| Internet-Draft | Evaluator Authority and Lifecycle | November 2026 |
| Das | Expires 20 May 2027 | [Page] |
Autonomous agents increasingly rely on an evaluator to review plans, evidence and outcomes before work continues, yet an evaluation result is often mistaken for permission to act. This document describes a vendor-neutral architecture in which the evaluator is a functional role, not a product: it may be a rule engine, model, program, enclave, service or human, placed locally, remotely or on a shared platform. It separates six independent lifecycles, requires a protected task-specific appointment derived from resource-owner authority, admits evidence by source, role, freshness and independence, preserves mandatory dissent, and emits typed determinations such as HOLD, DISPUTED and UNKNOWN. Its central rule is that a valid determination is not an executable authorization: separate owner-controlled authorization and a native enforcement boundary remain necessary. Twenty-seven embodiments, a threat model, data model and examiner index are provided.¶
Technical source: Indian Provisional Patent Application 202631131650, dated 3 November 2026 (as reported by the applicant; corroborated only by an eFiling payment record showing CBR date 03/11/2026, not by the official acknowledgment, which the editor has not seen), titled "Systems and Methods for Location-Independent Evaluator Architectures, Dynamic and Persistent Evaluation Lifecycles, Protected Appointment, Evidence Verification, and Authority-Separated Decision Control in Autonomous and Distributed Systems". The source states 27 embodiments and 540 numbered claims (60 independent, 480 dependent). Foundational reference: PCT/IB2026/055615, filed 4 June 2026, published 16 July 2026 as WO 2026/150382 A2. See Section 26.¶
Where to read the full PCT filing: search "Sangam Das" on WIPO PATENTSCOPE; search PCT/IB2026/055615 (pctib202655615); or read the PCT file directly in Technical Disclosure Commons at https://www.tdcommons.org/dpubs_series/12542/.¶
The IPR disclosure for this document will be made at the appropriate time through the official IETF process and in accordance with IETF guidelines (see Section 26).¶
AI-assisted drafting was used. Readers must verify patent numbers, dates, claim text, product capabilities and standards references against primary sources. This is an individual proposal, not an IETF standard and not an endorsed position.¶
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 20 May 2027.¶
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.¶
Autonomous and multi-agent systems increasingly place an evaluator between a proposed plan and the point where an effect occurs. The evaluator may inspect evidence, estimate consequences, check policy, review an outcome, and recommend whether work should continue. Three questions are frequently merged: who may evaluate, what evidence the evaluator may rely on, and who may actually cause the effect. This document keeps them apart.¶
The document describes one architecture, drawn from a single source filing with 27 embodiments, in an individual, informational form. It does not define a new wire protocol; it specifies functional requirements, candidate record schemas and test conditions that existing mechanisms (see Section 17) can carry.¶
Central rule: a valid evaluation determination is not automatically an executable authorization.¶
Existing practice commonly shows one or more of the following gaps: an evaluator is chosen by the agent it reviews; independence is counted as accounts rather than control groups; a signed report is treated as truth about the physical or ledger outcome; an approval is reused after its appointment, round or policy generation has changed; a majority hides a mandatory dissent; a lost acknowledgment is retried and duplicates an effect; and a coordinator's DONE message is treated as closure while effect-capable work remains.¶
The key words in this document follow BCP 14 [RFC2119] [RFC8174] only where requirement language is used, and sparingly. Most statements are architectural descriptions rather than protocol requirements.¶
A functional role that assesses evidence, a plan, a consequence or an outcome under a bounded mandate and produces a determination. It holds no effectuation authority by virtue of the role.¶
A role that organizes tasks or agents and may carry out limited, authorized coordination. It is not an evaluator and not an owner.¶
The party or process that, under resource-owner policy, issues a bounded permission for a specific act.¶
The enforcement boundary at which an effect is actually admitted or withheld by the resource owner.¶
A protected, task-specific designation of an evaluator derived from the authority of the mandatory resource owners.¶
A typed, non-executable output of an evaluation, bound to task, round, policy generation and validity interval.¶
The set of credentials, hosts or principals that are in fact under one administrative or economic control.¶
A bounded set of reviews over a fixed evidence set that closes with completeness, independence and dissent checks.¶
Evidence originating at the resource owner or by protected direct observation about what actually occurred.¶
State held inside a protected enforcement environment that is not exported as a portable signed artifact.¶
The principal architecture is shown in Figure 1. Evaluation authority ends at the boundary where a non-executable determination is produced; continuation is separately authorized and any effect is admitted only by the owner-controlled native boundary.¶
Agents / human actors
|
v
Task, plan and evaluation request
|
v
Authority and resource-owner resolution
|
v
Evaluator eligibility + appointment verification
|
mandate valid? --NO--> DENY / HOLD
| YES
v
Authorized evidence intake
|
v
Source, role, task, freshness, truth checks
|
v
Policy-bounded evaluation
|
v
Review completeness, independence, dissent
| incomplete --> HOLD contested --> DISPUTED
| unresolved --> UNKNOWN / REQUEST_EVIDENCE
v
Typed NON-EXECUTABLE determination
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
evaluator authority ends above; owner authority begins below
v
Separate protected continuation authorization
| (owner policy; may refuse)
v
Owner-controlled native Finality Sink
| admits exactly the authorized act or withholds
v
Native outcome evidence (success / partial / UNKNOWN)
A failed mandate yields DENY or HOLD. Incomplete review yields HOLD. Contested review yields DISPUTED. Unresolved or missing evidence yields UNKNOWN or REQUEST_EVIDENCE. None of these is an authorization.¶
Between agents and the evaluator; between the evaluator and the authorization issuer; between the issuer and the native gate. No authority is inherited across a boundary Figure 2.¶
At the typed determination. It is a premise that owner policy may consider, not a token that a gate honors.¶
+-----------+ +-------------+ +-------------+ +-----------+
| Evaluator | | Supervisor/ | | Authorization| | Native |
| assesses | | coordinator | | issuer/owner | | gate/Sink |
| evidence | | organizes | | grants bound | | admits |
| | | tasks | | permission | | the effect|
+-----+-----+ +------+------+ +------+-------+ +-----+-----+
| boundary 1 | boundary 2 | boundary 3 |
+----- no authority is inherited across a boundary ------+
One host may run several roles; co-location inherits nothing.
An evaluator is defined by what it does, not by what it is. The same function may be an in-process component, a local protected service, a hardware or firmware element, a trusted execution environment, a container or virtual machine, a remote service, a persistent platform, a human, a committee, or a hybrid. Rule engines, policy engines, solvers, statistical models, generative AI and verification programs are all admissible placements. A single-component evaluator is not excluded merely because it is a single component.¶
Same function, different placements:
[in-process component] [local protected service] [TEE/hardware]
[container/VM] [remote service] [human/committee]
\ | /
+-- protected intake path (versioned evidence) --+
|
bounded determination
|
protected output path (admissibility checked)
Co-location is not non-independence; remote is not independent.
Independence = real control relations, conflicts, access.
Location independence is not absence of control. Co-location with an agent does not necessarily defeat independence if control relationships, evidence access and conflicts are separated; third-party hosting does not prove independence if the host is under common control or economic dependence. Independence is assessed on actual control relationships, evidence access, conflicts of interest and the ability to influence the outcome, under policy that remains owner-specific.¶
Created for one task or round, retired when appointment, round or validity ends. Strong isolation; setup cost per task Figure 4.¶
A long-lived protected service that serves many independent projects, each with its own mandate, rounds, validity and retention Figure 5.¶
A persistent platform hosting task-specific contexts with separate expiry, namespaces and revocation Figure 6.¶
eligible? -> appoint -> activate instance -> intake evidence
|NO |
v v
refuse evaluate -> determination
|
handoff <------------------+
|
appointment expires or is revoked --> instance retired
|
evidence retained (protected); old authority NOT reusable
Host (long-lived) |-- Project A mandate --> rounds A1, A2 (expiry A) |-- Project B mandate --> rounds B1 (expiry B) Separate data and evidence per project. A expires: host and B continue; no inheritance from A. Platform compromise: revoke all mandates, revalidate each.
Persistent platform (shared maintenance) +- context X (org 1 task) -- own expiry, own namespace +- context Y (org 2 task) -- own expiry, own namespace per-context revocation | context migration -> revalidate shared-platform compromise -> all contexts untrusted until re-established
A persistent evaluator does not inherit authority across projects. If the platform is compromised, every context is treated as untrusted until re-established. Moving a context between hosts requires revalidation.¶
Six lifecycles are independent. None is assumed to nest inside another.¶
Host or platform exists. State: provisioned, active, suspended, decommissioned.¶
A running evaluator process. State: created, running, retired.¶
Designation for a task. State: proposed, owner-approved, active, expired, revoked.¶
Bounded review over a fixed evidence set. State: open, admitting, closed, disputed.¶
Interval during which a determination is usable as a premise. State: valid, stale, withdrawn. It can end before the appointment ends.¶
How long evidence and audit material are kept. It may outlive all other lifecycles and grants no permission.¶
time ------------------------------------------------------> infrastructure |=================================........... runtime instance| |=========| (retired early) appointment | |================| eval round | |====| determination | |========| (may go stale first) evidence retain | |==========================... Round may close while task is active; evidence outlives all permissions; a valid signature does not extend validity.
