| Internet-Draft | Consequence-Bound Execution Authority | November 2026 |
| Das | Expires 20 May 2027 | [Page] |
Authorization to compute, propose, prepare, simulate or request an operation is not, by itself, authorization to cause the operation's actual consequential effect. An authorized operation may afterwards be transformed, delegated, delayed, replicated, compiled, offloaded or executed under different technical conditions. This document describes an architecture, and a set of candidate interoperability requirements, for separating those authorizations: a proposed operation is held as a Candidate Act in a Non-Effective State; the Actual Pending Operation is reconstructed; current source authority, context, dependencies and policy and revocation epochs are evaluated by an independent validator; a scoped, non-bearer execution authorization is issued; and a protected Finality Sink at the actual effectuation boundary re-derives the operation, enforces the authorization, and supports reconciliation against authoritative outcome evidence.¶
The document is an individual, informational technical proposal. It credits OAuth, JWT/CWT, RBAC/ABAC, zero-trust, remote attestation, transaction logs, hardware isolation and cryptographic signatures for their established functions and does not characterize them as defective. It includes a threat model, 12 figures, 39 consequence-specific profiles, a non-normative European applicability discussion and an examiner-oriented terminology appendix. It is not an IETF standard, an EU regulatory instrument, a telecommunications or financial standard, or an endorsement by any named company.¶
The technical subject matter is attributed to the human inventor Sangam Kumar Das. Foundational reference: PCT/IB2026/055615, filed 4 June 2026, published 16 July 2026 as WO 2026/150382 A2. Later Indian provisional application 202631136555, dated 15 November 2026 (as reported by the applicant). See Section 20 for the full IPR and attribution statement.¶
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/.¶
AI-assisted drafting was used to prepare this explanatory text, keyword index and industry mapping. Readers must verify patent numbers, dates, claim quotations, product capabilities and standards references against primary sources.¶
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.¶
Systems increasingly separate the component that decides what to do from the component that makes it happen. An AI agent proposes a tool call; a workflow engine queues it; a database trigger fans it out; a SmartNIC compiles it into a hardware rule; a retry worker repeats it after a timeout. Each hop can be individually authenticated and individually permitted, yet the operation that finally produces an effect may differ from the operation that was approved, may be executed under authority that has since been revoked, or may have an effect that is only visible after composition with other work.¶
This document uses the term *consequence* for the externally effective result of an operation at the point where it can no longer be undone by the originating software alone. The architecture's organizing question is: at the boundary where a consequence would become real, what independent evidence shows that this exact effect, now, under current authority, is permitted?¶
The central principle preserved throughout this document is:¶
Authorization to compute, propose, prepare, simulate, or request an operation is not automatically authorization to cause its actual consequential effect.¶
The following are distinct questions with distinct evidence. A deployment can answer some and not others, and conflating them is the failure mode this document addresses.¶
| Question | Typical established mechanism | What it does not establish |
|---|---|---|
| Who is the caller? (identity authentication) | TLS client auth, OAuth client auth, workload identity | That this operation is permitted |
| May the caller invoke this tool? (tool-invocation permission) | OAuth scopes, RBAC, ABAC, API gateways | That the operation actually reaching the target is the one evaluated |
| May the caller stage this operation? (staging permission) | Draft/queue/propose APIs, change requests | That staged work may later become effective |
| May this exact effect be committed now? (effect-commit permission) | Proposed here: act-bound capability checked at a Finality Sink | That the intended outcome was achieved |
| Did the intended effect occur? (outcome evidence) | Destination-of-record confirmation, settlement evidence | That a later continuation is authorized |
The same author has published several individual Internet-Drafts on execution finality. This document was compared with the following, using the text available to the editor at the time of drafting:¶
The intended additional contribution of this document is limited to the following, and each should be checked for duplication:¶
