Internet-Draft Constitutional Action Gate November 2026
Watts Expires 18 May 2027 [Page]
Workgroup:
Network Working Group
Internet-Draft:
draft-watts-constitutional-action-gate-00
Published:
Intended Status:
Experimental
Expires:
Author:
D. O. Watts
GoodShyt Group Inc.

Constitutional Action Gate Profile for Agent-Mediated Internet Operations

Abstract

This document proposes CAG-1, an experimental, constrained decision profile for authorizing agent-mediated consequential effects at an enforcement point. It preserves the distinction between authenticated identity, qualified evidence, delegated permission, action approval, current epoch, and execution. The profile defines a strict JSON context, exact-effect binding, delegation attenuation, three-valued decision semantics, a replay/epoch admission example, and a synthetic conformance corpus. An ALLOW decision is not a new credential and does not prove execution, legal authority, scientific truth, or global revocation closure. This text is an unsubmitted working candidate; independent interoperability and security verification remain open.

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 18 May 2027.

▲

Table of Contents

1. Introduction

An autonomous service can correctly authenticate, verify a signed assertion, and observe a policy decision without possessing current authority for a particular consequential effect. Conflating these states can admit an unauthorized action even when every individual cryptographic check returns success.

This document defines CAG-1, an experimental, bounded, action-time admission-decision profile. CAG-1 consumes typed facts from authenticated, locally trusted adapters; checks principal, audience, scope, delegation attenuation, binding to the exact effect, witness status, policy version, revocation epoch, freshness, and an emergency stop; and derives ALLOW, DENY, or INDETERMINATE without conflating these outcomes.

The profile is a narrowly testable composition rule rather than a new OAuth grant, credential, HTTP header, token, revocation endpoint, trust framework, or law. The result is not a bearer credential and does not prove that a subsequent effect was committed, successful, safe, lawful, or beneficial.

This -00 text is an unsubmitted private working candidate. Its experimental status is a proposal, not recognition or endorsement by the IETF.

2. 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.

3. Terminology and Trust Boundary

Consequential effect: a write, permission change, monetary transfer, external message, credential issuance, irreversible resource use, or other operation for which current authority matters. The local application defines its effect types before evaluation.

Action gate: an enforcement point adjacent to a consequential sink that applies the CAG-1 predicate immediately before admission. An off-path evaluator is insufficient to claim protected sink enforcement.

Trusted adapter: a separately authorized component that authenticates an issuer, validates cryptographic proof, resolves an authoritative policy, checks audience and issuer trust, and supplies provenance-bound facts to the local gate. A client-controlled JSON statement claiming verification=pass MUST NOT be accepted as a trusted adapter result.

Witness: an independently sourced, typed assertion about evidence qualification or approval. A witness is not itself a transferable authorization grant.

Authority epoch: a monotonic generation identifier maintained by the relevant enforcement domain. Prior-epoch authority cannot be laundered into an admission under the current epoch.

Non-collapse: observed evidence, verified signatures, authenticated identity, issued permission, approved action, admission, execution, and successful outcome remain distinguishable states.

4. Scope and Non-Goals

CAG-1 specifies a local decision-context data model, deterministic decision semantics, and a bounded synthetic conformance corpus. The context is not a public credential, and transporting it over an untrusted channel is outside this version.

The profile does not define cross-domain issuer discovery, signing algorithms, a universal constitution, legal rights, consensus procedures, certification, agent consciousness, globally complete revocation, or a runtime side-effect commit protocol.

The profile does not claim that graph-based revocation closure has been established by a token-invalidation response. In particular, an older queued effect may still commit unless a sink enforces a suitable barrier.

5. CAG-1 Context Format

The input MUST be one UTF-8 JSON object no larger than 65536 octets, conforming to [RFC8259], with no duplicate member names. Unknown fields are rejected. All object shapes described below are exact. Arrays, maps, string lengths, and status enums have the finite limits described in this section.

All identifier strings, parameters, and JSON object member names in CAG-1 MUST be printable ASCII U+0020 through U+007E, without leading or trailing whitespace. Identifiers are nonempty and no longer than 128 characters. Effect parameter values MAY be up to 2048 characters. Numeric JSON values are prohibited; epochs are decimal uint64 strings with no leading zeros other than the string "0". Timestamp strings use YYYY-MM-DDTHH:MM:SSZ; this UTC-second profile excludes fractional seconds, offsets, and leap seconds. All timestamps denote UTC instants under [RFC3339].

