Internet-Draft verification.* family June 2026
Krausz Expires 8 December 2026 [Page]
Workgroup:
Network Working Group
Internet-Draft:
draft-krausz-verification-state-00
Published:
Intended Status:
Informational
Expires:
Author:
J. Krausz
TK Collective

The verification.* Constraint Family: Pre-Action Fail-Closed Gates for AI Agent Decisions

Abstract

This document specifies the verification.* constraint family --- a pre-action, fail-closed gate primitive for AI agent decisions, sibling in shape to the environment.* family used in Verifiable Intent specifications. A verification.* receipt is a JWS-signed artifact ([RFC7515]) carrying a canonical input, a derived binary act/halt output, and a versioned mapping identifier that binds them. A relying party recomputes the gate locally from signed primitives under the named mapping; the verifier never trusts the issuer's runtime. This shape provides decision explainability and traceability evidence aligned with EU AI Act Article 12 record-keeping obligations and with the Decision Explainability tier of Anthropic's Zero Trust for AI Agents framework [ANTHROPIC-ZT]. The format is forward-compatible across mapping revisions: receipts signed under one mapping ID remain verifiable as correct-under-that-mapping after newer mappings ship.

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 8 December 2026.

Table of Contents

1. Introduction

AI agents increasingly take economic and operational actions on probabilistic input. A customer-service agent posts content; a procurement agent commits to a supplier; a research agent files a regulatory report. Each action is preceded by a factual claim the agent believes to be true. Today, the infrastructure verifying whether the claim is true before the agent acts on it is fragmented across operator logs (after-the-fact), confidence scores (soft signals), and ad-hoc validation pipelines (not interoperable).

This document specifies the verification.* constraint family: a pre-action, fail-closed gate primitive shaped identically to the environment.* family used in Verifiable Intent specifications [VINTENT], but applied to probabilistic predicates rather than boolean state. A verification.* receipt asserts, for a specific claim under a specific ruleset at a specific moment, whether an agent SHOULD proceed or halt. The receipt is JWS-signed [RFC7515] and structured such that a relying party can recompute the gate decision locally from signed primitives, without trusting the issuer's runtime.

The four properties this primitive needs to satisfy together:

  1. Pre-action. The verification happens before the action.
  2. Fail-closed. Anything ambiguous halts; "impossible, not tedious."
  3. Third-party recomputable. The verifier never trusts the issuer's runtime; signature binds inputs, outputs, and mapping identifier together.
  4. Forward-compatible. Receipts signed under one ruleset remain verifiable as correct-under-that-ruleset even after newer rulesets ship.

Property 3 is the load-bearing one and the one most existing primitives stop short of. SCITT [SCITT] signs receipts (necessary but not sufficient). W3C Verifiable Credentials Confidence Method [VC-CM] exposes confidence as a verifiable property (useful but not gating). RATS [RFC9334] provides the Evidence->Verifier->AR->RP vocabulary (matched here). None fully recompose the gate verdict from signed inputs. Property 4 is the one that becomes a footgun if skipped: rule changes silently invalidate or mis-verify old artifacts.

This document is complementary to security frameworks for AI agent deployment, including the Zero Trust framework for AI Agents published by Anthropic in May 2026 [ANTHROPIC-ZT], which names decision explainability as non-optional for regulated AI systems. The verification.* constraint family is one credible artifact shape for satisfying that explainability requirement.

1.1. Requirements Language

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.

1.2. Scope

In scope:

  • Gate semantics for pre-action verification of factual claims.
  • Receipt envelope format and required claims.
  • Calibration anchor requirements and discipline.
  • Multi-axis freshness model.
  • Verification protocol for relying parties.

Out of scope:

  • Signing key management policy (defers to JWKS standards).
  • Specific verifier implementations or model architectures.
  • Claim-extraction techniques (how a claim is identified from agent input/output).
  • Adversarial-probe methodology specifics.

2. Terminology

Verifier --- Entity issuing a verification receipt.

Relying Party (RP) --- Agent or downstream system consuming the receipt at a gate.

Gate --- The decision point in the RP's code path where act / halt is consumed.

Calibration anchor --- The dataset and seed defining confidence semantics for the verifier's pipeline.

Multi-axis freshness --- Signature, calibration, and evidence each carry independent validity periods.

Verdict (raw) --- One of supported, refuted, unverifiable, unknown. The underlying truth label.