Formal predicates (informal notation):¶
MandateValid(e,t,now) := Appointed(e,t) AND NOT Revoked(e,t) AND now < Expiry(e,t) AND Uses(e,t) < Limit(e,t)¶
RoundOpen(r,now) := State(r)=open AND now in Window(r)¶
DetValid(d,now) := Bound(d,task,round,policyGen) AND now in Validity(d) AND NOT Withdrawn(d) AND SnapshotValid(d)¶
EvidenceAdmissible(x,r) := L(x)=ok for all ten checks in Section 9¶
Negative tests: (a) a retired instance presents a valid-looking signature (rejected: L2); (b) evidence arrives after round closure (rejected: L4); (c) a determination is reused after policy generation changes (stale: L5); (d) evidence retention continues while authority is gone (permitted, grants nothing: L6); (e) appointment expires while the host continues (no authority for that task: L3).¶
Four control roles are distinguished. A single host may run several roles; co-location inherits no authority.¶
| Role | May do | May not do |
|---|---|---|
| Evaluator | assess evidence; produce typed determination | admit an effect; issue a passport; self-appoint |
| Supervisor / coordinator | organize agents and tasks within a bounded lease | grant native authority; act as evaluator for its own plan |
| Authorization issuer / owner authority | issue bounded permission per owner policy | bypass the native gate; reuse consumed authority |
| Native gate / Finality Sink | admit or withhold the actual act | accept a determination or approval as sufficient |
request -> resolve actual mandatory owners
| |
ambiguous? resolved
YES: HOLD |
owner consents (all mandatory)
| |
missing/forged complete
-> REJECT -> appointment issued
-> scope, expiry, use count
Agent self-declaration of neutrality is never an input.
A protected appointment is task-specific and derived from owner authority. A candidate agent's self-declaration is never an input. The appointment record is expected to include the following elements:¶
Failure to verify any mandatory element results in refusal or HOLD, never in permissive default.¶
Ten distinctions apply to every evidence item:¶
source -> origin authenticated? -> role eligible? -> content
integrity? -> bound to task/operation? -> fresh? -> policy
generation? -> independent (control group)? -> correlated?
each NO -> excluded or flagged
admitted evidence != outcome truth
outcome truth needs resource-origin evidence or protected
direct observation; agent success claims, tool acks, transport
acks, coordinator flags and evaluator reports are weaker.
Determinations are typed and non-executable.¶
| Type | Meaning | Executable? |
|---|---|---|
| APPROVED_FOR_CONSIDERATION | premise for owner authorization process | No |
| APPROVED | evaluator finds conditions met under mandate | No; premise only |
| DENIED | conditions not met | n/a |
| HOLD | incomplete mandate, evidence or review | No |
| DISPUTED | reviewers or sources contest | No |
| UNKNOWN | outcome or evidence unresolved | No; not DENIED, not SUCCESS |
| REQUEST_EVIDENCE | named additional evidence needed | No |
| REVISE | plan change suggested | No; needs recommitment |
| RENEGOTIATE | parties must renegotiate | No |
| HUMAN_REVIEW | escalate to human | No |
| SUBSTITUTE | replace an agent | No; needs fresh authorization |
| ABORT | stop the task | Not a native effect by itself |
UNKNOWN is not DENIED and not SUCCESS. Dissent is preserved with the determination; it is never silently dropped.¶
reviewers r1..rn -> map credentials -> control groups g1..gk diversity rule met on k (not n)? NO -> HOLD mandatory reviewer present? NO -> HOLD dissent recorded and shown? NO -> reject round contested outcome -> DISPUTED (not a majority vote) round closed -> determination eligible as a premise only
The following table is a crosswalk of the 27 embodiments to the claims of the source filing. Claim numbers are those of the filing as supplied and should be verified against the official record.¶
| Emb. | Code | Indep. claims | Dep. claims | Title (source) |
|---|---|---|---|---|
| E01 | CE-NJC-01 | 1-2 | 61-76 | Protected Multi-Agent Negotiation, Joint-Plan Commitment, and Independent Evaluator-Governed Continuation |
| E02 | CE-BFQ-02 | 3-5 | 77-100 | Bilateral Communication Freeze, Distributed Quiescence, and Protected Evaluation-State Convergence |
| E03 | CE-NSI-01 | 6-7 | 101-116 | Negotiation-State Integrity, Causal Message Containment, and Continuous Evaluation-Snapshot Validity |
| E04 | CE-INE-01 | 8-9 | 117-132 | Independent Multi-Agent Consequence Evaluator, Protected Evidence Admission, and Next-Actor Selection |
| E05 | CE-VCR-01 | 10 | 133-140 | Isolated Virtual Consequence Reconstruction and Non-Effective Scenario Execution |
| E06 | CE-RBL-01 | 11-12 | 141-156 | Risk, Benefit-Loss and Tail-Exposure Evaluation With Independent Domain Thresholds |
| E07 | CE-NAS-01 | 13-14 | 157-172 | Evaluator-Directed Next-Actor Selection and Controlled Next-Turn Instructions |
| E08 | CE-ACR-01 | 15-16 | 173-188 | Corrective Instructions, Protected Plan Revision, and Controlled Multiparty Renegotiation |
| E09 | CE-ARR-01 | 17-18 | 189-204 | Agent Role Reconfiguration, Bounded Delegation, and Protected Capability Transition |
| E10 | CE-SCR-01 | 19-20 | 205-220 | Selective Communication Resumption, Single-Turn Authority, Automatic Re-Freezing, and Inter-Phase Hold |
| E11 | CE-NMP-01 | 21-22 | 221-236 | Protected Next-Move Passport, Actor-Bound Issuance, Single-Use Finality Verification, and Outcome-Uncertainty Reconciliation |
| E12 | CE-PFS-01 | 23-24 | 237-252 | Protected Finality Sink Enforcement Before Communication, Privileged Execution, External Dispatch and Physical Actuation |
| E13 | CE-NSA-01 | 25-26 | 253-268 | Risk-Triggered Neutral Substitute-Agent Selection, Appointment, and Protected Reauthorization |
| E14 | CE-PWM-01 | 27-28 | 269-284 | Protected Pending-Work Escrow, Task-State Migration, and Substitute Reauthorization |
| E15 | CE-PCR-01 | 29-31 | 285-308 | Phased Execution, Protected Completion Receipts, and Verified Phase Advancement |
| E16 | CE-VFO-01 | 32-33 | 309-324 | Verified Feedback, Actual-Outcome Evaluation, and Protected Continuation |
| E17 | CE-EAC-01 | 34-35 | 325-340 | Evaluator Integrity, Source Provenance, Collusion-Resistant Review, and Protected Continuation |
| E18 | CE-PAR-01 | 36-38 | 341-364 | Cross-Domain Network Partitions, Unknown Outcomes, Durable Fencing, Migration and Safe Recovery |
| E19 | CE-SCL-01 | 39-40 | 365-380 | Protected Session Closure, Obligation Reconciliation, Native Domain Sealing, and Irreversible Authority Retirement |
| E20 | CE-JSA-01 | 41-42 | 381-396 | Verified Mandatory Approvers, Authenticated Hosting and Installation, and Revocable Supervisor Identity |
| E21 | CE-IEO-01 | 43-44 | 397-412 | Independent Cross-Domain Evaluator Appointment, Actual-Outcome Verification, Reviewer Accountability and Dispute Containment |
| E22 | CE-CTP-01 | 45-46 | 413-428 | Cross-Domain Trust and Policy Compatibility |
| E23 | CE-LDC-01 | 47-48 | 429-444 | Leaderless Cross-Domain Agent Coordination |
| E24 | CE-SER-01 | 49-50 | 445-460 | Self-Expiring Supervisor and Evaluator Authority, Authenticated Revocation, and Generation-Fenced Replacement |
| E25 | CE-DMG-01 | 51-52 | 461-476 | Dynamic Multi-Organization Membership and Principal-Independent Threshold Governance |
| E26 | CE-PCE-01 | 53-55 | 477-500 | Purpose-Bound Cross-Organizational Outcome Evidence |
| E27 | CE-MEF-01 | 56-60 | 501-540 | Multi-Domain Protected Execution Commitment, Partial-Outcome Reconciliation and Cross-Organizational Recovery |
Embodiments E01-E16 establish the multi-agent evaluation handoff. Each is summarized with the required profile fields. E12 is the embodiment that restates the foundational Finality Sink in the evaluator setting; this document does not claim a new execution-finality protocol.¶
Agent A <--proposals/counters--> Agent B
\ /
candidate plan v1 (bound to accepted state)
|
required principals commit? --NO--> keep negotiating
| YES
plan = committed but NON-EFFECTIVE
|
freeze + pending-work reconciliation OK?
NO -> HOLD YES -> evaluator handoff
Resource effects still need separate owner authorization.
freeze request -> each edge acknowledges?
partial ack -> NOT frozen (unresolved)
all ack -> inventory in-flight, callbacks, queues, tools
-> late message arrives? YES -> classify, quarantine
-> all accounted? YES -> quiescent snapshot
NO -> UNRESOLVED (no evaluation)
Applies only to interfaces actually under enforcement.
msg m1 -> m2 -> m3 (per-sender sequence, causal links)
snapshot S bound to plan v1 and evidence set E1
material change (new message / evidence / policy gen)
-> S invalidated -> determination on S becomes stale
-> re-snapshot -> re-evaluate
superseded or unaccounted message cannot alter plan silently
verified plan + authentic observations
-> isolated non-effective simulator (no egress)
-> scenarios + uncertainty
-> per-domain risk, tail exposure, thresholds
owner veto? YES -> blocked unknown fidelity -> qualified
result = evidence only; never a commit or proof of behavior
evaluator -> select actor (eligible?) -> bounded instruction
owner approves -> passport (actor, task, op, plan, phase,
epoch, policy, single use)
gate checks passport: wrong actor / replay / expired -> REJECT
one turn admitted -> turn ends -> automatic refreeze
unknown delivery -> reconcile (no blind retry)
risk trigger -> screen candidates (neutral, qualified, no conflict) -> de-appoint original -> appoint substitute escrow unfinished work (canonical, minimal) -> fence source -> destination accepts -> fresh authorization old credentials and consumed authority are NOT inherited
phase N done? -> resource-origin receipt verified?