The top-level object MUST contain exactly profile, action, authority, policy, witnesses, and runtime. profile MUST equal "cag-1". action MUST contain principal, audience, operation, resource, effect, effect_digest, and nonce. effect MUST contain operation, resource, and parameters (a map of zero to 32 strings). action.effect.operation and action.effect.resource are verified against the outer operation and resource in evaluation.

authority MUST contain chain, a nonempty array of at most eight grant hops. Each hop MUST contain issuer, subject, audience, operations (one to 32 distinct strings), resources (one to 32 distinct strings), not_before, expires_at, epoch, and verification. The verification status is one of pass, fail, unknown. The first hop is a grant from a trusted issuer (trust resolved by the adapter). For each later hop, issuer MUST equal its predecessor subject and the later hop MUST attenuate its predecessor scope, audience, time window, and epoch.

policy MUST contain digest (SHA-256), verification (pass/fail/unknown), require_evidence (Boolean), and require_approval (Boolean). A gate MUST NOT accept these policy bits from the requesting client as a substitute for its authenticated policy source.

witnesses MUST contain evidence with status, effect_digest, policy_digest; and approval with status, principal, audience, operation, resource, effect_digest, policy_digest. A status field is one of pass, fail, unknown. A required PASS witness must bind to the exact action fields and policy version. Optional witnesses MAY be UNKNOWN but policy verification remains mandatory.

runtime MUST contain now, fresh_until, current_epoch, revocation (clear/revoked/unknown), emergency_stop (clear/engaged/unknown), sink_barrier (ready/blocked/unknown), policy_digest, and adapter_verification (pass/fail/unknown). All runtime fields MUST come from a trusted gate snapshot, not the untrusted caller. fresh_until is an upper bound on the validity of that snapshot.

effect_digest, policy.digest, policy_digest, and witness digests MUST use a lowercase ASCII string of the form "sha256:" followed by exactly 64 lowercase hexadecimal digits. Action effect serialization is the JSON Canonicalization Scheme [RFC8785] restricted to the ASCII-only, no-numeric effect-object subset. Bytes are hashed with SHA-256. The supplied effect digest MUST match the locally recomputed digest.

6. Illustrative Context and Decision

The following fixture contains only synthetic principals and data. It illustrates the context shape; the PASS flags stand for facts returned by trusted test adapters and are not cryptographic attestations.

{
  "profile": "cag-1",
  "action": {
    "principal": "agent-alpha",
    "audience": "sink-alpha",
    "operation": "update",
    "resource": "record-01",
    "effect": {
      "operation": "update",
      "resource": "record-01",
      "parameters": {
        "field": "status",
        "value": "active"
      }
    },
    "effect_digest": "sha256:4d7a691ecac5ac94a47752b98bb316af23bfe354d56528f6766137e0be702ebe",
    "nonce": "nonce-0001"
  },
  "authority": {
    "chain": [
      {
        "issuer": "owner-alpha",
        "subject": "agent-alpha",
        "audience": "sink-alpha",
        "operations": [
          "read",
          "update"
        ],
        "resources": [
          "record-01",
          "record-02"
        ],
        "not_before": "2026-11-14T11:00:00Z",
        "expires_at": "2026-11-14T13:00:00Z",
        "epoch": "7",
        "verification": "pass"
      }
    ]
  },
  "policy": {
    "digest": "sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
    "verification": "pass",
    "require_evidence": true,
    "require_approval": true
  },
  "witnesses": {
    "evidence": {
      "status": "pass",
      "effect_digest": "sha256:4d7a691ecac5ac94a47752b98bb316af23bfe354d56528f6766137e0be702ebe",
      "policy_digest": "sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
    },
    "approval": {
      "status": "pass",
      "principal": "agent-alpha",
      "audience": "sink-alpha",
      "operation": "update",
      "resource": "record-01",
      "effect_digest": "sha256:4d7a691ecac5ac94a47752b98bb316af23bfe354d56528f6766137e0be702ebe",
      "policy_digest": "sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
    }
  },
  "runtime": {
    "now": "2026-11-14T12:00:00Z",
    "fresh_until": "2026-11-14T12:05:00Z",
    "current_epoch": "7",
    "revocation": "clear",
    "emergency_stop": "clear",
    "sink_barrier": "ready",
    "policy_digest": "sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
    "adapter_verification": "pass"
  }
}
{
  "profile": "cag-1",
  "conformant": true,
  "decision": "ALLOW",
  "admitted": false,
  "reasons": []
}

