lists. Comments can also be sent to the author. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Ehstand Expires 30 March 2027 [Page 1] Internet-Draft Human Oversight Acts September 2026 Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at https://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 30 March 2027. Copyright Notice Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/ license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License. Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 1.1. Problem Statement . . . . . . . . . . . . . . . . . . . . 3 1.2. Scope . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.3. Conventions . . . . . . . . . . . . . . . . . . . . . . . 4 2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.1. Oversight Act . . . . . . . . . . . . . . . . . . . . . . 5 2.2. Overseer . . . . . . . . . . . . . . . . . . . . . . . . 6 2.3. Observation . . . . . . . . . . . . . . . . . . . . . . . 6 2.4. Named Property . . . . . . . . . . . . . . . . . . . . . 7 2.5. Check . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.6. Checked Property . . . . . . . . . . . . . . . . . . . . 8 2.7. Decision . . . . . . . . . . . . . . . . . . . . . . . . 8 2.8. Release . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.9. Standing Authority . . . . . . . . . . . . . . . . . . . 9 2.10. Oversight Record . . . . . . . . . . . . . . . . . . . . 10 2.11. Undifferentiated Approval . . . . . . . . . . . . . . . . 10 2.12. Check Step . . . . . . . . . . . . . . . . . . . . . . . 11 2.13. Error Detectability . . . . . . . . . . . . . . . . . . . 11 2.14. Fail-Open Check Step . . . . . . . . . . . . . . . . . . 12 2.15. Check Test . . . . . . . . . . . . . . . . . . . . . . . 12 Ehstand Expires 30 March 2027 [Page 2] Internet-Draft Human Oversight Acts September 2026 3. Oversight Acts as Evidence . . . . . . . . . . . . . . . . . 13 4. Relation to Existing Vocabularies . . . . . . . . . . . . . . 14 4.1. W3C PROV-O . . . . . . . . . . . . . . . . . . . . . . . 14 4.2. W3C Verifiable Credentials Data Model 2.0 . . . . . . . . 15 4.3. EU Artificial Intelligence Act, Article 14 . . . . . . . 15 4.4. NIST AI Risk Management Framework . . . . . . . . . . . . 16 4.5. Internet Security Glossary (RFC 4949) . . . . . . . . . . 16 4.6. IETF Work on Agent Authorization and Human Involvement . 17 5. Security Considerations . . . . . . . . . . . . . . . . . . . 18 6. Privacy Considerations . . . . . . . . . . . . . . . . . . . 20 7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 20 8. Informative References . . . . . . . . . . . . . . . . . . . 20 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . 23 Author's Address . . . . . . . . . . . . . . . . . . . . . . . . 23 1. Introduction This section describes the problem this document addresses, its scope, and the conventions it uses. 1.1. Problem Statement Automated and agentic systems increasingly act on behalf of people and organizations: they draft documents, move money, change configurations, and call other systems. Many deployments place a person somewhere in this process, and many records of the process note that a person was involved. Typical forms are an approval flag, a "reviewed" field, a signed confirmation, or a log entry saying that a named user pressed a button. Such records usually do not say what the person did. The same flag can stand for at least four different acts: the person was shown the output; the person determined whether a named property of the output holds; the person chose what should happen next; or the person, using authority they hold, permitted an action. These acts establish different things. Being shown an output establishes nothing about it. Determining that a named property holds establishes that property and no other. Permitting an action establishes that the action was permitted, not that it was correct. When the acts are not distinguished, consumers of the record (auditors, relying parties, and downstream systems and agents) cannot tell which of these things was established. Such a record is easily read as establishing more than it does, so that a confirmation click is treated as though it were a check. The weakness does not lie in any single act. It lies in a vocabulary that has one word for all of them. Ehstand Expires 30 March 2027 [Page 3] Internet-Draft Human Oversight Acts September 2026 Existing texts show the gap. Article 14 of the EU Artificial Intelligence Act [EU-AI-ACT] is headed "Human oversight" and describes what the persons to whom oversight is assigned are to be enabled to do; the definitions in Article 3 of that Regulation do not include the term. The NIST AI Risk Management Framework [NIST-AI-RMF] states in its Appendix C that "Human roles and responsibilities in decision making and overseeing AI systems need to be clearly defined and differentiated." An individual Internet- Draft, [I-D.schrock-human-authorization-binding], notes that agent- action record formats reserve a place for "the human authorization" and leave its semantics undefined. This document provides a small set of terms with which such records, and the texts that specify them, can say which act took place and what its record establishes. 