If, on comparison, the core chain material of this document substantially duplicates an existing draft, the recommended course is to revise that draft or to publish only the profile catalogue and threat analysis as a narrowly scoped companion document, rather than submitting a redundant document.¶
This is an individual technical proposal. It is not an endorsed IETF standard, not an EU regulatory approval, not a telecommunications standard, not a financial certification and not an implementation by any named company. References to companies and products identify potential technical integration points only. They are not evidence of adoption, sponsorship, copying, infringement or commercial affiliation. Regulatory references (Section 18) are engineering relevance statements and are not claims of legal compliance.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY" and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
Scope of normative language: because this document is Informational, normative keywords bind only an implementation that elects to claim the property to which the statement relates. Requirements are collected in Section 6; keywords used elsewhere restate those requirements, state architectural recommendations, or restate IETF process obligations. Such language does not oblige any deployment to claim a property and does not define a wire format, registry or transport. Statements marked [ARCH] are architectural recommendations.¶
Definitions are listed with industry-standard synonyms where they exist. This document avoids new acronyms where a plain term suffices; "IEV" is retained only because it is used by the cited companion draft.¶
A proposed consequential operation produced by a human, software or AI agent. Proposal confers no effectuation authority. Synonyms: proposed operation, deferred operation, staged transaction.¶
A protected state in which a Candidate Act exists without having produced its governed consequence. Synonym: pending non-committed state.¶
The operation that would really occur at the effectuation boundary, derived (reconstructed) from the Candidate Act, current resource state and applicable execution parameters, rather than taken from the proposer's description. Synonym: reconstructed effective operation.¶
The process of deriving the Actual Pending Operation, including the effective principal, target, parameters, material dependencies and derived child effects.¶
The record of the originating authority for an act: principal, delegation chain, permitted task and operation class, limits, purpose and applicable epochs. Synonym: originating authorization context.¶
An environment whose integrity and policy state are protected against the operation source, in which validation occurs and protected state is kept. Realizations include a trusted execution environment, a hardware security module, a separate process or host, or a trusted application boundary.¶
A component, outside the control of the operation source, that evaluates the Actual Pending Operation against current authority, policy, dependencies and evidence. Synonym: runtime policy revalidation service.¶
A non-bearer authorization bound to a specific act, scope, destination, validity period or consumption count, designated Finality Sink and current epochs. Possession alone is not sufficient to cause the effect. Synonym: operation-bound execution authorization.¶
The decision logic at or before the effectuation boundary that admits or rejects an operation.¶
The component at the first boundary at which the operation can become externally effective, which verifies a Scoped Execution Capability against the Actual Pending Operation and consumes it. Synonyms: protected effectuation boundary, execution enforcement point.¶
Authorization whose predicate is evaluated over the effect that would actually result, including aggregate and transitive effects, rather than over the request alone. Synonym: effect-aware authorization.¶
The property that originating restrictions and delegated limits remain binding through transformation, summarization, delegation, queueing, translation and retry. Synonym: delegated authority preservation.¶
A protected, monotonically advancing version of the effect-relevant policy. Synonym: policy generation counter.¶
A protected, monotonically advancing version of revocation state, used to reject authority issued before a revocation.¶
A record, bound to an act and phase, carrying authenticated evidence of what occurred. Self-reported receipts from the act source are not Authoritative Outcome Evidence.¶
A receipt whose evidence originates at the destination of record or an independent observer rather than at the operation source. Synonym: authenticated destination effect evidence.¶
A validation in which authenticated evidence of an earlier real (bounded) effect is evaluated before a continuation condition for a later effect is established. Synonym: receipt-gated continuation validation.¶