7. Deterministic Decision Algorithm

A CAG-1 evaluator MUST perform shape validation before computing any authorization result. Invalid input produces conformant=false, decision=null, admitted=false; INVALID is a parser/profile classification, not an authorization outcome.

For conformant input, the evaluator derives a vector of critical conditions. A condition is FALSE when an authoritative mismatch or failed check is established, UNKNOWN when the trusted source reports unresolved status, and TRUE only when an independent required check succeeds. Missing required facts are non-conformant if the field is absent, and INDETERMINATE if it exists with status unknown.

The evaluator MUST check: trusted-adapter and policy validation; exact policy digest; authoritative runtime freshness; effect-object operation, resource, and digest; actor binding to the terminal delegation subject; every hop verification, scope, time validity, and current epoch; chain attenuation; revocation clear; emergency stop clear; sink barrier ready; and, where policy requires, PASS evidence and approval witnesses bound to effect and policy.

The decision MUST be DENY if at least one critical condition is FALSE; otherwise it MUST be INDETERMINATE if at least one critical condition is UNKNOWN; otherwise it MUST be ALLOW. DENY and INDETERMINATE both prohibit admission. A conformant but unauthenticated fact stream MUST NOT yield ALLOW.

An implementation MUST NOT use a weighted average to compensate for failed or unknown critical controls. A model-generated explanation, evidence receipt, signature, prior test result, or declaration of competence MUST NOT independently change a status to PASS.

A result SHOULD identify each failed or unknown condition in deterministic, machine-readable reason codes. The reference implementation returns sorted reason-code strings; this code list is an experimental local artifact and establishes no IANA registry.

validate_strict_context(ctx) else conformant=false, decision=null
for each critical condition c in CAG-1: c := TRUE | FALSE | UNKNOWN
if any(c = FALSE): result := DENY
else if any(c = UNKNOWN): result := INDETERMINATE
else: result := ALLOW
admission := FALSE unless an authorized live sink atomically
  checks epoch and nonce, then reserves the permitted effect
never derive executed or successful from admission alone

8. Delegation, Scope, and Context Binding

For each hop i > 0: issuer_i = subject_(i-1), operations_i is a subset of operations_(i-1), resources_i is a subset of resources_(i-1), audience_i = audience_(i-1), not_before_i >= not_before_(i-1), expires_at_i <= expires_at_(i-1), and epoch_i = epoch_(i-1). The action principal MUST equal the last subject, and its operation, resource, and audience MUST be in every hop scope.

A verified signature on an out-of-scope hop does not override these checks. The first-hop authority issuer MUST be authorized by a separately configured trust and policy system; this profile does not define trust anchors.

9. Revocation, Replays, and Commit Boundaries

At a participating sink, current_epoch MUST be read from protected, monotonic local or coordinated authority state. A previously authorized snapshot MUST NOT be reusable after the corresponding epoch advances. Epoch rollback MUST be prohibited within the administrative scope.

A gate MUST ensure that for every protected effect, decision validation, authority-epoch comparison, nonce uniqueness, and the commitment to admission are atomic with respect to epoch advancement and nonce consumption. A denied or indeterminate decision MUST NOT execute the effect. A permit from an off-path library MUST NOT be treated as evidence that the sink performs this atomic step.

The nonce is unique within principal and audience at the local gate. A successfully admitted tuple (principal, audience, nonce) MUST be rejected on replay; keeping this state must cover the full relevant replay window. A resource server MUST also bind its actual outgoing effect bytes to the effect digest when the operation is executed.

Queue jobs, retries, delayed operations, child agents, and effect sinks not participating in a compatible barrier remain outside any claim of complete closure. An ACK from token revocation [RFC7009] is not by itself proof of closure. See the existing Watts revocation-closure I-D [I-D.watts-oauth-agent-revocation-closure] for broader pending-effect semantics.