1.2. Scope This document defines terminology. It does not: * specify a protocol, message, or data format; * state when human oversight is needed, or which kind of act a given situation calls for; * specify how checks are to be made, how error detectability is to be determined, or how check tests are to be designed; it gives no values, thresholds, or reference data; * interpret any law, regulation, or framework. The terms are intended for authors of protocols, record formats, and policies, and for those who audit systems against them. Within the IETF, they are intended in particular for specifications that record human involvement in the actions of software agents, such as the user confirmation described in Section 10.7 of [I-D.ietf-wimse-aims], the audit records of [I-D.gilda-wimse-agent-audit-record], and the human- authorization evidence of [I-D.schrock-human-authorization-binding] (Section 4.6). 1.3. Conventions This document does not use the requirement keywords of BCP 14. Where a sentence in the definitions reads as a condition (for example, that a check is over a named property), it states part of a definition: something that does not meet the condition is not an instance of the defined concept. Ehstand Expires 30 March 2027 [Page 4] Internet-Draft Human Oversight Acts September 2026 Each term is a noun. Each definition is a single phrase that can replace the term in running text; it names a broader concept and the characteristics that set the defined concept apart from other concepts under it, following the principles for the writing of definitions in [ISO704]. Notes and an example follow each definition. Notes add information; they do not change the definition. The corresponding verbs (observe, check, decide, and release) are used in running text with the same meanings. In ordinary usage, a release (Section 2.8) is often called an "approval" or an "authorization". This document avoids both words for it. In [RFC4949], and in the OAuth-based work cited in Section 4.6, "authorization" denotes an approval granted to a system entity, or the process of granting it (Section 4.5); and "approval" is commonly used for records that do not state the kind of act (Section 2.11). The words "evidence" and "relying party" are used in their ordinary senses, not in the senses defined for remote attestation in [RFC9334]. The word "verification" is used only in quotations and in the sense of the document cited where it occurs. In this document, "system" means an automated or agentic system: software that produces outputs or takes actions with some degree of autonomy, including software agents that act on behalf of people or organizations. The "object" of an oversight act is the output, action, state, or class of actions of the system to which the act is directed. 2. Terminology 2.1. Oversight Act Definition: An act of a natural person that is directed at an output, action, state, or class of actions of a system and that is recorded, or is meant to be relied on, as human involvement in the operation of that system. Note 1: This document distinguishes three kinds of oversight act by what the overseer does: observation (Section 2.3), check (Section 2.5), and decision (Section 2.7). It distinguishes one kind of decision further, by the authority within which it is made: release (Section 2.8). In this document, "the kinds of oversight act" refers to these four. Note 2: The kind of an act depends on what the overseer did, not on the interface element used to record it. A button labeled "Approve" can record any of the four kinds, several of them at once, or none. Ehstand Expires 30 March 2027 [Page 5] Internet-Draft Human Oversight Acts September 2026 Note 3: A single interaction can contain more than one act, and each act can be recorded separately. Example: A reviewer reads an agent's proposed database change, determines that it modifies only the intended table, and allows it to run within authority the reviewer holds. The interaction contains a check and a release; the release is also a decision. 2.2. Overseer Definition: The natural person who performs a given oversight act. Note 1: A system, a software agent, or an organization is not an overseer in the sense of this document, although an overseer may act on behalf of an organization. Note 2: Different acts concerning the same object can have different overseers. Example: The member of staff who releases a queued payment is the overseer of that act. A manager who later countersigns it is the overseer of a second act. 2.3. Observation Definition: An oversight act in which the overseer is presented with the object and makes no check, decision, or release concerning it. Note 1: Monitoring a dashboard, reading a log, watching an agent's activity stream, and having an output displayed on screen are observations. Whether the overseer perceived or understood the object is not part of the definition. Note 2: A record of an observation shows at most that the object was presented to the overseer. It does not show what the overseer noticed. Note 3: An observation may prompt a later act, such as a decision to interrupt the system. The later act is a separate oversight act. Example: A reviewer keeps a live view of an agent's actions open during a session and takes no action. The session contains observations only. Ehstand Expires 30 March 2027 [Page 6] Internet-Draft Human Oversight Acts September 2026 2.4. Named Property Definition: A property that the object of an oversight act may or may not have, stated before the outcome of that act is known and precisely enough that a person other than the overseer can tell what a determination of it establishes and what it does not. Note 1: "Looks correct" and "approved" are not named properties. "Every total in table 2 equals the sum of its rows" and "the recipient named in the message is the recipient named in the request" are. Note 2: A property that is stated, or changed, only after the outcome is known does not meet this definition for that act; see "Property substitution" in Section 5. Example: For a generated purchase order, "the unit prices match the current price list" is a named property. A determination of it establishes nothing about the quantities ordered. 