A validation receipt whose existence and ordering are anchored in an append-only or transaction-log structure, so that omission or reordering is detectable. Anchoring is optional and does not by itself establish validity.¶
A structured description of the effective operation as reconstructed, including, where relevant, installed match/action behavior of a hardware path (Effective Hardware Action Descriptor).¶
The graph of resources, services, derived operations and state on which the effect depends or which it affects materially. Synonym: causal effect graph.¶
A cryptographic or hardware-protected commitment to state used in validation (for example an epoch, ledger or watermark).¶
Evidence from the system of record, or from an independent observer with adequate coverage, sufficient to classify an operation as committed, conclusively not committed, or unresolved.¶
Authority that was valid at issuance but whose epoch, revocation state, target, context or time bound no longer holds at effectuation.¶
An effect reachable from the authorized operation through a chain of triggers, workflows, queues, delegations or calls, not named by the entry point.¶
The combined effect of several operations or work products that are each individually acceptable.¶
A real but intentionally limited effect, not a simulation, preview or dry run, performed before broader effectuation. Synonyms: canary effect, limited trial effect.¶
Enforcement at the point where the operation actually becomes effective, as opposed to enforcement only at request, gateway or logging layers.¶
The extent to which every discovered route capable of producing a governed effect is subject to the required control, demonstrated by evidence of rejection at the sink rather than by policy configuration or logs.¶
The following mechanisms are not criticized here. Each answers a real question, and the architecture below relies on them.¶
Delegated authorization of a client to call resource servers ([RFC6749], [RFC8693]); proof-of-possession and sender-constraining ([RFC9449], [RFC8705], [RFC7800]); rich authorization details ([RFC9396]). These can carry fine-grained scopes and bind tokens to keys. They do not, by themselves, make the resource server re-derive the effective operation or guarantee that every route to the effect is covered.¶
Compact signed or MACed claim formats ([RFC7519], [RFC8392], [RFC9052]). They are suitable carriers for a Scoped Execution Capability; this document does not select a format.¶
Role-based and attribute-based access control ([NIST-ABAC]). They decide whether a principal may perform an operation class. The profiles here concern whether the particular effect is within scope at effectuation.¶
Continuous verification and least privilege ([NIST-ZTA]). The architecture is consistent with it and adds a requirement to verify at the actual effect boundary.¶
Evidence about an attester's state, appraised by a verifier for a relying party ([RFC9334], [RFC9711]). The attestation profile (E9) concerns correspondence between remediation, evidence and policy per consequence class and does not replace RATS roles.¶
Durable records of what was done. Logging is detection and accountability; it is not prevention unless the log write gates the effect.¶
Isolation of keys and execution. They bound what compromise can do under intact-hardware assumptions; E35 considers the case where those assumptions fail.¶
Integrity and origin of a message. A valid signature over a request does not show that the request's effect is permitted now.¶
In scope: an architecture for controlling the transition from a proposed operation to its actual consequence; candidate interoperability requirements for implementations that claim specific properties; threat analysis; and a catalogue of consequence-specific profiles.¶
Non-goals:¶
Figure 1 shows the principal flow. The ten elements that the architecture relates are: (1) Originating Authority, (2) Candidate Act, (3) Non-Effective State, (4) Actual Pending Operation reconstruction, (5) current execution context and dependencies, (6) independent validation, (7) scoped execution authorization, (8) Protected Finality Sink, (9) actual resulting effect, and (10) authoritative outcome reconciliation.¶
+------------------------------+
| AI Agent / Software / Human |
| Operation Source |
+--------------+---------------+
| (1) propose
v
+------------------------------+
| Candidate Act |
| Proposed consequential act |
+--------------+---------------+
| (2) stage; no real effect yet
v
+------------------------------+
| Protected Non-Effective |
| State |
+--------------+---------------+
| (3) derive what would really happen
v
+------------------------------+
| Actual Pending Operation |
| Reconstruction |
+--------------+---------------+
| (4) bind current facts
v
+------------------------------+
| Source Authority + Context |