Recommendation (canonical) --- One of confident_supported, un_probed_not_cleared, vulnerable_supported, weak_supported, refuted, unverifiable, error. Derived from primitives under the named mapping.

Gate (derived) --- act or halt. Derived from recommendation under the named mapping.

Mapping document --- A published, immutable document specifying how recommendations are derived from primitives and how gates are derived from recommendations. Identified by a stable string identifier (e.g., v0.3.0-2026-05-30).

3. The verification.* Constraint Family

3.1. Sibling-not-Member Relationship to environment.*

Table 1
Property environment.* verification.*
Predicate Boolean (state matches or does not) Probabilistic in [0,1]
Threshold ownership Oracle-defined, fixed-semantic Calibration-anchored, mapping-versioned
Freshness Single TTL Multi-axis
Gate shape Binary halt Binary halt

Both families produce a binary fail-closed gate primitive. They differ in predicate shape (boolean vs probabilistic) and threshold provenance (fixed vs calibration-anchored). This difference is sufficient to warrant a sibling family rather than a member entry under environment.*: probabilistic predicates require calibration discipline that boolean predicates do not, and the version-binding mechanism described in Section 4 is specific to mappings between probabilistic primitives and binary outputs.

3.2. Composition Rules

Conjunction: Multiple verification.* constraints in a single mandate combine with AND (orthogonal+conjunctive). All constraints MUST resolve to act for the action to proceed.

Ordering with environment.*: environment.* constraints MUST short-circuit before verification.* constraints per [ENV-STATE] §5.5. Environment evaluation is typically cheaper (no oracle roundtrip) and a failed environment constraint moots the verification. See also Section 9 for the non-interference security property this ordering provides.

4. Receipt Format

4.1. JWS Envelope

Verification receipts MUST be issued as JSON Web Signatures (JWS) [RFC7515] in compact serialization.

  • alg: ES256 mandatory. Other algorithms optional and SHOULD follow IANA JOSE Algorithms registry.
  • kid: MUST resolve to a key in the issuer's published JWKS [RFC7517] at /.well-known/jwks.json unless an alternate JWKS location is published in the issuer's metadata.
  • typ: verification-receipt+jws.

4.2. Payload Claims --- Canonical, Derived, Mapping-Bound

The payload signs three structurally distinct field groups plus their bindings.

Canonical input (signed primitives):

  • v_verdict (string, REQUIRED) --- One of supported, refuted, unverifiable, unknown.
  • v_confidence (number, REQUIRED) --- Float in [0, 1].
  • v_adversarial_result (string, REQUIRED) --- One of resilient, vulnerable, not_checked.

Canonical derived (signed):

  • v_recommendation (string, REQUIRED) --- One of confident_supported, vulnerable_supported, weak_supported, refuted, unverifiable, error. Derived deterministically from the canonical input under the named mapping.

Derived output (signed):

  • v_gate (string, REQUIRED) --- One of act, halt. Derived from v_recommendation under the named mapping.

Binding (signed):

  • v_gate_mapping (string, REQUIRED) --- Stable identifier of the published mapping document used at issuance (e.g., v0.3.0-2026-05-30). The mapping document is immutable after publication; future revisions ship as new identifiers.
  • v_gate_mapping_hash (string, REQUIRED) --- SHA-256 hex digest of the canonical serialization of the mapping document identified by v_gate_mapping. MUST be present in every receipt. Receipts MUST bind to a content-addressed mapping; the absence of v_gate_mapping_hash is a malformed-receipt condition and MUST result in gate decision halt.

Provenance (signed, not gating):

  • v_method (string, OPTIONAL) --- Self-describing verifier pipeline identifier.
  • v_calibration (object, OPTIONAL) --- Calibration anchor metadata. See Section 7.
  • v_sources_used (array of strings, OPTIONAL) --- Source labels actually consulted in this evaluation.
  • v_evidence (string, OPTIONAL) --- URI pointer to evidence corpus, if persisted.

Standard JWT claims [RFC7519]:

  • iss (REQUIRED), sub (RECOMMENDED), iat (REQUIRED), exp (REQUIRED), nbf (OPTIONAL).