2.5. Check Definition: An oversight act in which the overseer examines the object in order to determine whether a named property holds for it. Note 1: A check is always over a named property (Section 2.4). An act in which a person looks at an object and forms a general impression, without a named property, is an observation, however carefully it is done. Note 2: A check has an outcome: the property holds, the property does not hold, or the overseer could not determine which. The outcome concerns the checked property only. It establishes nothing about other properties of the same object. Note 3: An overseer may use tools, including automated ones, in making a check. Where the overseer adopts a tool's result without determining the outcome, the overseer's act is an observation of that result; any determination made by the tool is not a check in the sense of this document. Note 4: A check is a single act. The point in a process at which checks are to be made is a check step (Section 2.12). Example: A reviewer determines whether every document identifier cited in a generated report resolves to an existing document. The outcome says nothing about whether the report's conclusions follow from those documents. Ehstand Expires 30 March 2027 [Page 7] Internet-Draft Human Oversight Acts September 2026 2.6. Checked Property Definition: The named property over which a given check is made. Note 1: Each check has exactly one checked property. An overseer who determines several properties of the same object makes several checks. Example: In the example in Section 2.5, the checked property is "every document identifier cited in the report resolves to an existing document". 2.7. Decision Definition: An oversight act in which the overseer selects, from the alternatives available at that point, the course of action to be taken with respect to the object. Note 1: Alternatives include using, disregarding, overriding, or reversing an output; letting an operation continue; interrupting it; and referring the matter to another person. A decision to let an action be carried out is a release only where it is made within standing authority (Section 2.8). Note 2: A decision can be made without any check. A record of a decision does not imply that a check was made. Note 3: Responsibility for a decision remains with the overseer, or with the person or organization on whose behalf the overseer decides. Recording a decision does not transfer that responsibility to the system. Note 4: Where a record format assigns the choice of a course of action to a relying party, a decision in the sense of this document corresponds to that choice when the overseer is, or acts for, the relying party. Example: After reading a proposed refund, a member of staff chooses to hold it for a colleague instead of releasing or rejecting it. 2.8. Release Definition: A decision in which the overseer, acting within standing authority they hold, permits a specified action or class of actions to be carried out. Ehstand Expires 30 March 2027 [Page 8] Internet-Draft Human Oversight Acts September 2026 Note 1: The standing authority (Section 2.9) exists before the act. A release exercises authority; it cannot create the authority it exercises. A decision to permit an action, made by a person who holds no standing authority covering it, is a decision but not a release, whatever its record calls it. Note 2: A release is bounded by what it specifies. Permission for one action does not extend to a different action, or to the same action with different parameters, unless the release specifies a class of actions that includes it. Note 3: Releasing is not checking. A release establishes that an action was permitted. It does not establish that the action, its inputs, or its effects are correct. Note 4: The reasons for not calling this act an "approval" or an "authorization" are given in Section 1.3 and Section 4.5. Example: A team lead who holds signing authority up to a stated amount permits an agent to submit one specific purchase order within that amount. The team lead's act is a release. 2.9. Standing Authority Definition: Authority to permit actions of a class that a person holds before a given oversight act and that was conferred by a source other than that act. Note 1: Typical sources are an employment role, a documented delegation, a contract, law, or the person's authority over their own affairs. Note 2: Standing authority can itself result from delegation. [PROV-O] describes delegation as "the assignment of authority and responsibility to an agent (by itself or by another agent) to carry out a specific activity as a delegate or representative". Note 3: Other documents use "mandate" for a related but different concept. In [I-D.yossif-agent-mandate-problem], a mandate is "The set of constraints the Principal authorizes at T0, expressed as data the Principal signs"; it is created by the principal's act. In the terms of this document, that act is a release over a class of actions, made within the principal's standing authority. This document does not use the word "mandate" for standing authority, because in that usage a mandate results from the act instead of preceding it. Ehstand Expires 30 March 2027 [Page 9] Internet-Draft Human