If a queued operation cannot be fenced at commit time, implementations MUST NOT claim this profile prevents the operation. The reference SyntheticGate protects only its in-memory epoch/nonce reservation; it does NOT commit external effects.

10. Outcome and Receipt Semantics

An evaluator exposes separate fields profile, conformant, decision, admitted, and reasons. decision is ALLOW, DENY, INDETERMINATE, or null for nonconforming input. admitted remains false for pure evaluations. A stateful gate may set admitted=true only when it atomically reserves authorization under a current epoch and unused nonce.

ALLOW signifies that the specified checks passed at the named gate snapshot. It is not a bearer token, delegation, payment guarantee, scientific proof, or indication that a downstream operation has completed or succeeded. If a result is logged or signed, integrity of that log or receipt does not elevate its semantic meaning.

Historical approvals, revocations, and corrections SHOULD be append-only with signed or integrity-protected provenance according to the deployment threat model. Independent operational assurance requires separate evaluation of the actual effect sink.

11. Conformance Classes and Synthetic Tests

Syntax conformance demonstrates strict context parsing and finite field rules. Semantic conformance demonstrates identical TRUE/FALSE/UNKNOWN combination semantics. Vector-profile conformance means executing every declared fixture under an exact corpus digest. Deployment enforcement conformance requires independently observed sink-side atomicity and cryptographic source authentication. Independent reproduction requires independently controlled code and execution.

A claim at one level MUST NOT be promoted to a higher level merely because that lower level passed. A published test report SHOULD include source and corpus SHA-256 digests, implementation version, fixture counts and denominator, local versus independent status, fail cases, and bounded threat model.

The companion package defines 88 synthetic input vectors and tests pure decision semantics, nonce replay in a lock-protected reference class, monotonic epoch changes, and digest determinism. These tests are deliberately not evidence of production authorization, credential verification, multi-host linearizability, or independent reproduction.

12. Security Considerations

The dominant failure mode is fact-source spoofing: an attacker controlling the JSON document could set grant.verification=pass and runtime.revocation=clear. Implementations MUST supply these fields only from trusted, authenticated and policy-authorized adapters. When source trust cannot be established, status MUST NOT be PASS.

An attacker may perform digest or principal substitution, audience confusion, replay, policy downgrade, delegation scope amplification, current-epoch rollback, clock manipulation, stale-source acceptance, forged approval, lost revocation messages, concurrent races, or post-admission queue execution. This profile tests some decision boundaries but does not eliminate these classes of attack without operational controls.

Protected clocks, issuer/audience checks, source authentication, bounded freshness windows, credential rotation, a durable replay ledger where needed, failure-atomic sinks, and independent controls for administrative bypass SHOULD be deployed. Secrets MUST NOT appear in public test vectors, receipts, or decision logs.

Resource exhaustion is bounded by a 64KiB document cap, eight-hop delegation cap, finite arrays, and constrained parameter fields. Production parsers MUST additionally enforce depth, CPU, and memory limits appropriate to their environments.

Unknown revocation or sink state MUST NOT be promoted to CLEAR or READY. A deployment MUST treat unobserved downstream sinks and cached authorizations as excluded from any security guarantee. A signed decision record cannot replace an active authorization check.

The OAuth Security Best Current Practice [RFC9700] and sender-constrained-token mechanisms such as [RFC9449] remain relevant controls; CAG-1 does not supersede them.

13. Privacy Considerations

Principal, delegation, resource, policy, and action identifiers may expose sensitive structure. Deployments SHOULD use scoped opaque identifiers, minimize logged effect details, apply access controls, define retention/deletion obligations, and avoid correlating identifiers across unrelated tenants.

The companion corpus uses invented data and has no real credential values. Proprietary source research and real account identifiers MUST NOT be included in publicly shared fixtures without separately authorized disclosure and privacy review.

14. IANA Considerations

This experimental document requests no new IANA registries, HTTP fields, media types, OAuth grants, token types, error codes, or endpoint registrations.