NO -> block advancement (no skipping)
YES -> feedback evaluated (untrusted instructions contained)
-> outcome classified: success / partial / UNKNOWN
-> proposed next step -> separate authorization
coordinator "done" and agent self-report are not receipts
E01 (CE-NJC-01), source page 39. Protected Multi-Agent Negotiation, Joint-Plan Commitment, and Independent Evaluator-Governed Continuation¶
Independently governed agents agree a joint plan, but an agreement is not an authorization to act; continuation must pass an independent evaluator after a verified freeze.¶
Two or more agents under different administrative domains; mandatory authority principals; an independent neutral evaluator.¶
Versioned candidate plan bound to an accepted communication state and required principals; commitment records; freeze and pending-work records.¶
In: proposals, counteroffers, constraints, approvals. Out: a committed non-effective candidate plan and an evaluator-governed handoff.¶
negotiating, committed plan, frozen, evaluating, continuation proposed.¶
No commitment without required principals; no effect from a jointly signed plan; handoff withheld until freeze and reconciliation succeed.¶
Forged consent; missing mandatory owner; plan altered after commitment; one agent's self-declared authority.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
A jointly signed plan stays non-effective with respect to real resource operations; effects are admitted only at actual enforcement points.¶
Negative tests per Section 16; source reports: 16 and 33 test runs, and 42 tests passed on modeled code paths (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
First embodiment; basis for E02 to E16.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
joint plan commitment; multi-agent negotiation; evaluator handoff; non-effective plan¶
Independent claims 1, 2. Claim 1: Protected Multi-Agent Agreement and Independently Controlled Evaluation Handoff; Claim 2: Protected Multi-Agent Negotiation and Evaluator-Governed Continuation Method Dependent claims 61-76 (indexed, not individually analyzed here).¶
E02 (CE-BFQ-02), source page 173. Bilateral Communication Freeze, Distributed Quiescence, and Protected Evaluation-State Convergence¶
A declared stop is not a quiescent snapshot: in-flight messages, delayed callbacks, deferred tool calls and queue deliveries can still change state.¶
Two or more agents; freeze controller; independently administered enforcement interfaces.¶
Freeze epoch, per-edge acknowledgments, pending-work inventory, late-operation disposition records.¶
In: commitment to one plan version. Out: verified freeze with accounted pending work, or unresolved status.¶
freeze requested, partial acknowledgment, quiescent, unresolved.¶
Partial freeze acknowledgment is not a freeze; late or duplicated messages are not silently admitted; unresolved delivery blocks evaluation.¶
In-flight message after freeze; unfenced background work; different state generations; replayed deliveries.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Actual covered communication interfaces and queues; the restriction is asserted only for interfaces actually subject to enforcement.¶
Negative tests per Section 16; source reports: 28 and 17 test runs (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E01 commitment.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
communication freeze; distributed quiescence; barrier; message fencing¶
Independent claims 3, 4, 5. Claim 3: Independently Enforced Cross-Domain Communication Freeze and Distributed Quiescence System; Claim 4: Atomic Fencing, Pending-Work Reconciliation, and Verified Freeze Method; Claim 5: Protected Admission Fencing and Late-Operation Disposition Dependent claims 77-100 (indexed, not individually analyzed here).¶
E03 (CE-NSI-01), source page 309. Negotiation-State Integrity, Causal Message Containment, and Continuous Evaluation-Snapshot Validity¶
A later message can silently change the plan an evaluator is reviewing unless message causality and snapshot validity are tracked.¶
Agents; evaluator; state-integrity controller.¶
Causal event records (per-sender sequencing), snapshot identifiers, invalidation records, version-bound plan and evidence references.¶
In: messages and accepted task states. Out: a stable verifiable reference to the exact plan and interaction state, or invalidation.¶
snapshot valid, material change observed, snapshot invalidated.¶
Superseded or unaccounted communications cannot silently alter the evaluated plan; invalid snapshot cannot support a determination.¶
Out-of-order or replayed messages; snapshot reuse after material change.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Protected state controller for plan and evidence references; does not itself admit external effects.¶
Negative tests per Section 16; source reports: no test-run output located by the editor in the pages examined. These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E01, E02.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
causal containment; snapshot validity; negotiation-state integrity¶
Independent claims 6, 7. Claim 6: Protected Multi-Agent Causal-State and Evaluation-Snapshot Integrity System; Claim 7: Causal Message Containment and Continuous Snapshot-Validity Method Dependent claims 101-116 (indexed, not individually analyzed here).¶
E04 (CE-INE-01), source page 375. Independent Multi-Agent Consequence Evaluator, Protected Evidence Admission, and Next-Actor Selection¶
The evaluator must receive exactly the plan and evidence state without gaining the resource credentials needed to effectuate anything.¶
Independent evaluator; evidence sources; agents; organizational policy owners.¶
Admitted evidence set with source roles; evaluator appointment reference; typed recommendation record.¶
In: frozen plan, authenticated evidence. Out: a non-effective decision and a risk-constrained next-actor or action-class recommendation.¶
evidence admitted, evaluated, recommendation issued.¶
Evaluator recommendation is not permission; inadmissible evidence is excluded; missing evidence yields a hold or request, not approval.¶
Evaluator evidence-access escalation; malicious instructions in evidence; evaluator self-appointment.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
The evaluator is separated from the effectuation credentials; permission is handled elsewhere.¶
Negative tests per Section 16; source reports: 50 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E01 to E03.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
independent consequence evaluator; evidence admission; non-effective decision¶
Independent claims 8, 9. Claim 8: Independent Multi-Agent Consequence Evaluation and Non-Effective Decision System; Claim 9: Source-Aware Evaluation, Risk-Constrained Actor Selection, and Protected Decision-Handoff Method Dependent claims 117-132 (indexed, not individually analyzed here).¶
E05 (CE-VCR-01), source page 422. Isolated Virtual Consequence Reconstruction and Non-Effective Scenario Execution¶
A simulation can mislead: it is neither a real commit nor proof that destination infrastructure behaves as predicted.¶
Evaluator; isolated simulation environment; resource-state observers.¶
Virtual state reconstruction, scenario records, uncertainty annotations, phase-bound decision evidence.¶
In: verified plan, authentic observations. Out: predicted consequences with stated uncertainty.¶
reconstructed, scenario run, evidence recorded.¶
Simulation result is never treated as proof of actual execution; model uncertainty and fidelity limits are preserved.¶
False simulation-as-completion assertion; data egress from the simulator.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Isolated non-effective environment with guarded egress; no resource effect.¶
Negative tests per Section 16; source reports: 62 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E04.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
virtual consequence reconstruction; non-effective simulation; digital twin¶
Independent claims 10. Claim 10: Non-Effective Consequence Evaluation of Multi-Agent Candidate Operations Dependent claims 133-140 (indexed, not individually analyzed here).¶
E06 (CE-RBL-01), source page 486. Risk, Benefit-Loss and Tail-Exposure Evaluation With Independent Domain Thresholds¶
One organization's acceptable risk score must not override another owner's mandatory restriction, and no universal scalar risk exists.¶
Evaluator; each domain owner with its own thresholds; agents.¶
Per-domain thresholds, loss and tail-exposure measures, veto records, calibration records.¶
In: consequence evidence and policies. Out: risk classification per domain and conditioned continuation recommendation.¶
measured, thresholds applied, veto or pass.¶
A domain veto is not outvoted; missing calibration limits conclusions; no invented universal policy.¶
Threshold manipulation; miscalibrated risk; cumulative exposure ignored.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Each owner's own enforcement boundary applies its own threshold.¶
Negative tests per Section 16; source reports: 69 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E04, E05.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
tail exposure; risk benefit loss; domain threshold; owner veto¶
Independent claims 11, 12. Claim 11: Owner-Specific Consequence Bound Enforcement for Cooperative Operations; Claim 12: Calibration-Bounded Risk Reassessment With Adverse-Outcome Feedback Dependent claims 141-156 (indexed, not individually analyzed here).¶
E07 (CE-NAS-01), source page 538. Evaluator-Directed Next-Actor Selection and Controlled Next-Turn Instructions¶
Selecting an actor is not granting its credentials; routing needs eligibility, restrictions and a separately enforced continuation.¶
Evaluator; candidate next actors; owners.¶
Actor eligibility records, bounded next-turn instruction, task-type class.¶
In: evaluation and eligibility. Out: selected next actor and bounded instruction.¶
candidates screened, actor selected, instruction bounded.¶
Selection of an agent does not confer resource credentials; conflicting concurrent moves are arbitrated, not both admitted.¶
Next-actor impersonation; conflicting concurrent next moves.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Continuation enforced separately at the actor's resource or communication gate.¶
Negative tests per Section 16; source reports: 62 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E04 to E06.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
next actor selection; bounded instruction; arbitration of concurrent moves¶
Independent claims 13, 14. Claim 13: Protected Eligibility-Based Next-Actor Allocation; Claim 14: Conflict-Aware Arbitration of Concurrent Agent Next Moves Dependent claims 157-172 (indexed, not individually analyzed here).¶
E08 (CE-ACR-01), source page 582. Corrective Instructions, Protected Plan Revision, and Controlled Multiparty Renegotiation¶
A corrective instruction must not rewrite an already protected joint plan without an authorized transition.¶
Evaluator; agents; owners who must re-commit.¶
Patch records scoped to fields, version invalidation, fresh recommitment records.¶
In: evaluation finding. Out: authorized plan revision after fresh multi-owner recommitment, or hold.¶
correction proposed, versions invalidated, renegotiated, recommitted.¶
A model-generated proposal does not overwrite the committed plan; revision requires new consents.¶
Unauthorized renegotiation; stale commitments surviving revision.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Re-entry into a valid evaluation cycle with the same owner-side enforcement.¶
Negative tests per Section 16; source reports: 40 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E01, E04, E07.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