Oversight Acts September 2026 Example: An organization's approval policy that allows members of its finance team to release payments below a stated amount confers standing authority on each of them. 2.10. Oversight Record Definition: A record of a single oversight act that is issued by, or attributable to, the overseer and that identifies the overseer, the kind of act, the object, and the time of the act, and, for a check, the checked property and the outcome. Note 1: For a release, a record that also identifies the standing authority relied on and the action or class of actions permitted allows a consumer to see whether the act was within that authority. Note 2: A record that the system writes about the overseer, without any means of attributing it to the overseer, is not an oversight record in this sense. It states the system's account of the act. Note 3: This document does not define a format. The same record can be carried in several existing formats; see Section 4. Example: A record stating that a named reviewer, at a given time, checked a generated report for the property "every cited identifier resolves", with the outcome "holds", is an oversight record. A second record, stating that the same reviewer then released the report for publication under a named publication policy, is another. 2.11. Undifferentiated Approval Definition: A record of human involvement with a system that does not state which kind of oversight act was performed. Note 1: Approval flags, "reviewed" fields, and confirmation events are usually undifferentiated approvals. Note 2: Without further information, an undifferentiated approval shows only that an interaction was recorded. The interface element that produced it can record any kind of act, several, or none (Section 2.1, Note 2); see also Section 3. Example: A log entry "approved_by: user-17" is an undifferentiated approval. Ehstand Expires 30 March 2027 [Page 10] Internet-Draft Human Oversight Acts September 2026 2.12. Check Step Definition: A point in a process at which checks over a stated named property are to be made, by stated persons under stated conditions, on the objects that reach that point. Note 1: A check step is a procedure; a check is a single act made at it. A check step can produce records without any check being made, for example where a review screen is closed without the object being examined. Note 2: Error detectability (Section 2.13), the property of being fail-open (Section 2.14), and check tests (Section 2.15) concern check steps, not single checks. Example: In a contract workflow, the point at which a member of the legal team examines each generated contract for the property "all required clauses are present" is a check step. 2.13. Error Detectability Definition: A property of a check step that expresses how reliably checks at that step, as made by the persons and under the conditions stated for it, distinguish objects for which its named property does not hold from objects for which it holds, relative to a stated reference population of objects. Note 1: An error, in this entry, is a respect in which an object does not have the named property of the check step. Note 2: The reference population is the set of objects, and of errors in them, on which the checks were made. The same check step can reveal most errors of a kind in one population and few in another. Note 3: Error detectability is distinct from the proportion of objects that a check step flags. A check step that flags every object reveals every error and distinguishes nothing. Note 4: Errors of omission, where something that is required is absent from the object, leave nothing in the object to notice. A statement of error detectability that covers only errors present in the object says nothing about them. Note 5: Where it needs to be distinguished from corresponding properties of automated tools, this property can be called "human error detectability". Ehstand Expires 30 March 2027 [Page 11] Internet-Draft Human Oversight Acts September 2026 Note 6: This document specifies no method for determining error detectability, and no values or thresholds. Example: "Reviewers catch most errors" does not state error detectability: it names no check step, no named property, and no reference population. 2.14. Fail-Open Check Step Definition: A check step at which the failure of a check, whether because the check was not made or because it could not tell an object for which the named property holds from one for which it does not, produces the same record as a check in which the property was found to hold. Note 1: A human check step is fail-open wherever the record is written independently of whether the check was made. In that case, a check that was not in fact made still produces the record "checked", and ordinary operation gives no sign of the failure. Note 2: A check step that is not fail-open is fail-closed: the failure of a check produces a record that differs from the record of a check in which the property was found to hold, for example no record, or the outcome "could not determine". Note 3: [RFC4949] notes that "fail-safe" has two opposing meanings. This document uses "fail-open" and "fail-closed" only with the meanings given here. [I-D.schrock-human-authorization-binding] uses "fail-closed absence" for a related but different rule: a relying party is to treat an absent human-authorization binding as insufficient evidence. Example: A review step for the property "all required clauses are present", in which the field "checked" is set whenever the review screen is closed, is a fail-open check step: closing the screen without reading produces the same record as reading and finding every required clause. 