15. Prior Art and Research Scope

This proposal reuses the established RFC mechanisms for JSON [RFC8259], canonical JSON [RFC8785], normative keywords [RFC2119] [RFC8174], and timestamp representations [RFC3339]. It composes with OAuth token revocation [RFC7009], Token Exchange [RFC8693], DPoP [RFC9449], OAuth security practice [RFC9700], rich authorization requests [RFC9396], and RATS attestation roles [RFC9334], without replacing any of them.

Existing Internet-Drafts [I-D.ietf-wimse-aims], [I-D.watts-oauth-agent-revocation-closure], and related Watts non-collapse and evidence research describe neighboring problems. The proposed candidate contribution is the exact-action conjunction and three-outcome admission profile with an explicit reference test corpus and local nonce/epoch reservation example. Novelty, interoperability, and independent security benefit remain unproven.

The internal AAAS constitutional architecture and GSG governance model motivated design constraints but do not acquire ratified legal, IETF, or scientific authority through inclusion here.

16. Experimental Questions and Falsifiers

H1 [CONJECTURE]: independent CAG-1 implementations given the same authenticated fact contexts produce identical conformant and decision values. This dies if any valid vector generates different results under independent implementations.

H2 [CONJECTURE]: a correctly integrated sink that performs current-epoch and nonce admission atomically rejects stale or replayed operations. This dies if a controlled concurrent revocation and effect-commit test observes an admitted old-epoch effect after the declared barrier.

H3 [CONJECTURE]: explicit three-valued outcomes reduce false authorization promotion under missing witnesses compared with a chosen binary baseline. This dies if a blinded, preregistered comparison fails to show improvement on an independently labeled corpus or exposes higher false-ALLOW rates.

These hypotheses do not by themselves claim an original scientific discovery. A local self-test is not an independent implementation. A result at a library boundary is not evidence of effect-sink correctness.

17. Changes in -00

Initial private experimental candidate. Defines constrained CAG-1 typed context, exact-action checks, three-valued decision semantics, bounded delegation, local atomic epoch/nonce reference, a synthetic test corpus, and explicit release gates.

18. Normative References

[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", RFC 2119, , <https://www.rfc-editor.org/info/rfc2119>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", RFC 8174, , <https://www.rfc-editor.org/info/rfc8174>.
[RFC8259]
Bray, T., "The JavaScript Object Notation (JSON) Data Interchange Format", RFC 8259, , <https://www.rfc-editor.org/info/rfc8259>.
[RFC8785]
Rundgren, A., "JSON Canonicalization Scheme (JCS)", RFC 8785, , <https://www.rfc-editor.org/info/rfc8785>.
[RFC3339]
Klyne, G., "Date and Time on the Internet: Timestamps", RFC 3339, , <https://www.rfc-editor.org/info/rfc3339>.

19. Informative References

[RFC7009]
Lodderstedt, T., "OAuth 2.0 Token Revocation", RFC 7009, , <https://www.rfc-editor.org/info/rfc7009>.
[RFC8693]
Jones, M., "OAuth 2.0 Token Exchange", RFC 8693, , <https://www.rfc-editor.org/info/rfc8693>.
[RFC9449]
Fett, D., "OAuth 2.0 Demonstrating Proof of Possession (DPoP)", RFC 9449, , <https://www.rfc-editor.org/info/rfc9449>.
[RFC9700]
Lodderstedt, T., "Best Current Practice for OAuth 2.0 Security", RFC 9700, , <https://www.rfc-editor.org/info/rfc9700>.
[RFC9396]
Lodderstedt, T., "OAuth 2.0 Rich Authorization Requests", RFC 9396, , <https://www.rfc-editor.org/info/rfc9396>.
[RFC9334]
Birkholz, H., "Remote ATtestation procedureS (RATS) Architecture", RFC 9334, , <https://www.rfc-editor.org/info/rfc9334>.
[I-D.ietf-wimse-aims]
IETF, "AI Identity Management System", , <https://datatracker.ietf.org/doc/draft-ietf-wimse-aims/>.
[I-D.watts-oauth-agent-revocation-closure]
IETF, "Revocation Closure and Authority Epochs for Agent Delegation", , <https://datatracker.ietf.org/doc/draft-watts-oauth-agent-revocation-closure/>.

Author's Address

Deonté O'dell Watts
GoodShyt Group Inc.
San Francisco, CA
United States of America