4.3. Verification Protocol

A relying party verifying a receipt MUST execute the following sequence and treat any failure as a malformed receipt with gate decision halt:

1. Verify JWS signature against issuer's published JWKS (RFC 7515).
2. Resolve v_gate_mapping -> fetch the named immutable mapping document.
   MUST verify the SHA-256 digest of the fetched document matches
   v_gate_mapping_hash (hex-encoded). Digest mismatch -> malformed;
   gate decision = halt. This binding is mandatory; receipts without
   v_gate_mapping_hash are malformed.
3. Recompute candidate_recommendation from
   (v_verdict, v_confidence, v_adversarial_result)
   using the mapping's rules and the threshold recovered from the
   mapping document.
4. Confirm candidate_recommendation == v_recommendation.
5. Compute candidate_gate = mapping(v_recommendation).
6. Confirm candidate_gate == v_gate.
7. Verify exp/nbf against current time, subject to clock tolerance
   per Section 6.
8. If all match -> receipt is valid AND internally consistent under
   the named mapping. Any mismatch -> malformed; gate decision = halt.

The relying party never trusts the issuer's runtime to have applied the mapping correctly. The signature binds inputs, outputs, and mapping identifier together; the verifier recomputes locally.

4.4. Claim Binding

The verified claim MUST be bound to the receipt via v_claim:

"v_claim": {
  "text": "string (OPTIONAL when caller marks input as PII-sensitive)",
  "hash": "hex-encoded SHA-256 of canonicalized claim text (REQUIRED when text is omitted)"
}

Hash-only mode permits PII-sensitive verification while preserving the receipt's auditability --- the verifier can confirm a future caller's claim hash matches the issued receipt's bound hash without exposing the claim text.

4.5. Receipt Envelope Example

The following is a non-normative example of a receipt envelope payload illustrating the mandatory v_gate_mapping_hash binding alongside v_gate_mapping:

{
  "iss": "https://verifier.example.com",
  "sub": "claim:sha256:a3f1...",
  "iat": 1748649600,
  "exp": 1748736000,
  "v_verdict": "supported",
  "v_confidence": 0.91,
  "v_adversarial_result": "resilient",
  "v_recommendation": "confident_supported",
  "v_gate": "act",
  "v_gate_mapping": "v0.3.0-2026-05-30",
  "v_gate_mapping_hash": "sha256:1ad513cd0cfcc1fcd78136375268ba85c50cc267de8d9f92a9a3e61f5d672288",
  "v_claim": {
    "hash": "a3f1c2d4e5b6789012345678abcdef0123456789abcdef0123456789abcdef01"
  }
}

The v_gate_mapping_hash field carries the hex-encoded SHA-256 digest of the canonical serialization of the mapping document identified by v_gate_mapping. This field MUST be present in all conforming receipts; its absence is a malformed-receipt condition.

4.6. Version-Binding Rationale

Gate-derivation rules will evolve. A receipt issued under mapping v0.3.0-... MUST remain verifiable as correct-under-v0.3.0 after a newer mapping ships. The mapping identifier in v_gate_mapping is the binding that makes this true: a verifier fetches the same mapping document the issuer used at issuance, regardless of newer revisions. Receipts never silently re-verify to a different gate.

Mapping document publishers MUST treat published mappings as immutable. New rules ship under new mapping identifiers. Errata (typographical corrections only) MAY be appended to a published mapping document's metadata, but the normative rule tables MUST NOT change after first publication.

5. Binary-Halt Gate Semantics

5.1. Decision Table

The reference mapping v0.3.0-2026-05-30 defines the following decision table. Future mappings MAY tighten or extend this table; the structural shape (binary act/halt output derived from canonical recommendation) MUST be preserved.

Table 2
v_verdict v_confidence vs threshold v_adversarial_result v_recommendation v_gate
supported >= threshold resilient confident_supported act
supported >= threshold not_checked un_probed_not_cleared halt
supported (any) vulnerable vulnerable_supported halt
supported < threshold resilient or not_checked weak_supported halt
refuted (any) (any) refuted halt
unverifiable or unknown (any) (any) unverifiable halt
(error condition) N/A N/A error halt

5.2. Threshold Rules

The confidence threshold used to derive v_recommendation is specified in the named mapping document, not in the individual receipt. Receipts reference the mapping by v_gate_mapping identifier, and the threshold is recovered from that mapping. This ensures that the threshold is auditable, versioned, and consistent across all receipts issued under the same mapping: two receipts under the same mapping ID MUST gate against the same threshold; threshold changes require a new mapping version with a new ID.