2.15. Check Test Definition: A test of a check step in which a check at that step is given an object for which the step's named property is known not to hold, in order to establish whether the check reports that the property does not hold. Note 1: Testing a detection process with inputs known to contain faults is long established, for example in mutation testing of software [DeMillo1978] and, for human inspectors, in threat image Ehstand Expires 30 March 2027 [Page 12] Internet-Draft Human Oversight Acts September 2026 projection in aviation security screening [Hofer2005]. A recent statement of the principle for fail-open properties in general is [AIKR-FAILOPEN]. This document uses the term for check steps in the oversight of automated and agentic systems. Note 2: A check test tests the check step, not the object. Note 3: Check tests matter most for fail-open check steps, whose failures do not otherwise show in their records. Note 4: This document does not specify how such objects are made, how many are used, or what result is acceptable. Example: A contract that lacks a required clause, given to a check step for the property "all required clauses are present", is the object of a check test. 3. Oversight Acts as Evidence The kinds of oversight act differ in what their records can be used to establish. This section states those differences. It assumes that the record is authentic; authenticity is a separate question (Section 5). * Only a check over a named property is evidence that something is true of the object: namely, that the checked property held, or did not hold, for the object as presented to the overseer at the time of the check. How much weight that evidence carries depends on the error detectability of the check step at which the check was made and, for a fail-open check step, on its check tests. * A release is evidence that the overseer permitted an action. It shows that the permission was within the overseer's authority only to the extent that it identifies the standing authority relied on. It is not evidence that the action was correct. * A decision is evidence of which course of action was chosen, and by whom. It is not, by itself, evidence that the choice was correct, or, unless it is a release, that it was permitted. * An observation is evidence of neither truth nor permission. It shows at most that the object was presented to the overseer. * An undifferentiated approval is evidence only that an interaction was recorded. It is evidence of an observation only if other information shows that the object was presented to the overseer, and of another kind of act only if other information establishes that kind. Ehstand Expires 30 March 2027 [Page 13] Internet-Draft Human Oversight Acts September 2026 +=============+===============+====================+===============+ | Kind of act | About a | That an action was | Of the course | | | property of | permitted | of action | | | the object | | chosen | +=============+===============+====================+===============+ | Observation | No | No | No | +-------------+---------------+--------------------+---------------+ | Check | Yes, for the | No | No | | | checked | | | | | property only | | | +-------------+---------------+--------------------+---------------+ | Decision | No | No | Yes | +-------------+---------------+--------------------+---------------+ | Release | No | Yes, as far as the | Yes | | | | standing authority | | | | | is identified | | +-------------+---------------+--------------------+---------------+ Table 1: What a Record of Each Kind of Act Is Evidence Of A consumer who needs to know both that an action was permitted and that its object had a certain property needs two records: a release and a check. Neither act stands in for the other. Adding records of the same kind does not change their kind. Several observations do not amount to a check, and several releases do not establish that an action was correct. Agreement between several checks of the same named property adds weight only to the extent that the checks are independent of one another. 4. Relation to Existing Vocabularies The terms in this document are meant to be used alongside existing vocabularies, not to replace them. This section notes how they relate to several of them. None of the correspondences below is normative, and none is an interpretation of the documents cited. 4.1. W3C PROV-O PROV-O [PROV-O] describes an activity as "something that occurs over a period of time and acts upon or with entities", an agent as "something that bears some form of responsibility for an activity taking place, for the existence of an entity, or for another agent's activity", and an association as "an assignment of responsibility to an agent for an activity". Ehstand Expires 30 March 2027 [Page 14] Internet-Draft Human Oversight Acts September 2026 An oversight act can be described as a prov:Activity associated with the overseer, a prov:Person, through prov:wasAssociatedWith; the activity uses the object (prov:used) and generates an oversight record, a prov:Entity attributed to the overseer (prov:wasAttributedTo). Where the standing authority relied on in a release results from delegation, it can be described with PROV-O delegation (prov:actedOnBehalfOf, prov:Delegation). PROV-O records that a person was responsible for an activity. It does not distinguish kinds of oversight act, and it has no terms for a named property or