| Dependencies + Policy Epoch |
+--------------+---------------+
| (5) evaluate independently
v
+------------------------------+
| Independent Consequence- |
| Bound Validation |
+--------------+---------------+
|
v
+-------------+
| Valid now? |
+------+------+
NO | YES
+-------+-------+
v v
+--------------+ +------------------+
| HOLD / DENY | | Scoped Execution |
| / RECONCILE | | Authorization |
+--------------+ +--------+---------+
| (6) present at the boundary
v
+-------------------+
| Protected Finality|
| Sink |
+--------+----------+
| (7) sink re-derives and consumes
v
+-------------------+
| Actual Consequence|
+--------+----------+
| (8) observe at the destination
v
+-------------------+
| Outcome Evidence |
| & Reconciliation |
+-------------------+
Each numbered transition in Figure 1 is described with its mandatory condition, the trust assumption it relies on and its principal failure. Unless stated, a failed condition results in HOLD or DENY (fail closed).¶
Condition: the proposal is recorded with a stable act identity and the source's authority envelope. Trust: none placed in the source's own description of the operation. Failure: source supplies an inaccurate description, or an unbounded number of proposals (rate and quota limits apply).¶
Condition: staging produces no governed consequence; staging area is itself protected from tampering by the source after submission. Trust: staging store integrity. Failure: staged act modified after validation; staged act becomes effective through an unmodelled path (route coverage failure).¶
Condition: the effective principal, target, parameters, dependencies and derived child effects are derived from current resource state, not copied from the proposal. Trust: the reconstruction logic and its inputs are accurate. Failure: incomplete reconstruction (an effect branch is not modelled), reconstruction on stale state, or state changing between reconstruction and effect (time-of-check to time-of-use).¶
Condition: the current Source-Authority Envelope, material dependencies, and policy and revocation epochs are bound to the operation. Trust: epoch sources are authentic and fresh. Failure: epoch source unreachable or stale (treated as not current).¶
Condition: the validator is independent of the source and evaluates the reconstructed operation, not the proposal. Trust: validator integrity. Failure: validator compromise or substitution (Section 12, T7); mitigated by attestation of the validator and by the sink's own re-derivation.¶
Condition: every required predicate for the consequence class is true. UNKNOWN is not true. Failure: a predicate cannot be evaluated (HOLD); a predicate is mis-specified (policy error; residual risk).¶
Condition: no capability is issued. RECONCILE applies where an earlier unresolved outcome must be classified first. Failure: denial of service by inducing HOLD (Section 12, T17); mitigated by bounded holds and safe-state rules.¶
Condition: a Scoped Execution Capability is created, bound to the act, scope, destination, epochs, validity period or consumption count, and the designated sink. Trust: capability is non-bearer, so possession alone does not authorize. Failure: capability replay or widening (T5).¶
Condition: the sink independently reconstructs the operation it is about to make effective, compares it with the capability, checks sink-local epoch and consumption state, then consumes the capability atomically with the effect where technically required. Trust: the sink controls the first effectual boundary. Failure: alternate route around the sink (T11), crash between consumption and effect (REQ-13).¶
Condition: the outcome is classified from Authoritative Outcome Evidence as committed, conclusively not committed, or unresolved. Trust: evidence originates at the destination or an independent observer. Failure: forged, missing, delayed or contradictory evidence (T8, T9); unresolved outcomes remain quarantined (REQ-12).¶
Figure 2 illustrates why the validator compares the authorized envelope with the reconstructed operation. The nominal target, derived children, effective principal, and epochs may differ between approval and effectuation.¶
WHAT WAS AUTHORIZED WHAT WILL ACTUALLY HAPPEN
(Source-Authority Envelope) (Actual Pending Operation)
+---------------------------+ +---------------------------+
| principal / delegator | | effective principal |
| permitted task class | | effective operation |
| nominal target "T" | | resolved target "T'" |
| scope, limits, purpose | | derived children, writes |
| policy epoch Pe=17 | | current epoch Pe=19 |
| revocation epoch Re=4 | | current Re=5 |
+-------------+-------------+ +-------------+-------------+
| |
+--------------+---------------+
v
+------------------------------+
| Consequence-bound compare |
| T == T'? children inside |
| scope? epochs current? |
| limits composed correctly? |
+--------------+---------------+