corrective patch; plan revision; renegotiation; recommitment¶
Independent claims 15, 16. Claim 15: Field-Scoped Corrective Patch Admission for Committed Plans; Claim 16: Causal Renegotiation and Fresh Multi-Owner Recommitment Dependent claims 173-188 (indexed, not individually analyzed here).¶
E09 (CE-ARR-01), source page 624. Agent Role Reconfiguration, Bounded Delegation, and Protected Capability Transition¶
Changing an agent's role or tools is a capability change that must be separate from authority for an external effect.¶
Evaluator; owners; agents; capability enforcement controller.¶
Role and capability generations, delegation records, outstanding-work and credential fencing records.¶
In: evaluator recommendation, owner approval. Out: versioned capability transition.¶
role proposed, owner-approved, transition enforced.¶
Old credentials, retained workers and cached role assignments are fenced; role change does not authorize effects.¶
Credential retention; stale role cache; unauthorized delegation.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Protected versioned capability-enforcement point.¶
Negative tests per Section 16; source reports: 59 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E07, E08.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
role reconfiguration; bounded delegation; capability generation¶
Independent claims 17, 18. Claim 17: Owner-Governed Dynamic Agent Role Reconfiguration; Claim 18: Safe Strategy Replacement With Outstanding-Work and Credential Fencing Dependent claims 189-204 (indexed, not individually analyzed here).¶
E10 (CE-SCR-01), source page 670. Selective Communication Resumption, Single-Turn Authority, Automatic Re-Freezing, and Inter-Phase Hold¶
Allowing one needed message should not reopen unrestricted communication.¶
Evaluator; agents; communication gates.¶
Single-turn authority record, communication-edge permission, refreeze state.¶
In: evaluator need for a specific communication. Out: one permitted message then automatic return to hold.¶
held, one turn admitted, refrozen.¶
Turn termination returns to a protected evaluation state; unknown delivery is handled, not assumed.¶
Race between reopening and refreezing; reuse of the turn authority.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Ingress and egress admission at the actual communication boundary.¶
Negative tests per Section 16; source reports: 53 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E02, E07.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
single-turn authority; selective resumption; automatic refreezing; inter-phase hold¶
Independent claims 19, 20. Claim 19: One-Message Cross-Domain Communication Reopening; Claim 20: Race-Safe Re-Freezing and Unknown-Delivery Handling Dependent claims 205-220 (indexed, not individually analyzed here).¶
E11 (CE-NMP-01), source page 719. Protected Next-Move Passport, Actor-Bound Issuance, Single-Use Finality Verification, and Outcome-Uncertainty Reconciliation¶
A next-move authorization must not become a bearer token that any holder can reuse.¶
Issuer, selected actor, owners, consuming gate.¶
Passport (or protected-local slot) bound to actor, task, operation, plan version, phase, freeze epoch, policy, resource state and single use; consumption state.¶
In: evaluator selection and owner approval. Out: a single-use bound authorization and consumption record.¶
issued, presented, consumed, expired or revoked, unknown.¶
Replay, wrong actor, wrong operation or consumed passport are rejected; unknown outcome goes to reconciliation, not retry.¶
Passport replay; next-actor impersonation; attenuation errors.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Verified at the consuming gate before the act proceeds.¶
Negative tests per Section 16; source reports: 62 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E07, E10.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
next-move passport; single-use authorization; non-bearer; consumption state¶
Independent claims 21, 22. Claim 21: Protected Actor- and Operation-Bound Continuation Authority; Claim 22: Protected Next-Move Grant Attenuation, Revocation and Consumption State Dependent claims 221-236 (indexed, not individually analyzed here).¶
E12 (CE-PFS-01), source page 768. Protected Finality Sink Enforcement Before Communication, Privileged Execution, External Dispatch and Physical Actuation¶
A valid evaluation and authorization are not the same as admission of the actual pending act at the resource.¶
Resource owner; Finality Sink; agents.¶
Reconstructed pending act, authorization reference, alternate-path closure records, external-dispatch outbox.¶
In: authorization and pending act. Out: admit exactly the authorized act or withhold.¶
act reconstructed, compared, admitted or withheld.¶
Non-matching act is withheld; model evaluation does not bypass owner admission.¶
Alternate unmonitored paths; act altered after authorization.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
The participating effectuation boundary before communication release, data export, privileged execution, remote dispatch, payment, network configuration or actuation.¶
Negative tests per Section 16; source reports: 53 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Applies the foundational Finality Sink to the evaluator setting.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
finality sink; pending act reconstruction; point of effect; outbox¶
Independent claims 23, 24. Claim 23: Native Pending-Act Reconstruction and Exact-State Authorization; Claim 24: Alternate-Path Closure and Protected External-Dispatch Outbox Dependent claims 237-252 (indexed, not individually analyzed here).¶
E13 (CE-NSA-01), source page 815. Risk-Triggered Neutral Substitute-Agent Selection, Appointment, and Protected Reauthorization¶
When risk triggers replacement of an agent, the substitute must be neutral, qualified and separately authorized, not an inheritor of credentials.¶
Evaluator; original agents; substitute candidates; owners.¶
Eligibility and conflict screening records, de-appointment record, substitute designation.¶
In: risk trigger. Out: appointed substitute with fresh authorization.¶
risk triggered, candidates screened, substitute appointed.¶
No automatic inheritance of the replaced agent's credentials; conflicted candidates rejected.¶
Substitute credential inheritance; hidden conflict of interest.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Substitute acts only after reauthorization at resource gates.¶
Negative tests per Section 16; source reports: 56 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E04, E09.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
substitute agent; neutrality; de-appointment; reauthorization¶
Independent claims 25, 26. Claim 25: Neutral Substitute-Agent Eligibility and Conflict Screening; Claim 26: Original-Agent De-Appointment and Protected Substitute Designation Dependent claims 253-268 (indexed, not individually analyzed here).¶
E14 (CE-PWM-01), source page 863. Protected Pending-Work Escrow, Task-State Migration, and Substitute Reauthorization¶
Transferring task data must not transfer native effect permission or consumed authority.¶
Source agent or controller; destination; owners; evaluator.¶
Canonical minimal-disclosure escrow of unfinished work, source fencing, destination acceptance records.¶
In: unfinished work and chosen substitute. Out: committed escrow, destination acceptance, fresh authorization.¶
escrowed, source fenced, destination accepts, reauthorized.¶
Unproven outcomes and consumed authorizations are not inherited; unaccepted escrow is not live.¶
Escrow substitution; source continuing to act after fence.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Source fencing and destination-side admission.¶
Negative tests per Section 16; source reports: 61 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E13.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
pending-work escrow; task-state migration; source fencing¶
Independent claims 27, 28. Claim 27: Canonical Minimal-Disclosure Escrow of Unfinished Multi-Agent Work; Claim 28: Recoverable Cross-Domain Task Escrow Commit and Destination Acceptance Dependent claims 269-284 (indexed, not individually analyzed here).¶
E15 (CE-PCR-01), source page 928. Phased Execution, Protected Completion Receipts, and Verified Phase Advancement¶
A coordinator's claim that a phase is done is not evidence; phases need resource-origin receipts and prerequisites.¶
Phase scheduler; resource owners issuing receipts; evaluator.¶
Phase identities, source receipts, prerequisite records, advancement state.¶
In: receipts. Out: verified phase closure and separately authorized advancement.¶
phase open, receipt verified, closed, next phase authorized.¶
Missing receipt blocks advancement; phases cannot be skipped; unknown outcome is reconciled.¶
Missing phase receipt; forged completion evidence.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Resource-originated receipt verification before advancement.¶
Negative tests per Section 16; source reports: 66 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E12, E14.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
phased execution; completion receipt; phase advancement¶
Independent claims 29, 30, 31. Claim 29: Resource-Originated Atomic Task Completion Evidence; Claim 30: External Effect Intent Correlation and Unknown-Outcome Reconciliation; Claim 31: Protected Evidence-Based Phase Closure and Separately Governed Advancement Dependent claims 285-308 (indexed, not individually analyzed here).¶
E16 (CE-VFO-01), source page 1005. Verified Feedback, Actual-Outcome Evaluation, and Protected Continuation¶
Agent self-reported success must not stand in for actual outcome evidence; feedback can also carry hostile instructions.¶
Evaluator; resource owners; feedback sources.¶
Source-attributed feedback records, evidence-origin correlation, outcome classification.¶
In: feedback and outcome evidence. Out: outcome classification supporting a proposed continuation.¶
feedback received, origin correlated, outcome classified.¶
Untrusted feedback instructions are contained; unknown is not success; continuation needs fresh authorization.¶
Authenticated but false outcome reports; instruction injection in feedback.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Continuation is separately authorized at owner gates.¶
Negative tests per Section 16; source reports: 61 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E15.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
verified feedback; actual outcome; outcome truth; feedback containment¶
Independent claims 32, 33. Claim 32: Independent Evidence-Origin Correlation and Actual-Outcome Classification; Claim 33: Untrusted Feedback Instruction Containment and Outcome-Truth Continuity Dependent claims 309-324 (indexed, not individually analyzed here).¶
Embodiments E17-E27 are treated individually. E15 relates to phase receipts, E18 to partition recovery, and E27 to native partial outcomes; all three are read in relation to the existing execution-finality drafts and add no new protocol.¶
act sent -> response lost -> status unknown read-only source-authoritative status inquiry committed -> record; not committed -> new authorization unreachable -> UNKNOWN (no automatic retry) old coordinator generation g1 < new g2 -> resource rejects g1 recovery and retry are separately authorized
close request -> inventory obligations in each mandatory domain, queues, delegated work, grants every domain sealed with source-authenticated evidence? NO -> closure withheld YES -> revoke/exhaust grants -> tombstone (no reopen) late completion evidence recorded; terminated authority is never revived. Coordinator DONE is not closure.