for the outcome of a check. These are what the terms in Section 2 add. 4.2. W3C Verifiable Credentials Data Model 2.0 An oversight record can be expressed as a claim in a verifiable credential [VC-DATA-MODEL-2.0] whose issuer is the overseer. The Data Model distinguishes verification from validation and states: "Verification of a credential does not imply evaluation of the truth of claims encoded in the credential." The same separation applies to oversight records. Verifying an oversight record carried as a credential establishes that the overseer issued it. What the record then establishes depends on the kind of act it records (Section 3): a verified record of an observation is still evidence of an observation only. Validation, which the Data Model describes as "The assurance that a claim from a specific issuer satisfies the business requirements of a verifier for a particular use", is where a verifier would look for the kind of act, the checked property, and, for a release, the standing authority. 4.3. EU Artificial Intelligence Act, Article 14 Article 14 of Regulation (EU) 2024/1689 [EU-AI-ACT] is headed "Human oversight". Article 14(4) lists what the natural persons to whom human oversight is assigned are to be enabled to do. Several items on that list correspond to kinds of act defined here: * to "duly monitor its operation" (Article 14(4)(a)) corresponds to observation, and to checks where monitoring is against a named property; Ehstand Expires 30 March 2027 [Page 15] Internet-Draft Human Oversight Acts September 2026 * to "decide, in any particular situation, not to use the high-risk AI system or to otherwise disregard, override or reverse the output of the high-risk AI system" (Article 14(4)(d)) and to "intervene in the operation of the high-risk AI system or interrupt the system" (Article 14(4)(e)) are decisions in the sense of Section 2.7; * Article 14(4)(b) refers to the tendency of "automatically relying or over-relying on the output" (automation bias), which is one reason why checks fail; where a check step is fail-open, such failures do not show in its records (Section 5). Article 14(5) provides, for certain systems, that an identification is to be "separately verified and confirmed by at least two natural persons". The terms of this document can be used to state which kind of act each of these persons performed and, for a check, over which named property. These correspondences are terminological. This document does not interpret the Regulation or state what it requires. 4.4. NIST AI Risk Management Framework The NIST AI Risk Management Framework [NIST-AI-RMF] lists under MAP 3.5: "Processes for human oversight are defined, assessed, and documented in accordance with organizational policies from the GOVERN function." Its Appendix C states that "Human roles and responsibilities in decision making and overseeing AI systems need to be clearly defined and differentiated." The terms in this document can be used to state which kinds of oversight act a process contains and what their records establish. Error detectability and check tests name what an assessment of whether a check step functions would report. The Framework does not use these terms, and this document does not describe how the Framework is to be applied. 4.5. Internet Security Glossary (RFC 4949) [RFC4949] defines "authorization" as "An approval that is granted to a system entity to access a system resource" and as "A process for granting approval to a system entity to access a system resource". In both senses the system entity is the party that receives the authorization; who grants it is not part of the definition. A release in this document is made by a natural person acting within standing authority, and it permits an action of a system rather than only access to a resource. Ehstand Expires 30 March 2027 [Page 16] Internet-Draft Human Oversight Acts September 2026 [RFC4949] also records, as a definition of non-Internet origin that it does not recommend for Internet documents, a sense taken from the SET payment specifications: "The process by which a properly appointed person or persons grants permission to perform some action on behalf of an organization." A release is close to that sense. This document uses a different word so that its term is not read in the recommended sense above. [RFC4949] defines "verification", in its first sense, which is marked as applying in the context of authentication, as "The process of examining information to establish the truth of a claimed fact or value". "Check" is close to that sense; it differs in that a check is an act of a person, is always relative to a named property, and can have the outcome "could not determine". "Accountability", which [RFC4949] defines as the property of a system or system resource "that ensures that the actions of a system entity may be traced uniquely to that entity", is served by records of human involvement only to the extent that they identify the overseer and the kind of act. 4.6. IETF Work on Agent Authorization and Human Involvement Several Internet-Drafts address human involvement in the actions of software agents. The terms in this document can be used with them as follows. * Section 10.7 of [I-D.ietf-wimse-aims] states that user confirmation solicited by an interactive agent framework does not by itself constitute authorization and is to be bound to a verifiable authorization grant issued by the authorization server. In the terms of this document, a record of such a confirmation is an undifferentiated approval unless it states