Pre-Action Risk-Graded Assurance for Agent Interactions

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶

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 25 February 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

Emerging agent-governance systems evaluate policy at action boundaries and return, in the common case, a binary decision. This is adequate for uniform allow/deny gating but does not capture a property that liability-bearing and safety-sensitive deployments increasingly require: assurance should scale with risk, be evaluated before the action proceeds, be expressible uniformly across different kinds of action, and leave a provable record of why a given assurance level was required and whether it was met.¶

Consider three actions an agent may take: (a) re-authenticate the human on whose behalf it acts; (b) execute a delegated task with external effect; (c) select which computational resource (for instance, which model) will handle a request. These belong to different domains, yet each admits a natural notion of "assurance": authentication depth for (a), oversight degree for (b), and selection rigor for (c). Today each is governed by a separate mechanism, if at all.¶

This document describes a single pre-action stage that unifies them at the interface level. It does not prescribe how risk is computed, nor the control law that maps risk to requirement; those are implementation and product matters, and in some deployments proprietary. What it standardizes is the shape of the decision and its record, so that audit consumers and interoperating systems can reason about assurance uniformly.¶

This version (-01) additionally describes the autonomy-asymmetry control law (Section 4) as a reference design for the mapping from risk signal to assurance requirement. The control law is informative; an implementation is free to use any mapping. It is described here because it captures a property — asymmetric response to failure versus success — that recurs across safety-sensitive systems and that the phase-seal chain makes observable.¶

2. Conventions and Definitions

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶

3. The Pre-Action Assurance Stage

This section describes the decision whose record is the subject of this document. It is descriptive context for the audit-record fields of Section 6, not a mandated enforcement architecture: a deployment records these fields wherever it makes such a decision, however its own enforcement is structured.¶

An implementation that records these fields evaluates, for an incoming agent action request and before the action proceeds:¶

1. a risk signal for the request (the method of computing it is out of scope);¶

2. an assurance requirement derived from that signal, expressed in a form appropriate to the request's action-domain (Section 5);¶

3. whether the requirement is satisfied; and¶

4. a decision that gates the action at an enforcement point (a PEP in the sense of [RFC2748]): the action proceeds only if the requirement is met, and cannot bypass the stage.¶

The decision, its inputs, and its outcome are recorded (Section 6).¶

The requirement is "risk-graded": it may take one of several ordered levels rather than a single permit/deny outcome. The number of levels, and the mapping from risk signal to level, are out of scope for this document; a deployment MAY use as few as two levels, but an interoperating consumer MUST NOT assume only two levels exist.¶

4. Autonomy-Asymmetry Control Law

This section describes the autonomy-asymmetry control law: a reference design for the mapping from risk signal to assurance requirement. It is informative. The control law captures the property that loss of autonomy SHOULD be immediate and recovery of autonomy SHOULD be gradual — an asymmetry that reflects the greater cost of a missed safety event versus the modest cost of temporarily elevated assurance.¶

5. Domain-Appropriate Assurance Forms

The assurance requirement is one primitive expressed differently per domain. This document defines three domains and leaves the set extensible.¶

Human-authentication domain:

The requirement is an authentication depth, ranging from a light single-factor re-authentication for low-risk actions to a full multi-modal re-authentication for high-risk or non-repudiable actions.¶

Agent-action-execution domain:

The requirement is an oversight degree, ranging from autonomous execution, through draft-then-human-confirmation, to block-pending-human-authorization.¶

Computational-resource-selection domain:

The requirement is a selection rigor, ranging from single-resource selection to multi-resource cross-validation or consensus of a specified depth.¶

An implementation MAY define additional domains. A domain definition MUST specify the ordered set of assurance levels for that domain and how a level is determined to be satisfied.¶

6. Audit Record Fields

For each evaluation, an implementation SHOULD record, in the deployment's audit-record model:¶

action_ref:

A reference to the action request being gated.¶

risk_signal:

The risk value that drove the decision, and a reference to the method or metric identity that produced it (the metric's internals are out of scope).¶

domain:

The action-domain of the request.¶

assurance_required:

The derived requirement, in the domain-appropriate form of Section 5, including the level within that domain's ordered set.¶

assurance_outcome:

Whether the requirement was satisfied, and by what evidence reference.¶

pre_action:

A boolean or structural indicator that the evaluation preceded and gated the action (as opposed to a post-hoc log).¶

verified_human_root:

OPTIONAL. A reference per [I-D.zagarella-verified-human-root] binding the decision to an accountable human.¶

control_law_state:

OPTIONAL. Where the autonomy-asymmetry control law (Section 4) is in use: the assurance level at which the control law stood when this decision was made, the most recent VERIFY-phase outcome that influenced it, and the iteration counter of the current packet. This field makes the control law's state observable in the audit record without requiring out-of-band state.¶