owners consent -> joint appointment -> host bound to instance -> installed at each domain -> per-call authorization lease expires / owner withdraws -> revocation serial n+1 -> local fence acknowledged disconnected domain: authority ends by lease bound, not by an assumed instant signal; successor needs new generation
domain A policy domain B policy
\ /
authenticated trust formation
|
verified policy assertions -> schema + semantic check
compatible? NO -> refuse (no optimistic merge)
YES -> constraint intersection (conservative)
-> bounded joint authority; non-delegable stay local
drift (owner or policy generation) -> re-verify
org X op1 --native receipt--> org Y checks predecessor org Y admits op2 only if receipt verifies at Y no common supervisor required; no global authority convergence evidence is not permission forged or missing predecessor -> refuse at owner side
change proposed -> snapshot eligible electorate (epoch e) ballots (source-authenticated, one per organization) mandatory vetoes honored? joiner consents? NO -> blocked YES -> stage epoch e+1 old epoch fenced; unresolved effects settled local installation complete -> activate e+1 resource-local authorization still required
evaluator request (purpose, audience) -> custodian checks purpose + audience match? NO -> refuse YES -> reveal attribute subset + commitment proof evidence lineage + disclosure audit (minimal retention) proof of origin or inclusion != proof of physical truth
domain 1: PREPARE ok domain 2: PREPARE ok domain 3: denied owner-specific authorization per domain -> COMMIT at gates results: committed / denied / UNKNOWN any domain not committed -> PARTIAL_EFFECT manifest recovery choices: inquire, retry (authorized), compensate no global atomic transaction is assumed
partial effect recorded -> compensation proposed owner of affected resource consents? NO -> refused YES -> compensation is a NEW native effect (own receipt) reversible? -> may reverse; irreversible -> corrective act original committed record stays in history
roster of mandatory domains for each: resource-origin receipt present and verified? any missing -> composite status INCOMPLETE all present -> composite completion record record is evidence only (not executable authority)
E17 (CE-EAC-01), source page 1078. Evaluator Integrity, Source Provenance, Collusion-Resistant Review, and Protected Continuation¶
A signature or majority vote does not establish a reliable evaluation; independence and dissent need protected treatment.¶
Evaluator; reviewers; evidence sources; owners.¶
Appointment, reviewer roster, control-group mapping, review-round record, dissent record.¶
In: evidence and reviews. Out: closed round with preserved dissent, or fail-closed hold.¶
round open, reviews admitted, completeness checked, closed.¶
Mandatory dissent is not suppressed; correlated sources are not counted as independent; failure of mandatory conditions fails closed.¶
Reviewer collusion; compromised evaluator; correlated observations.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Continuation depends on a complete, independent review and separate authorization.¶
Negative tests per Section 16; source reports: 120 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E04, E16.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
evaluator integrity; control group; mandatory dissent; review round closure¶
Independent claims 34, 35. Claim 34: Evaluator Eligibility With Underlying Control-Group Independence; Claim 35: Closed Review-Round Completeness and Mandatory Dissent Preservation Dependent claims 325-340 (indexed, not individually analyzed here).¶
E18 (CE-PAR-01), source page 1133. Cross-Domain Network Partitions, Unknown Outcomes, Durable Fencing, Migration and Safe Recovery¶
A lost acknowledgment is not a failed effect, and retries or recovery need separate authorization and fencing.¶
Coordinators (old and new); resource owners; workers.¶
Original-operation identity, native-effect fingerprints, idempotency keys, durable fencing generations, checkpoint lineage.¶
In: partition or lost response. Out: authoritative status or unknown; fenced stale leaders.¶
executing, response lost, status queried, resolved or unknown.¶
Unknown or partial effects are not assumed rolled back or retried; stale workers rejected.¶
Split-brain coordination; unknown effect retried; stale leader.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Read-only source-authoritative status inquiry and resource-side monotonic fencing.¶
Negative tests per Section 16; source reports: 123 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E11, E15.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
unknown outcome; durable fencing; idempotency key; partition recovery¶
Independent claims 36, 37, 38. Claim 36: Resource-Authoritative Reconciliation After Lost Execution Response; Claim 37: Monotonic Resource-Side Stale-Leader Fencing; Claim 38: Checkpoint-Lineage-Controlled Agent Migration and Recovered Execution Dependent claims 341-364 (indexed, not individually analyzed here).¶
E19 (CE-SCL-01), source page 1187. Protected Session Closure, Obligation Reconciliation, Native Domain Sealing, and Irreversible Authority Retirement¶
A coordinator's DONE message does not prove every effect-capable path is closed.¶
Coordinators; each mandatory domain owner.¶
Obligation inventory, per-domain seals, closure tombstone, grant revocation records.¶
In: closure request. Out: sealed closure or withheld closure.¶
closing, domains sealed, tombstone set.¶
Closure withheld while obligations or effect-capable grants remain; terminated session authority is not revived.¶
False complete-session closure; residual grants after closure.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Native-domain sealing at each domain.¶
Negative tests per Section 16; source reports: 84 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E15, E18.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
session closure; domain seal; tombstone; authority retirement¶
Independent claims 39, 40. Claim 39: Complete Native-Obligation Inventory for Agent Session Termination; Claim 40: Independently Sealed Domains and Non-Reopenable Session Tombstone Dependent claims 365-380 (indexed, not individually analyzed here).¶
E20 (CE-JSA-01), source page 1226. Verified Mandatory Approvers, Authenticated Hosting and Installation, and Revocable Supervisor Identity¶
A supervisor cannot appoint itself; mandatory owners must consent and the supervisor must be installed at each domain.¶
Mandatory owners; supervisor; hosting runtime.¶
Appointment, host-instance binding, installation records, per-call authorization, lease.¶
In: owner consents. Out: authenticated supervisor installed with limited coordination functions.¶
proposed, approved, hosted, installed, active, withdrawn.¶
No self-appointed supervisor; service creation is not authorization to operate; coordination is not native effect authority.¶
Unauthorized supervisor creation or installation.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Local installation and per-call authorization at each domain.¶
Negative tests per Section 16; source reports: 91 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E12, E19.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
joint supervisor; mandatory approver; host binding¶
Independent claims 41, 42. Claim 41: Owner-Authorized Bounded Cross-Organization Supervisor; Claim 42: Supervisor Host Instance Binding and Independent Local Installation Dependent claims 381-396 (indexed, not individually analyzed here).¶
E21 (CE-IEO-01), source page 1286. Independent Cross-Domain Evaluator Appointment, Actual-Outcome Verification, Reviewer Accountability and Dispute Containment¶
Evaluators across organizations need owner-derived appointment, outcome sources independent by control group, and dispute containment.¶
Owners; roster of evaluators; reviewers; outcome sources.¶
Evaluator roster, qualification and consent records, review round, dispute and appeal records.¶
In: owner roster. Out: protected review result, including UNKNOWN and HOLD, with separate continuation.¶
appointed, round open, closed, disputed, appealed.¶
Disputes are contained and new rounds isolated; unavailable evaluator never grants permission.¶
Collusion; false independence; dispute spillover.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Continuation separately authorized at owner gates.¶
Negative tests per Section 16; source reports: 82 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E17, E20.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
cross-domain evaluator; reviewer accountability; dispute containment; appeal¶
Independent claims 43, 44. Claim 43: Cross-Organization Authorization of Independent Outcome Evaluation; Claim 44: Authenticated Evaluation Dispute, Appeal and New-Round Isolation Dependent claims 397-412 (indexed, not individually analyzed here).¶
E22 (CE-CTP-01), source page 1356. Cross-Domain Trust and Policy Compatibility¶
Organizations do not share one authority root or one universal policy; compatibility must be verified, not assumed.¶
Independent organizations; resource owners.¶
Trust records, verified policy assertions, schema and semantic compatibility results, non-delegable conditions, policy generations.¶
In: policy assertions. Out: constraint intersection and bounded joint authority, or refusal for incompatibility.¶
trust formed, policies compared, joint authority bounded, drift detected.¶