which kind of oversight act the user performed. The authorization grant to which it is bound is issued by the authorization server and is not itself an oversight act. * [I-D.schrock-human-authorization-binding] defines how records bind named-human authorization evidence. The kinds of act defined here can state which act such evidence records: a place reserved for "the human authorization" can hold a record of a check, a decision, or a release, and these establish different things (Section 3). * [I-D.yossif-agent-mandate-problem] characterizes the gap between an authorized agent and an authorized action. It uses "mandate" for the set of constraints that a principal authorizes and signs (Section 2.9, Note 3). In the terms of this document, the Ehstand Expires 30 March 2027 [Page 17] Internet-Draft Human Oversight Acts September 2026 principal's act is a release over a class of actions. The definition of release in Section 2.8, which requires a specified action or class of actions and standing authority, names the human act whose record would bound an action. * [I-D.rosenberg-aiproto-cheq] describes human confirmation of decisions and actions proposed by agents before they are acted upon. In the terms of this document, such a confirmation can be a decision or a release; it is a check only where a named property is stated. * [I-D.sato-soos-hem] specifies what happens when an agent session requires human judgment before it may continue, and defines human decision types. It also addresses approval fatigue (Section 5). * [I-D.gilda-wimse-agent-audit-record] defines a record format for AI agent authorization decisions. Oversight records are records of human acts; they can be referenced from, or kept alongside, such audit records. 5. Security Considerations This document defines terms and introduces no protocol. The considerations below concern the use of the terms, and of records that apply them. Rubber-stamping: A person asked to approve many items can come to approve them without checking. In the terms of this document, the checks at a check step then fail; where the check step is fail-open, its records do not show the failure, and decisions and releases are made without the check that consumers of their records assume. Article 14(4)(b) of [EU-AI-ACT] refers to the related tendency to over-rely on system output, and [I-D.sato-soos-hem] addresses approval fatigue. Recording the kind of act does not prevent rubber-stamping; it prevents a rubber stamp from being read as a check. Error detectability and check tests are the terms for finding out whether a check step still functions. Spoofed and inflated records: Ehstand Expires 30 March 2027 [Page 18] Internet-Draft Human Oversight Acts September 2026 A system under oversight, or an attacker, can produce records that claim human acts which did not take place, or can record an act as a kind whose record establishes more than the act did: for example, an observation as a check, or an act in which no property was named as a check over a named property. Records that are issued by, and verifiably attributable to, the overseer make claims of acts that did not take place harder to produce. Records that state the kind of act and the checked property make an inflated kind visible. As Section 4.2 notes, an authentic record is not thereby a true one. Property substitution: If a property can be stated or changed after the outcome is known, a check can be made to appear to have established whatever the outcome supports. The definition in Section 2.4 excludes such properties. Records in which the named property is fixed before the check is made make such substitution visible. Presentation by the system: A check establishes a property of the object as it was presented to the overseer. A system that controls the presentation can show an object that differs from the one it acts on, or can steer what the overseer attends to. Records that bind the act to the object actually acted on, for example by a digest of that object, narrow this gap. Release beyond standing authority or scope: A release used for an action it did not specify, or a decision to permit an action made by a person without standing authority for it, is not a release in the sense of this document, but its record may look the same. Records that identify the standing authority relied on and the permitted action or class of actions allow a consumer to detect this; see also [I-D.yossif-agent-mandate-problem]. Oversight records as audit targets: Oversight records determine who is held responsible for an outcome. They are therefore of value to anyone who wants to shift that responsibility, and can be altered, deleted, selectively retained, or back-dated. [I-D.ietf-wimse-aims] requires audit records to be tamper-evident; the same consideration applies to oversight records. Check tests: An object made for a check test is defective by construction. If it enters ordinary processing, it is acted on as a defective output. Ehstand Expires 30 March 2027 [Page 19] Internet-Draft Human Oversight Acts September 2026 Shared blind spots: Several overseers who rely on the same source, tool, or presentation can share one blind spot, so that their checks agree without being independent (Section 3). 