|
any mismatch | all hold
+---------------------+---------------------+
v v
+-------------+ +------------------+
| HOLD / DENY | | scope-limited |
| fail closed | | authorization |
+-------------+ +------------------+
Permission to stage an operation (to write it into a queue, change request or draft) MUST be treated, in any implementation claiming effect-commit control, as separate from permission to commit its effect. Permission to commit the effect is in turn separate from evidence that the intended effect was achieved: a continuation or customer-facing statement that depends on the outcome requires the outcome evidence (Section 6, REQ-07 and REQ-12).¶
Figure 3 shows trust boundaries B1 (protected enforcement domain), B2 (Finality Sink) and B3 (effect domain).¶
+-----------------------------------------------------------------+
| UNTRUSTED / PARTIALLY TRUSTED ZONE |
| model, agent runtime, orchestrator, tool adapters, retry |
| workers, caches, application-side gateways |
+--------------------------------+--------------------------------+
| Candidate Act (no authority)
v
+-----------------------------------------------------------------+
| PROTECTED ENFORCEMENT DOMAIN (trust boundary B1) |
| +-------------------+ +--------------------------------+ |
| | Independent |<--->| Protected state | |
| | Validator | | policy epoch, revocation | |
| | (re-derives facts)| | epoch, consumption record, | |
| +---------+---------+ | exposure ledger, commitments | |
| | +--------------------------------+ |
| | Scoped Execution Capability (non-bearer) |
+------------+----------------------------------------------------+
v
+-----------------------------------------------------------------+
| FINALITY SINK (trust boundary B2) |
| verifies capability vs ACTUAL operation, checks sink-local |
| epoch/consumption, then admits or rejects |
+--------------------------------+--------------------------------+
| only admitted operations
v
+-----------------------------------------------------------------+
| EFFECT DOMAIN: database, network element, payment rail, device, |
| actuator, accelerator, message channel (trust boundary B3) |
+-----------------------------------------------------------------+
Assumption: B2 is the first boundary where the act can become
externally effective. If another route reaches B3 around B2, the
route-coverage requirements (REQ-09, REQ-16) are violated.
AI agent, software component or human. Untrusted for the purpose of describing its own operation. May hold identity credentials and tool-invocation permissions.¶
Holds Candidate Acts in the Non-Effective State. Integrity-protected; the source can append but not mark an act validated.¶
Derives the Actual Pending Operation and Material Dependency Graph from current state. May be a part of the validator or of the sink, or both; the sink MUST be able to re-derive for the operation it admits where the sink claims to enforce REQ-01 to REQ-04.¶
Evaluates predicates and issues the Scoped Execution Capability. Runs in a Protected Enforcement Domain. May be co-located with the sink for latency (see Section 7).¶
Epochs, consumption records, exposure ledgers, watermarks, commitments. Crash-consistent where replay protection is claimed.¶
The enforcement point at the first effectual boundary. Holds the credential or capability to cause the effect; the operation source does not.¶
Produces Authoritative Outcome Evidence. Independent of the operation source.¶
Classifies unresolved outcomes, authorizes retry, compensation or recovery under current authority.¶
Evaluates evidence of a bounded real effect before continuation authority is created (Figure 4).¶
This section lists candidate interoperability requirements. Each applies only to an implementation that elects to claim the named property. Each has an enforcement rationale and a testable failure condition. The tag PROFILE denotes a candidate profile requirement; ARCH denotes an architectural recommendation. No mandatory wire format, registry or transport is imposed.¶
PROFILE¶
An implementation claiming this property MUST derive a stable act identity from a canonical representation of the operation, such that two representations that cause the same effect map to the same identity and a materially different effect maps to a different one. The canonicalization MUST be deterministic and MUST be applied at validation and again at the sink.¶
Without a canonical form, an attacker or a translation layer can present a different serialization to the validator and to the sink (semantic or encoding substitution).¶
Present the sink with an operation that differs only in a field the validator ignored; the sink MUST reject. Present two encodings of one effect; both MUST map to one identity.¶
E1-E38; T1, T14¶
PROFILE¶