For a strict fail-closed gate, un-probed adversarial state is not equivalent to resilient --- it represents uncertainty about a dimension that can be exploited by adversarial input. Per the fail-closed property, uncertainty MUST halt. A claim may be confidently supported on its face, but if adversarial probing was not performed, the confidence applies only to the base claim, not to the claim under adversarial pressure. The un_probed_not_cleared recommendation reflects this: the gate treats the absence of probing as a distinct unresolved risk, not as a cleared risk.

Relying parties MAY require a higher threshold mapping for their own gate policies by requiring receipts under a different mapping ID that specifies a higher threshold. They MUST NOT treat a receipt as conformant under a mapping that specifies a different threshold than the one in that mapping document.

5.3. Fail-Closed Mandate

If a receipt is missing, malformed, expired, signature-invalid, or its mapping ID cannot be resolved, the relying party MUST treat the gate decision as halt. Implementations MUST NOT default to act under any error condition. This is the "impossible, not tedious" design principle: friction-based controls bypass under adversarial pressure; hard barriers do not.

6. Multi-Axis Freshness

6.1. Three Independent Axes

Verification receipts have three independent staleness axes, each with its own validity window:

Table 3
Axis Field Remediation
Signature exp Key rotation; verifier MUST reject expired signatures
Calibration v_calibration.valid_until Recalibrate verifier pipeline against current anchor
Evidence v_evidence.valid_until (OPTIONAL) Re-fetch evidence; usually shorter window than calibration

6.2. Staleness Semantics

  • Stale signature --- Receipt is invalid; relying party MUST treat as halt.
  • Stale calibration --- Receipt SHOULD be re-evaluated; relying party MAY honor as soft-stale per its own policy. Verifier issuers SHOULD publish recalibration cadence in metadata.
  • Stale evidence --- Receipt SHOULD be re-evaluated. MAY be honored for retrospective audit purposes (e.g., post-incident review of a past action) but MUST NOT be honored as a gate signal for new actions.

Each axis has a different remediation path; collapsing them into a single TTL would force the most-restrictive cadence to govern all three.

7. Calibration Anchor Requirements

7.1. Reproducibility Mandate

Verifier issuers MUST publish:

  • Anchor dataset name and version.
  • Anchor seed (or seeded run identifier).
  • Anchor scoring methodology.

Verifier issuers SHOULD publish a reference reproduction harness under a permissive open-source license (Apache 2.0, MIT, or BSD). A working reference implementation reproducing 57.6% on AVeriTeC [AVERITEC] is available at https://github.com/TKCollective/agentoracle-eval-harness as one example of conformance.

7.2. Drift Detection

Verifier issuers SHOULD publish a recalibration cadence in the JWKS metadata or the verifier's well-known endpoint. The v_calibration.valid_until field in receipts SHOULD reflect this cadence. Relying parties consuming receipts with stale calibration SHOULD log and surface the staleness rather than silently accept.

9. Security Considerations

9.1. Composition Non-Interference with environment.* Family

When verification.* composes alongside environment.* on the same mandate, the ordering specified in §3.2 (environment.* short-circuits before verification.*) MUST hold. This document inherits that ordering directly from [ENV-STATE] §5.5, which specifies that environment.* is not subject to ordering preemption by any other constraint family. A verification.* ACT outcome MUST NOT mask an environment.* HALT outcome; if environment.* halts on any constraint, that halt is final regardless of the verification.* state. Implementations MUST evaluate environment.* to its terminal state before evaluating any verification.* constraint, and a verification.* state SHALL NOT be reached if environment.* has already halted.

9.2. Other Security Considerations

Key compromise. Verifier issuers MUST rotate JWKS keys on a published cadence. Receipts signed under a compromised key remain verifiable against historical key state during the rotation horizon. RPs MUST honor key revocation lists where published.

Replay attacks. Receipts include iat and exp. RPs MUST verify both against current time, subject to clock tolerance. Claim-binding via v_claim.hash provides resistance to receipt-reuse against a different claim payload.

Confidence inflation attacks. A verifier issuer who inflates v_confidence to push more receipts to act would diverge from their published calibration anchor. The reproducibility mandate (Section 7) makes this detectable. RPs SHOULD periodically sample receipts against the issuer's published harness.

Selective disclosure. Hash-only claim binding (Section 4.4) permits PII-sensitive verification without exposing claim text in the receipt envelope.

Downgrade attacks. A future relying party MUST NOT accept a v0.3-spec receipt against a v0.4-spec gate. The receipt format version is implicitly bound by v_gate_mapping; mismatches between expected and signed mapping IDs are malformed-receipt conditions per Section 4.3.