6. Privacy Considerations Oversight records identify people. A record of an oversight act names, or can be linked to, the overseer, and states what that person did, when, and with what outcome. Collected over time, such records describe a person's work: how quickly they act, how often they check, and how often their checks miss errors. Statements of error detectability and results of check tests that concern identified persons are assessments of those persons' performance. [RFC6973] describes surveillance and identification as privacy threats; records of oversight acts can give rise to both. Those who specify or deploy records that use these terms may wish to consider the following: * Such records are personal data in many jurisdictions, and their use to monitor or evaluate the persons concerned can be subject to specific legal rules, including rules on employee monitoring. * The identity a record needs depends on its use. Showing that a release was within standing authority generally requires the holder of that authority to be identified. Showing that a property was checked may need only a role or a pseudonymous identifier, with identification available when it is needed. * Records of observation, such as records of screen time or attention, can amount to surveillance of the overseer while establishing little about the system (Section 3). * Error detectability can be stated for a check step as carried out by a group of persons rather than for each individual. * An oversight record can contain, or point to, the object of the act, which may itself contain personal data of third parties. 7. IANA Considerations This document has no IANA actions. 8. Informative References Ehstand Expires 30 March 2027 [Page 20] Internet-Draft Human Oversight Acts September 2026 [AIKR-FAILOPEN] Parle, A., "Re: Knowledge Representation for the Trust Layer", Message to the public-aikr@w3.org mailing list, W3C AI KR (Artificial Intelligence Knowledge Representation) Community Group; includes a quoted message by N. Templeman, 17 September 2026, . [DeMillo1978] DeMillo, R. A., Lipton, R. J., and F. G. Sayward, "Hints on Test Data Selection: Help for the Practicing Programmer", Computer, vol. 11, no. 4, pp. 34-41, DOI 10.1109/C-M.1978.218136, April 1978, . [EU-AI-ACT] European Parliament and Council of the European Union, "Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act)", Official Journal of the European Union, OJ L, 2024/1689, 12 July 2024, . [Hofer2005] Hofer, F. and A. Schwaninger, "Using threat image projection data for assessing individual screener performance", WIT Transactions on the Built Environment, vol. 82, pp. 417-426, DOI 10.2495/SAFE050411, May 2005, . [I-D.gilda-wimse-agent-audit-record] Gilda, S., "An Audit Record Format for AI Agent Authorization Decisions", Work in Progress, Internet- Draft, draft-gilda-wimse-agent-audit-record-00, 21 September 2026, . Ehstand Expires 30 March 2027 [Page 21] Internet-Draft Human Oversight Acts September 2026 [I-D.ietf-wimse-aims] Kasselman, P., Lombardo, J., Rosomakho, Y., Campbell, B., Steele, N., and A. Parecki, "AI Identity Management System", Work in Progress, Internet-Draft, draft-ietf- wimse-aims-00, 15 September 2026, . [I-D.rosenberg-aiproto-cheq] Rosenberg, J., White, P., and C. F. Jennings, "CHEQ: A Protocol for Confirmation AI Agent Decisions with Human in the Loop (HITL)", Work in Progress, Internet-Draft, draft- rosenberg-aiproto-cheq-00, 19 October 2025, . [I-D.sato-soos-hem] Sato, T., "The Human Escalation Mechanism (HEM) for Agentic AI Systems", Work in Progress, Internet-Draft, draft-sato-soos-hem-07, 5 September 2026, . [I-D.schrock-human-authorization-binding] Schrock, I., "Binding Named-Human Authorization Evidence into Agent-Action Records", Work in Progress, Internet- Draft, draft-schrock-human-authorization-binding-00, 3 July 2026, . [I-D.yossif-agent-mandate-problem] Yossif, M. K., "Problem Statement: Verifiable Human Mandates for Autonomous Agent Actions", Work in Progress, Internet-Draft, draft-yossif-agent-mandate-problem-00, 22 July 2026, . [ISO704] International Organization for Standardization, "Terminology work - Principles and methods", ISO 704:2022, July 2022, . [NIST-AI-RMF] National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)", NIST AI 100-1, DOI 10.6028/NIST.AI.100-1, January 2023, . Ehstand Expires 30 March 2027 [Page 22] Internet-Draft Human Oversight Acts September 2026 [PROV-O] Lebo, T., Ed., Sahoo, S., Ed., and D. McGuinness, Ed., "PROV-O: The PROV Ontology", W3C REC REC-prov-o-20130430, 30 April 2013, . [RFC4949] Shirey, R., "Internet Security Glossary, Version 2", FYI 36, RFC 4949, DOI 10.17487/RFC4949, August 2007, . [RFC6973] Cooper, A., Tschofenig, H., Aboba, B., Peterson, J., Morris, J., Hansen, M., and R. Smith, "Privacy Considerations for Internet Protocols", RFC 6973, DOI 10.17487/RFC6973, July 2013, . [RFC9334] Birkholz, H., Thaler, D., Richardson, M., Smith, N., and W. Pan, "Remote ATtestation procedureS (RATS) Architecture", RFC 9334, DOI 10.17487/RFC9334, January 2023, . [VC-DATA-MODEL-2.0] Sporny, M., Ed., Thibodeau Jr, T., Ed., Herman, I., Ed., Cohen, G., Ed., and M. B. Jones, Ed., "Verifiable Credentials Data Model v2.0", W3C REC REC-vc-data-model- 2.0-20250515, 15 May 2025, . Acknowledgments The treatment of fail-open check steps and check tests in this document was prompted by a discussion of fail-open properties in the W3C AI Knowledge Representation Community Group, in which N. Templeman stated, and A. Parle restated, that a property that fails open can be shown to work only by testing it with an input built to make it fail, since ordinary, honest operation never produces the failing case [AIKR-FAILOPEN]. Author's Address Andreas Ehstand Independent Researcher Starnberg Germany Email: ehstand.schule@gmail.com URI: https://orcid.org/0009-0006-3773-7796 Ehstand Expires 30 March 2027 [Page 23]