The validator MUST evaluate the current Source-Authority Envelope including the delegation chain, and MUST ensure that each delegation narrows or preserves, and never widens, the permitted scope. Restrictions present at origin MUST remain binding after summarization, handoff, forwarding or translation.¶
Authority that is reconstructed from model-visible or forwarded content can be lost or invented in transit; authority must come from a record the transforming component cannot rewrite.¶
Delegate A to B to C and have C request a wider scope; the request MUST be denied. Summarize context and drop a prohibition; the later effect MUST still be denied.¶
E4, E15, E17, E20, E24, E36; T3, T1¶
PROFILE¶
The sink MUST resolve and compare the actual destination and effective target (including effective principal) with those bound in the capability. A nominal target name MUST NOT be accepted where resolution can change between approval and effect.¶
Aliases, redirects, re-resolution, account reassignment and target drift are the commonest TOCTOU form.¶
Change DNS, alias or account mapping after validation; the sink MUST reject the changed destination.¶
E22, E23, E24, E28, E34; T6¶
PROFILE¶
The capability MUST bind the policy, revocation and (where relevant) configuration epochs under which it was issued, and the sink MUST reject a capability issued under an epoch older than the sink's current epoch for that consequence class.¶
Epoch binding makes revocation and policy change effective for already issued but unconsumed authority.¶
Advance the policy epoch after issuance; presenting the capability MUST fail.¶
E3, E9, E10, E13, E25; T4, T5, T13¶
PROFILE¶
A capability MUST be single-use or bounded-use as designated, and the sink MUST record consumption so that a second presentation, including from a restored replica or a retry worker, is rejected.¶
Possession of a valid capability must not allow repeated effect.¶
Present the same capability twice, including after a failover; the second MUST be rejected.¶
E5, E14, E32, E33; T5, T13, T15¶
PROFILE¶
Where revocation is claimed to be effective, the sink MUST check the current Revocation Epoch (or equivalent) at effectuation, and MUST define the maximum staleness it tolerates for revocation data.¶
Revocation that is only evaluated at issuance does not bind authority already in flight.¶
Revoke after issuance and before presentation; presentation MUST fail within the stated staleness bound.¶
E1, E3, E9; T4, T5¶
PROFILE¶
Evidence relied upon for a validation or continuation decision MUST originate outside the operation source, MUST be bound to the act and phase, and MUST be checked for coverage and freshness (including, for absence-of-event evidence, sensing opportunity). Self-reported completion by the source MUST NOT be accepted as Authoritative Outcome Evidence.¶
Evidence the source can write or withhold gives the source control of the decision that is supposed to constrain it.¶
Have the source write a PASS record; it MUST NOT advance the act. Remove a sensor from coverage; a 'no event' inference MUST NOT be admitted.¶
E12A, E14, E16, E19, E27; T8, T9¶
PROFILE¶
The authorization predicate MUST be evaluated over the reconstructed effect, including aggregate and transitive effects relevant to the consequence class, and the capability's scope MUST be no wider than the validated effect.¶
A request-level check can pass while the combined or derived effect is outside scope.¶
Submit two individually acceptable work products whose combination violates a limit; the aggregate MUST be withheld.¶
E2, E4, E7, E21, E26, E31, E33; T3¶
PROFILE¶
Enforcement MUST be located at the actual effectuation boundary for each covered route. A deployment claiming coverage MUST identify the routes it covers and MUST demonstrate, for each, that an operation lacking valid authority is rejected at the sink.¶
Controls located upstream of an alternate route do not constrain the alternate route.¶
Invoke the effect through an alternate API, credential or hardware path; if the route is declared covered, the sink MUST reject.¶
E25, E10, E36; T11, T12¶
ARCH¶
Implementations SHOULD allow the capability to be realized by cryptographic means (for example signed or MACed tokens), by protected local state, by hardware interlocks, or by a hybrid, and SHOULD state which is used for which consequence class. Fast paths SHOULD use pre-validated local state with cryptographic validation at lease creation, renewal, change and revocation rather than per event.¶
A single mechanism cannot meet all latency and trust constraints; the property to preserve is the binding, not a specific cipher.¶
Run the same scenario under cryptographic and local-state realizations; observable accept/reject outcomes MUST be equivalent.¶
E1, E5, E10, E13, E38; T7¶
ARCH¶
Where a consequence is progressive, implementations SHOULD permit a bounded real effect followed by authenticated effect evidence and a continuation condition (Figure 4). A later phase MUST NOT exceed the maximum authorized cumulative envelope merely because earlier phases succeeded.¶