Incompatible policies are not merged optimistically; non-delegable conditions stay with owners.¶
Incompatible cross-domain policies; stale policy generation.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Each owner keeps local resource admission.¶
Negative tests per Section 16; source reports: 121 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E20, E21.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
trust formation; policy semantic compatibility; constraint intersection¶
Independent claims 45, 46. Claim 45: Semantic Compatibility and Owner-Bounded Cross-Domain Cooperation; Claim 46: Continuous Cross-Domain Owner and Policy Drift Containment Dependent claims 413-428 (indexed, not individually analyzed here).¶
E23 (CE-LDC-01), source page 1425. Leaderless Cross-Domain Agent Coordination¶
Cooperation without a common supervisor needs dependencies enforced at the resource owner, not just a task graph in an agent.¶
Independent resource owners; local supervisors.¶
Causal dependency evidence, predecessor receipts, remote-status leases.¶
In: predecessor receipts. Out: owner-side admission or refusal; convergence evidence only.¶
predecessor required, receipt verified, admitted.¶
Convergence evidence is not executable permission; spoofed dependency is refused.¶
Dependency spoofing; split coordination views.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Resource-owner-side predecessor check.¶
Negative tests per Section 16; source reports: 103 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E15, E22.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
leaderless coordination; causal dependency; predecessor receipt¶
Independent claims 47, 48. Claim 47: Leaderless Causal Task Execution Across Independent Resource Owners; Claim 48: Owner-Side Causal Receipt and Remote-Status Lease Validation Dependent claims 429-444 (indexed, not individually analyzed here).¶
E24 (CE-SER-01), source page 1515. Self-Expiring Supervisor and Evaluator Authority, Authenticated Revocation, and Generation-Fenced Replacement¶
Delegated authority must expire or be withdrawn safely without claiming instant remote revocation.¶
Owners; supervisor or evaluator; successor.¶
Lease, monotonic revocation serial, tombstones, local fencing acknowledgments, generation-fenced successor appointment.¶
In: lease or withdrawal. Out: expired or revoked authority; safe successor.¶
active, expiring, revoked, fenced, successor active.¶
Disconnected domain not yet informed is not claimed revoked; stale host rejected.¶
Revocation delay; stale supervisor during replacement; key compromise.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Local fencing at each domain.¶
Negative tests per Section 16; source reports: 115 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E20, E21.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
authority lease; revocation convergence; generation fencing; successor activation¶
Independent claims 49, 50. Claim 49: Bounded Supervisor Lease and Authenticated Revocation Convergence; Claim 50: Safe Successor Activation After Partitioned Supervisory Withdrawal Dependent claims 445-460 (indexed, not individually analyzed here).¶
E25 (CE-DMG-01), source page 1579. Dynamic Multi-Organization Membership and Principal-Independent Threshold Governance¶
Voting authority belongs to organizations, not agent counts, and membership changes need epochs.¶
Member organizations; joiners; mandatory-veto holders.¶
Eligible-electorate snapshot, membership epoch, ballots, joiner consent.¶
In: membership change and ballots. Out: reconstituted membership with new epoch activation.¶
proposed, voted, staged, old epoch fenced, new epoch active.¶
One organization with many agents does not gain many votes; non-delegable vetoes hold; unresolved effects block activation.¶
Vote inflation; mandatory veto bypass.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Resource-local authorization after reconstitution.¶
Negative tests per Section 16; source reports: 178 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E20, E22, E24.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
membership epoch; principal-independent threshold; owner veto; reconstitution¶
Independent claims 51, 52. Claim 51: Principal-Independent Threshold Governance With Non-Delegable Owner Veto; Claim 52: Protected Membership Reconstitution and Epoch-Safe Join/Leave Dependent claims 461-476 (indexed, not individually analyzed here).¶
E26 (CE-PCE-01), source page 1650. Purpose-Bound Cross-Organizational Outcome Evidence¶
Evaluators should receive no more confidential evidence than needed, and origin proof is not proof of physical truth.¶
Evidence custodians; verifier audience; owners.¶
Purpose-bound requests, source commitments, attribute proofs, disclosure lineage, audit anchors.¶
In: purpose-bound request. Out: selective, audience-bound evidence and audit record.¶
requested, committed, disclosed, audited.¶
Out-of-purpose or wrong-audience release refused; inclusion proof not treated as truth.¶
Selective-disclosure abuse; metadata leakage; forged inclusion proofs.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Evidence release gate at the custodian.¶
Negative tests per Section 16; source reports: 98 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E16, E21.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
minimal disclosure; purpose-bound evidence; selective disclosure; evidence lineage¶
Independent claims 53, 54, 55. Claim 53: Source-Authenticated Selective Evidence Commitments; Claim 54: Audience- and Purpose-Bound Confidential Outcome Evidence Release; Claim 55: Minimal-Retention Evidence Audit With Source Sequence and Anchor Continuity Dependent claims 477-500 (indexed, not individually analyzed here).¶
E27 (CE-MEF-01), source page 1715. Multi-Domain Protected Execution Commitment, Partial-Outcome Reconciliation and Cross-Organizational Recovery¶
There is no global atomic transaction across independently administered domains; partial results and compensation need honest handling.¶
Resource owners in several domains; coordinator; evaluator.¶
PREPARE and COMMIT records, predecessor manifests, partial-outcome manifest, compensation authorization, completeness roster.¶
In: composite plan. Out: per-domain native outcomes and a non-executable completion record.¶
prepared, committed, partial, unknown, compensated, complete.¶
PARTIAL_EFFECT and UNKNOWN are shown honestly; compensation is a new authorized act; completion record is not executable.¶
Partial commitment; compensation without consent; irreversible effect treated as rolled back.¶
Alternatives are other placements of the same function and existing mechanisms described in Section 17; selection is policy-specific.¶
Owner-specific protected prepare and commit at each native domain gate.¶
Negative tests per Section 16; source reports: 106 tests (reported). These are local tests of the source's modeled code paths and do not establish production security.¶
Builds on E12, E15, E18, E19.¶
The evaluator output stays non-executable; effects are admitted only at a participating native boundary as described in [REL-FOUND].¶
prepare commit; partial effect; compensation; completeness evidence¶
Independent claims 56, 57, 58, 59, 60. Claim 56: Independent Prepare-and-Commit Authorization at Native Domain Gates; Claim 57: Owner-Installed Mandatory Predecessor Manifest Enforcement; Claim 58: Partial Composite-Effect Reconciliation and Controlled Resume; Claim 59: Separately Authorized Compensation as a New Native Effect; Claim 60: Composite Native-Evidence Completeness Without Composite Execution Authority Dependent claims 501-540 (indexed, not individually analyzed here).¶
Illustrative, protocol-independent records follow. Field names are candidates only; JSON or CBOR encodings are examples. Signed portable artifacts are distinguished from protected local state, which is not exported.¶
Illustrative example (portable):¶
{
"id": "req-1",
"task": "t-9",
"plan_ref": "sha256:...",
"owners": [
"o1",
"o2"
],
"purpose": "pre-effect review"
}¶
Illustrative example (portable, signed by owners):¶
{
"evaluator": "ev-3",
"task": "t-9",
"owners_consent": [
"o1",
"o2"
],
"policy_gen": 7,
"expiry": "2026-12-01T00:00:00Z",
"uses": 1,
"revocation_serial": 0
}¶
Illustrative example (protected local plus signed summary):¶
{
"round": "r-1",
"evidence_set": "sha256:...",
"state": "closed",
"reviewers": 3,
"control_groups": 3,
"dissent": []
}¶
Illustrative example (portable, source-signed):¶
{
"id": "ev-12",
"source": "o1",
"role": "resource_origin",
"task": "t-9",
"fresh_until": "...",
"policy_gen": 7,
"class": "native_receipt"
}¶
Illustrative example (portable, signed):¶
{
"type": "HOLD",
"round": "r-1",
"task": "t-9",
"valid": [
"..",
"..."
],
"executable": false
}¶
Illustrative example (portable, signed):¶
{
"supervisor": "sv-1",
"owners_consent": [
"o1",
"o2"
],
"host_binding": "inst-5",
"lease": "...",
"serial": 3
}¶
Illustrative example (resource-origin, signed or protected local):¶
{
"op": "op-77",
"domain": "d1",
"result": "COMMITTED",
"fingerprint": "...",
"gen": 12
}¶
Illustrative example (portable, signed):¶
{
"composite": "c-2",
"domains": {
"d1": "COMMITTED",
"d2": "UNKNOWN",
"d3": "DENIED"
},
"status": "PARTIAL_EFFECT"
}¶
Illustrative example (portable, per-domain signed):¶
{
"session": "s-4",
"domains_sealed": [
"d1",
"d2"
],
"tombstone": true
}¶
Illustrative example (portable, source-authenticated):¶
{
"purpose": "outcome audit",
"audience": "rev-1",
"attributes": [
"amount_range"
],
"commitment": "..."