Mapping document tampering. Receipts MUST bind to a content-addressed mapping via the v_gate_mapping_hash field (SHA-256). Mapping documents MUST be hosted such that the SHA-256 digest of the canonical serialization is stable and independently verifiable (e.g., at a content-addressed URL or via a git-tagged manifest). Verifier implementations MUST verify the SHA-256 digest of the fetched mapping document against v_gate_mapping_hash before use; a digest mismatch MUST result in gate decision halt. A relying party that cannot independently verify the mapping document's hash MUST treat the receipt as malformed.

10. IANA Considerations

This document requests the following IANA registrations:

Media type: verification-receipt+jws (to be registered in the IANA Media Types registry following the procedures of RFC 6838).

JWT claim names: v_verdict, v_confidence, v_adversarial_result, v_recommendation, v_gate, v_gate_mapping, v_gate_mapping_hash, v_method, v_calibration, v_sources_used, v_evidence, v_claim (to be registered in the JSON Web Token Claims registry).

Well-known URI: Verification issuers using HTTPS SHOULD publish their JWKS at /.well-known/jwks.json per existing [RFC7517] Section 4.7 conventions. No new well-known URI is requested.

The verification.* and environment.* constraint families are related sibling namespaces. Coordination on a joint constraint-family registry shared with the environment.* family ([ENV-STATE]) is deferred to all interested parties (Krausz, Borthwick, Msebenzi) for resolution outside the scope of this document. This document does not commit to a specific registry structure, nor does it assert agreement from the environment.* authors on any registry arrangement. The cross-referencing between verification.* and environment.* as related constraint families stands as documented, independent of any future registry decision.

11. Informative References

[ANTHROPIC-ZT]
Anthropic, "Zero Trust for AI Agents: A security framework for deploying autonomous AI agents in the enterprise", PDF, 36 pages, .
[AVERITEC]
Schlichtkrull, M., "AVeriTeC: A Dataset for Real-World Claim Verification with Evidence from the Web", .
[ENV-STATE]
Borthwick, D. and M. Msebenzi, "Verifiable Intent --- environment.* Constraint Family", Work in Progress, Internet-Draft, draft-borthwick-msebenzi-environment-state-00, , <https://datatracker.ietf.org/doc/html/draft-borthwick-msebenzi-environment-state-00>.
[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <https://www.rfc-editor.org/info/rfc2119>.
[RFC7515]
Jones, M., Bradley, J., and N. Sakimura, "JSON Web Signature (JWS)", RFC 7515, DOI 10.17487/RFC7515, , <https://www.rfc-editor.org/info/rfc7515>.
[RFC7517]
Jones, M., "JSON Web Key (JWK)", RFC 7517, DOI 10.17487/RFC7517, , <https://www.rfc-editor.org/info/rfc7517>.
[RFC7519]
Jones, M., Bradley, J., and N. Sakimura, "JSON Web Token (JWT)", RFC 7519, DOI 10.17487/RFC7519, , <https://www.rfc-editor.org/info/rfc7519>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <https://www.rfc-editor.org/info/rfc8174>.
[RFC9334]
Birkholz, H., Thaler, D., Richardson, M., Smith, N., and W. Pan, "Remote ATtestation procedureS (RATS) Architecture", RFC 9334, DOI 10.17487/RFC9334, , <https://www.rfc-editor.org/info/rfc9334>.
[SCITT]
IETF SCITT Working Group, "An Architecture for Trustworthy and Transparent Digital Supply Chains", Work in Progress, Internet-Draft, draft-ietf-scitt-architecture, , <https://datatracker.ietf.org/doc/html/draft-ietf-scitt-architecture>.
[VAP-FW]
Kamimura, K., "Verifier Abstraction Pattern Framework", Work in Progress, Internet-Draft, draft-kamimura-vap-framework, , <https://datatracker.ietf.org/doc/html/draft-kamimura-vap-framework>.
[VC-CM]
World Wide Web Consortium, "Verifiable Credentials Confidence Method", W3C Working Draft, .
[VINTENT]
Verifiable Intent Working Group, "Verifiable Intent --- Constraint Type Definitions and Validation Rules", Work in Progress, Internet-Draft, v0.1-draft, , <https://datatracker.ietf.org/doc/html/v0.1-draft>.

Author's Address

Joe Krausz
TK Collective
Los Angeles, CA
United States of America