These fields let an audit consumer answer, for any recorded action: what was the assessed risk, what assurance did policy therefore require, was it met, and was the check made before the action — uniformly across domains. Where the control_law_state field is present, a consumer can additionally answer: at what autonomy level was this action authorized, and what sequence of VERIFY outcomes drove the system to that level.¶

7. Interoperability Considerations

Two systems interoperate at this interface when they agree on: the set of action-domains in use; for each domain, the ordered set of assurance levels and their satisfaction criteria; and the audit-record field names and encodings. This document fixes the field semantics and the three baseline domains; the concrete level sets and encodings are expected to be profiled by the consuming ecosystem or working group.¶

Systems that evaluate policy only at action boundaries can expose a degenerate two-level grading through this interface without adopting continuous grading internally; consumers therefore MUST treat two-level behavior as a valid special case and MUST NOT infer richer semantics than a producer advertises.¶

Where the autonomy-asymmetry control law (Section 4) is used, the thresholds (lower_threshold, upper_threshold) and hysteresis window MUST be agreed between producer and consumer when they are used to derive the assurance requirement exposed in the audit record. A consumer that does not understand the control_law_state field MUST ignore it rather than fail; the field is OPTIONAL.¶

8. Security Considerations

Non-bypass:

The value of a pre-action stage is that it cannot be skipped. An implementation MUST ensure the action does not proceed unless the assurance requirement is satisfied; a stage that can be bypassed provides no assurance regardless of how it grades risk.¶

Risk-signal integrity:

The assurance requirement is only as trustworthy as the risk signal. Manipulation of the signal manipulates the requirement. Deployments SHOULD protect the integrity and provenance of the risk input; this document's audit fields record which metric was used so that later review can detect anomalous inputs.¶

Record integrity:

Assurance decisions are security-relevant events. The audit records SHOULD be carried in a tamper-evident record model so that the "required vs met" history cannot be silently rewritten.¶

Binding to human root:

Where the verified_human_root field is used, the security considerations of [I-D.zagarella-verified-human-root] apply to that binding, including its replay and revocation properties.¶

Downgrade:

An attacker who can lower the derived requirement (by influencing the risk signal or the domain classification) weakens assurance without triggering a deny. Deployments SHOULD monitor for anomalous downward drift in required assurance relative to comparable historical actions.¶

Control-law manipulation:

The autonomy-asymmetry control law (Section 4) relies on the integrity of the VERIFY-phase outcomes read from the seal chain. An attacker who can inject false passing outcomes into the seal chain can accelerate slow-up and obtain a lower assurance requirement without earning it. Implementations MUST verify chain integrity (Section 4.4) before treating any seal as a sensor reading. A chain with a failed integrity check MUST be treated as if all VERIFY outcomes in it were failures.¶

Hysteresis parameter choice:

A hysteresis window that is too short weakens the slow-up property; a window that is too long makes recovery from a false VERIFY failure disproportionately slow. Deployments SHOULD tune the hysteresis window to the consequence severity of the action class being governed.¶

9. IANA Considerations

This document has no IANA actions. A future version may register the audit-record field names of Section 6, subject to working-group adoption.¶

10. References

Design Rationale

The interface deliberately stops at the decision boundary. Whether risk is computed by a single model, an ensemble, a calibrated cross-model metric, or a human-tuned heuristic is a deployment choice; whether the mapping from risk to requirement is a step function, a continuous curve, or a control loop with feedback is likewise out of scope for the normative parts of this document.¶

The autonomy-asymmetry control law (Section 4) is described here as a reference design precisely because it captures a pattern — asymmetric response to failure — that is well-motivated in safety-sensitive agent systems. The fast-down property reflects that the cost of a missed safety event is typically much higher than the cost of temporarily elevated assurance. The slow-up property with hysteresis reflects that trust, once lost, should be re-established by demonstrated performance rather than by the mere passage of time or a single success.¶

The choice of the phase-seal chain as the sensor (Section 4.4) reflects a design constraint: the control law's state must be fully observable from the chain record alone, without requiring a separate control-state store. This makes the loop's history auditable by any third party with access to the chain, which is consistent with the tamper-evident record requirement of Section 8.¶

Changes from -00

• Added Section 4 (Autonomy-Asymmetry Control Law): the fast-down/slow-up control law, VERIFY-pass-rate coupling, the iteration governor as the per-packet instance, and the phase-seal chain as sensor.¶

• No change to the assurance interface or the audit-record fields.¶

Acknowledgments

This document responds to the direction of the proposed AUDIT work on authorization state and provenance, and is designed to compose with [I-D.zagarella-verified-human-root].¶

Author's Address