}¶
Pseudocode is illustrative and non-normative; it is not a reference implementation and the source's attached tests are local tests only.¶
def admit_evaluator(e, task, now):
a = appointments.get(e, task)
if a is None or a.revoked or now >= a.expiry: return HOLD
if not owners_consented(a, task.mandatory_owners): return HOLD
if not independent(e, task.parties): return DENY
return OK¶
def admit_evidence(x, r):
checks = [origin(x), role(x), integrity(x),
binds(x, r.task), fresh(x), policy_gen(x),
independent(x), uncorrelated(x)]
if not all(checks): return EXCLUDE
return ADMIT # admitted != outcome truth¶
def verify_det(d, now):
return (sig_ok(d) and d.executable is False and
now in d.validity and snapshot_valid(d.snapshot)
and not withdrawn(d))¶
def continue_(d, owner_policy):
if not verify_det(d, now()): return HOLD
grant = owner_policy.authorize(d) # separate decision
return grant # gate still admits or withholds¶
def revoke(a):
a.serial += 1; publish(a)
for dom in domains(a): fence(dom, a.generation)
# disconnected domains end by lease expiry, not by signal¶
def close_round(r):
if not mandatory_present(r): return HOLD
if control_groups(r) < r.policy.min_groups: return HOLD
if not dissent_included(r): return REJECT
r.state = 'closed'; return seal(r)¶
def reconcile(op):
s = source_status(op) # read-only, source-authoritative
if s is None: return UNKNOWN # never auto-retry
return COMMITTED if s.committed else NEEDS_NEW_AUTH¶
Seven central relationships are NOT implied:¶
The following threats each carry eight fields. They are generic categories; mitigation depends on deployment and residual limitations are stated.¶
act as evaluator¶
controls an agent¶
agent to appointment authority¶
appointment comes from the required authority source¶
check appointment issuer and scope¶
reject; HOLD¶
issuer compromise¶
agent declares itself neutral; determination refused¶
fake an owner's approval¶
forges or replays approval¶
owner to appointment¶
consent is current and from the actual owner¶
verify consent source and freshness¶
reject appointment¶
stolen credentials¶
forged consent record; appointment refused¶
omit an owner¶
controls owner list¶
owner resolution¶
all mandatory owners resolved¶
compare against owner manifest¶
HOLD¶
manifest error¶
one owner removed; no approval¶
create a supervisor¶
controls one domain¶
owner to supervisor¶
supervisor needs mandatory owner consent¶
check joint appointment¶
reject¶
weak owner process¶
single-owner creation; refused¶
install a rogue supervisor¶
access to a host¶
host to domain¶
installation bound to approved instance¶
verify host-instance binding¶
reject installation¶
host compromise¶
unapproved instance install; refused¶
hold evaluator and gate roles unrestrained¶
controls a shared host¶
role boundary¶
co-location grants no inherited authority¶
check separation of role authority¶
deny inherited authority¶
weak isolation¶
same host evaluator tries gate use; denied¶
pass off a related evaluator as neutral¶
claims neutrality¶
independence check¶
independence from actual control relationships¶
check control groups¶
reject evaluator¶
hidden control¶
self-declared neutral; rejected¶
inflate reviewers¶
holds many credentials¶
reviewer diversity¶
count control groups not accounts¶
map credentials to groups¶
count once¶
unknown affiliation¶
ten keys one group; counted once¶
force a result¶
colluding reviewers¶
review round¶
dissent and independence preserved¶
check source independence¶
HOLD or DISPUTED¶
covert collusion¶
colluding pair; round not approved¶
hide a dissent¶
controls aggregator¶
review aggregation¶
mandatory dissent cannot be dropped¶
verify dissent included¶
reject round¶
unrecorded dissent¶
dissent removed; closure refused¶
use expired mandate¶
retains old credential¶
appointment lifetime¶
mandate currently valid¶
check validity at use¶
reject¶
clock error¶
expired mandate presented; refused¶
admit evidence after closure¶
retains round token¶
round lifetime¶
closed round admits nothing¶
check round state¶
reject¶
delayed closure¶
late evidence after close; refused¶
reuse approval elsewhere¶
holds a determination¶
determination scope¶
bound to task and round¶
check scope binding¶
reject¶
loose binding¶
project A result used for B; refused¶
poison many tasks¶
controls platform¶
platform to tasks¶
no automatic inheritance across tasks¶
per-task mandates; platform compromise handling¶
revoke all contexts¶
platform key compromise¶
platform compromised; all tasks revalidated¶
reuse retired instance¶
holds old instance key¶
instance lifetime¶
retired instance has no authority¶
check instance state¶
reject¶
retention of keys¶
retired instance acts; refused¶
read more evidence¶
evaluator process¶
evidence access¶
access limited to mandate¶
check evidence classes¶
deny¶
side channels¶
evaluator requests unlisted class; denied¶
steer the evaluator¶
controls a source¶
evidence to evaluator¶
evidence is data not instruction¶
treat content as untrusted¶
ignore instruction; flag¶
model susceptibility¶
embedded command; no behavior change¶
claim false success¶
holds valid keys¶
source to evaluator¶
origin is not outcome truth¶
require resource-origin evidence¶
UNKNOWN¶
dishonest source¶
signed false success; not accepted as truth¶
fake diversity¶
two reports one origin¶
evidence independence¶
underlying provenance independent¶
check provenance groups¶
count once¶
hidden provenance¶
two APIs one source; counted once¶
use old policy¶
holds old policy¶
policy to evaluation¶
current generation applied¶
compare generations¶
reject¶
propagation delay¶
old generation determination; refused¶
alter plan after freeze¶
queued messages¶
communication edge¶
unaccounted messages not admitted¶
pending-work inventory¶
hold; reconcile¶
unobserved path¶
late message; snapshot invalidated¶
act as selected actor¶
stolen identity¶
actor binding¶
passport bound to actor¶
check actor identity¶
reject¶
identity compromise¶
wrong actor presents passport; refused¶
inherit old credentials¶
new agent¶
substitute to resource¶
no inheritance¶
fresh authorization required¶
deny¶
weak revocation¶
substitute uses old capability; denied¶
swap escrowed work¶
controls escrow store¶
escrow commit¶
escrow canonical and committed¶
verify escrow digest or record¶
reject¶
store compromise¶
altered escrow; refused¶
claim simulation as done¶
controls reporter¶
simulation to record¶
simulation not completion¶
require native receipt¶
reject completion¶
fidelity gap¶
simulated success; no completion¶
two coordinators¶
partition¶
coordinator to resource¶
one valid generation¶
monotonic fencing¶
fence older¶
partition detection limits¶
two leaders; stale one rejected¶
act before revocation arrives¶
disconnected domain¶
revocation to domain¶
no claim of instant revocation¶
bounded freshness¶
fail closed after bound¶
irreducible delay¶
disconnected domain; authority expires by lease¶
merge unlike policies¶
policy mismatch¶
policy compatibility¶
semantic compatibility verified¶
compare semantics¶
refuse¶
semantic ambiguity¶
same word different meaning; refused¶
reverse an effect unilaterally¶
coordinator¶
compensation¶
compensation is a new authorized act¶
check owner authorization¶
refuse¶
urgent pressure¶
unauthorized compensation; refused¶
claim rollback¶
coordinator¶
history¶
irreversible effects stay in record¶
reversibility classes¶
mark irreversible¶
classification error¶
irreversible shown as rolled back; rejected¶
lie at source¶
owner source¶
source honesty¶
limits of source honesty stated¶
cross-source checks¶
UNKNOWN or DISPUTED¶
source dishonesty¶
contradicting sources; DISPUTED¶
block progress¶
designs impossible rules¶
policy¶
rules satisfiable or escalate¶
deadlock detection¶
escalate to human¶
policy design¶
unsatisfiable quorum; escalates¶
extend authority¶
controls clock¶
time¶
expiry uses protected time or events¶
protected time source¶
fail closed¶
time dependency¶
clock rewind; authority not extended¶
reuse consumed state¶
restores checkpoint¶
state restoration¶
consumption survives restore¶
checkpoint lineage¶
reject¶
backup practices¶
restored checkpoint reuses grant; refused¶
unlawful transfer¶
evaluator in other region¶
evidence export¶
jurisdiction rules checked¶
check constraints¶
deny¶
legal ambiguity¶
evidence to forbidden region; denied¶
treat approval as authority¶
integrator¶
determination to gate¶
determination non-executable¶
gate ignores determinations¶
reject¶
integration error¶
gate receives determination alone; refuses¶
bypass the gate¶
finds other path¶
coverage¶
claims limited to covered paths¶
coverage inventory¶
state limitation¶
unknown paths¶
effect via uncovered path; limitation stated¶
No new wire protocol is proposed where existing standards suffice. Candidate carriers, with architecture requirements kept separate from candidate schemas and normative requirements:¶
| Mechanism | Possible use | Note |
|---|---|---|
| HTTP and TLS | transport of requests and determinations | does not supply appointment semantics |
| OAuth 2.0 and token exchange | delegation of bounded evaluator access | token is not an effect authorization |
| JWT and CWT with COSE/JOSE | signed appointment and determination artifacts | signature is not outcome truth |
| Workload identity | evaluator and supervisor identity | identity is not independence |
| Remote attestation (RATS) | host and instance binding | attests state, not truthfulness of evidence |
| Distributed tracing | evidence lineage correlation | traces are claims, not receipts |
| Transparency logs | audit continuity | inclusion is not truth |
| Message queues and agent protocols | carrying requests | delivery is not admission |
Hypothetical and non-endorsed. These descriptions identify potential research contexts only and were not verified against vendor documentation by the editor; readers must verify against public product documentation. Established product functionality is separated from proposed integration, and the alignment is AI-generated and must be cross-checked against the original filings. It is not an endorsement, validation, adoption claim or infringement statement.¶
Established: tool-call approval and human review steps exist in various platforms. Proposed: a gate placed at the tool-execution broker would reject an act whose appointment, round or determination is stale, and would require a separate owner-issued authorization before the call is released. Why it matters: model-approved text then cannot by itself cause an effect.¶
Established: policy decision and enforcement points, workload identity, audit logs. Proposed: an evaluator service whose appointment is derived from resource-owner consent and whose determinations are inputs to, not substitutes for, an existing PEP check at the resource API.¶
Established: dual control, maker-checker, reconciliation. Proposed: a reviewer appointed per transaction class, with minimal-disclosure outcome evidence, UNKNOWN on lost acknowledgments and compensation as a separately authorized act (Examples F, I, J, K).¶
Established: orchestration and slice policies. Proposed: cross-domain evaluator for multi-operator changes, with owner-side predecessor checks and partial-outcome manifests when one domain does not commit.¶
Established: safety interlocks and change management. Proposed: a persistent evaluator serving many maintenance projects with per-project mandates (Example B), with actuation only through the owner's existing controlled interface.¶
Each example gives initial authority, protected state, eligible evidence, determination, separate native authority and failure outcomes.¶
Initial authority: each company owns its purchasing and payment authority. Protected state: a committed, non-effective joint plan, frozen. Eligible evidence: inventory and price receipts from each owner. Determination: APPROVED_FOR_CONSIDERATION by an evaluator appointed by both. Native authority: each company's gate separately admits purchase and payment. Failures: stale price evidence yields HOLD; missing consent yields refusal; unknown payment outcome yields UNKNOWN.¶
Initial authority: one service, many projects, separate mandates. Protected state: per-project namespaces, rounds and validity. Evidence: sensor logs and work-order receipts. Determination: per-project, valid for a bounded interval. Native authority: plant controls admit changes. Failures: project A expiry gives no authority to B; platform compromise suspends all contexts.¶
Initial authority: three owners set reviewer qualifications. Protected state: round with control-group map. Evidence: provenance-diverse reports. Determination: closed only with mandatory reviewer present and dissent preserved. Native authority: owners authorize separately. Failures: three accounts in one group count once; suppressed dissent rejects the round.¶
Initial authority: plan committed. Protected state: freeze epoch. Evidence: pending-work inventory. Determination: HOLD while a late message is unaccounted; snapshot invalidated. Native authority: unchanged, none granted. Failures: partial acknowledgment is not a freeze.¶
Initial authority: old agent's consumed capability. Protected state: escrowed task data, source fenced. Evidence: neutrality screening. Determination: SUBSTITUTE recommended. Native authority: the substitute has no valid permission to continue consuming the old capability. Failures: credential inheritance denied.¶
Initial authority: one authorized payment. Protected state: operation fingerprint and idempotency key. Evidence: none received. Determination: UNKNOWN. Native authority: retry needs new authorization. Failures: the coordinator must not repeat; a read-only status inquiry runs first.¶
Initial authority: epoch e with two owners. Protected state: electorate snapshot. Evidence: joiner consent; unresolved earlier operation. Determination: HOLD until the earlier operation settles. Native authority: local installation required. Failures: disconnected owner's veto not bypassed; activation withheld.¶
Initial authority: two independent owners. Protected state: predecessor manifest at each owner. Evidence: native receipt of operation one at owner two. Determination: convergence evidence only. Native authority: owner two admits operation two only if the receipt verifies. Failures: spoofed predecessor refused.¶
Initial authority: bank controls records. Protected state: attribute commitments. Evidence: selective attributes for a stated purpose and audience. Determination: verified outcome attribute. Native authority: none gained. Failures: out-of-purpose request refused; metadata minimized.¶
Initial authority: two domain owners. Protected state: PREPARE records. Evidence: ledger committed, inventory not released. Determination: PARTIAL_EFFECT. Native authority: each domain separately. Failures: no global atomicity assumed; completion record not executable.¶
Initial authority: a committed operation. Protected state: history record. Evidence: owner consent. Determination: compensation proposed. Native authority: compensation is a new native effect with its own receipt. Failures: irreversible effects cannot be recorded as rolled back.¶
Initial authority: task apparently ended. Protected state: obligation inventory. Evidence: queued work and old grants found. Determination: closure withheld. Native authority: grants revoked or exhausted, then seals. Failures: DONE message alone is rejected.¶
The document makes no universal latency claim. Three paths are distinguished: a fast path (valid cached determination within validity), an evaluation path (new review), and a recovery path (partition, UNKNOWN, revocation). Legacy systems may adopt the architecture by placing the gate at an existing enforcement point, with the evaluator offline or batched.¶
Measurement plan: p50, p95, p99 and p99.9 latency; throughput; setup cost per appointment; reviewer count effect; revocation propagation time; partition recovery time; false acceptance and false denial rates; UNKNOWN frequency; and privacy leakage. No figures are asserted here. Source-reported test counts (see Section 11) come from local Python reference code only and are not performance or production evidence.¶
Each topic below is established prior art or practice in its own right. The applicant does not claim that any isolated primitive, signature, quorum, AI evaluator or distributed-transaction mechanism was invented by the applicant. The architecture's distinction, if any, lies in the particular cooperation of protected appointment, role independence, lifecycle separation, evidence eligibility, bounded determination, owner-specific authorization and native-effect admission. That statement is not a legal conclusion.¶
Related individual drafts by the same author: [REL-FOUND], [REL-CBEA], [REL-CAUSAL], [REL-IEV] (identifiers as supplied by the author; verify on the Datatracker). Others' work on agent delegation, handoff and requirements ([I-D.agentic-ai-usecases-requirements]) is acknowledged; the editor has not claimed a full survey.¶
This is not a legal conclusion on patentability.¶
The architecture reduces the chance that an evaluation result is mistaken for authority, but it depends on correct enforcement at participating boundaries. Non-bypassable native-effect control applies only to actual participating protected enforcement paths. See Section 16 for 60 threat categories. Residual limits: a compromised trusted base, a dishonest resource owner, undetected correlated sources, algorithm compromise, and uncovered alternate effect paths. No implementation is assumed to guarantee globally atomic transactions, instantaneous revocation during disconnection, physically infallible source evidence, zero-latency evaluation or universal interoperability.¶
Evidence is disclosed only under authorized purpose and audience constraints and with minimal retention. Metadata, logs and traces can leak information and are minimized. Cross-jurisdiction evidence transfer follows applicable constraints; the document gives no legal advice.¶
This document requests no IANA actions.¶
The underlying technical invention and patent disclosure are attributed to human inventor Sangam Kumar Das. AI-assisted preparation of the draft, terminology mapping, diagrams, industry comparisons or explanatory text does not independently establish AI inventorship.¶
AI-assisted material can contain mistakes, omissions, inaccurate product references or misinterpretations of technical claims. All assertions must be checked against the original filed patent documents, verified RFCs, primary standards documentation and official industry publications.¶
The document is an individual technical proposal and does not constitute an endorsed IETF standard, approved RFC, adopted working-group specification or IETF consensus position. No IETF-assigned identifier is claimed.¶
References to OpenAI, Anthropic, Google, Microsoft, NVIDIA, AWS, Alibaba, Huawei, SAP, IBM, Nokia, Ericsson, financial companies or other organizations identify potential research contexts only. No adoption, endorsement, affiliation, copying, infringement or commercial implementation is asserted.¶
A positive evaluator determination does not independently establish a real-world effect or confer authority for such an effect. Non-bypassable native-effect control applies only to actual participating protected enforcement paths.¶
The author must comply with applicable IETF BCP 78 and BCP 79 requirements. Mentioning the patent inside a draft does not replace any required formal IETF IPR disclosure. No FRAND, royalty-free or other licensing commitment is made or implied here.¶
Current source disclosure: inventor Sangam Kumar Das; Indian Provisional Patent Application 202631131650; applicant-reported filing date 3 November 2026 (the source PDF is named with a 'NOV 4' label but contains a digital-signature date of 3 November; no 4 November filing is inferred). Supplied source: 1,905 pages, 27 embodiments, 540 numbered claims.¶
Lineage: PCT/IB2026/055615 filed 4 June 2026, published 16 July 2026 as WO 2026/150382 A2. Related Indian provisionals reported by the applicant: 202631134241 (8 November 2026), 202631136555 (15 November 2026). Later features are not assigned the June priority date automatically; entitlement to any priority date is a legal question outside this document.¶
Public research records: https://doi.org/10.5281/zenodo.24274259 and https://doi.org/10.5281/zenodo.24076231; Technical Disclosure Commons records 12542, 12934 and 12973. These were not all fetched by the editor; verify. An IETF IPR disclosure (reference 7610) exists for earlier drafts of the author; whether it covers this draft is not established. IPR will be disclosed at the appropriate time as per IETF guidelines.¶
The complete profiles for E01-E27 are in Section 12 and Section 13; the crosswalk of all 60 independent claims (mapped to the embodiment of each) and the indexed dependent claims 61-540 is in Section 11. Independent claim titles are listed in each profile's Claim references field.¶
27 diagrams appear in the body; their captions are listed here for reference.¶
Appointment state machine: proposed -> owner-approved -> active -> (expired | revoked); evaluation round: open -> admitting -> (closed | disputed); determination: valid -> (stale | withdrawn). Transitions to active require all mandatory owner consents; any lifecycle can end without ending the others (see Section 6).¶
Illustrative records are in Section 14. Example evidence chain: a resource-origin receipt (EvidenceRecord, class native_receipt) is admitted into a closed EvaluationRound, which yields a non-executable EvaluatorDetermination; owner policy separately issues a bounded authorization; the native gate admits or withholds the act and emits a NativeOutcomeReceipt.¶
The 60 negative-test categories, each with a concrete test, are in Section 16 (T01-T60). Test results reported in the source are local reference-code tests and are not reproduced as evidence of security.¶
Counts: 212 keywords and 115 phrases. The index serves searches on WIPO PATENTSCOPE, EPO Espacenet, Google Patents, IEEE, ACM, arXiv, the IETF Datatracker and general technical literature. Presence of a term does not imply novelty or infringement. Cross-references give the embodiment (E) and, through Section 11, the independent claims.¶
Boolean phrases (quoted terms in practice); each lists the embodiments.¶
See Section 18. The mapping is hypothetical, non-endorsed, AI-generated and unverified against vendor documentation; cross-check against the original filings before drawing any conclusion. It is not an endorsement, validation or infringement statement.¶
Where a tool broker, API gateway, payment rail adapter, network controller or actuator interface already enforces access, a gate that additionally requires a current owner-issued authorization could sit there.¶
appointment, round, determination validity, policy generation, actor, operation and single-use state.¶
stale, replayed, wrong-actor or determination-only requests.¶
a model or reviewer approval cannot itself cause an effect.¶
See Section 26, [IN-131650], [PCT-055615], [TDC-12542]. Related drafts: [REL-FOUND], [REL-CBEA], [REL-CAUSAL], [REL-IEV]. This document is distinguishable from them: causal diagnosis locates drift; execution finality governs admission of an effect; this document governs evaluator appointment, evidence eligibility, lifecycles, typed determinations and cross-domain evaluation governance, and uses E12, E15, E18 and E27 only in relation to those existing drafts.¶