Evidence of a small real effect can justify a larger one without granting authority for the whole session.¶
Request a next phase exceeding the remaining envelope; it MUST be denied even after a successful first phase.¶
E37, E38; T8¶
PROFILE¶
A timeout, lost acknowledgment, worker failure or restart MUST NOT be treated as evidence that an effect did not occur. The implementation MUST keep a logical effect identity distinct from attempt identity, MUST hold conflicting continuations while the outcome is unresolved, and MUST classify the outcome from Authoritative Outcome Evidence before authorizing retry, compensation or recovery.¶
Blind retry after an unknown outcome can duplicate an irreversible effect.¶
Drop the acknowledgment of a committed payment; a retry MUST NOT be admitted until the outcome is reconciled.¶
E14, E32, E33, E34; T9, T15, T18¶
PROFILE¶
Where single-use is claimed, consumption of the capability and the governed effect MUST be atomic or recoverable such that a crash cannot yield both a retained valid capability and a completed effect, nor a consumed capability without a recorded effect state.¶
Non-atomic consume-then-act or act-then-consume orders create duplicate or lost effects across crashes.¶
Kill the sink between consume and effect and between effect and record; on recovery the state MUST be classifiable and no duplicate MUST occur.¶
E5, E14, E32; T5, T13¶
PROFILE¶
Where an operation, policy or authority contract is translated between protocols, engines or platforms, the implementation MUST verify, per consequence class, that the destination preserves the semantics the contract requires, and MUST fail closed (deny or declare uncovered) for features the destination does not preserve.¶
Syntactic acceptance of a translated request does not show that restrictions survived.¶
Translate a request with a prohibition into a protocol lacking that feature; the result MUST be denial or an explicit uncovered declaration.¶
E20, E30; T14¶
PROFILE¶
Where a tool runs in a provider-hosted environment that an application-side gateway cannot intercept (Figure 6), authority MUST be enforced at the end-target sink, or the tool MUST be disabled for covered consequence classes, or the path MUST be declared uncovered.¶
An application-side control does not see a call that never passes through the application.¶
Enable a hosted tool that calls a covered target; the end target MUST reject without valid authority, or the tool MUST be unavailable.¶
E20, E25; T10, T11¶
PROFILE¶
A deployment claiming non-bypassability MUST perform and document an alternate-path analysis covering credentials, APIs, hardware paths, management planes, retry workers and administrative routes, and MUST treat any known act-equivalent path as disabled, mediated, restricted or declared uncovered.¶
Claims of non-bypassability are only as strong as the route inventory.¶
Add a new credential with direct access to the target; a route-coverage test MUST fail or the credential MUST be rejected at the sink.¶
E25, E10, E6, E36; T11, T12¶
ARCH¶
Implementations SHOULD place full validation at slower consequential boundaries (management, configuration, authorization, settlement, deployment) and SHOULD use pre-validated bounded local state for low-latency loops, validating at lease creation, renewal, change and revocation. No latency figure is specified or claimed here.¶
Per-event cryptographic validation can be infeasible in microsecond loops; the control point must match the time scale of the consequence.¶
Measure that a revocation takes effect within the declared bound for a fast path without per-event signature verification.¶
E2, E10, E13, E38; T17¶
The architecture distinguishes two time scales. First, consequential management, configuration, authorization, settlement and deployment boundaries, where a validation step can be placed before the effect without making the operation infeasible. Second, low-latency radio, scheduling and safety-control loops, where a per-event cryptographic check may be infeasible. For the second class this document proposes only that validation occur when protected local state is created, renewed, changed or revoked, with the fast path enforcing the resulting bounded state using local protected checks (Section 6, REQ-10 and REQ-17). No latency figure is stated or implied. Whether a given loop can be covered at all, and with what safety consequences, is a matter for measurement in the deployment and for the safety case of that system.¶
Safety consequence: for cyber-physical effects, validator unavailability MUST NOT translate into an unsafe default. The safe state (hold, continue last safe configuration, or stop) is system specific and is chosen by the safety case, not by this document.¶
The following profiles apply the architecture to specific boundaries. They are illustrative of disclosed mechanisms and are not wire protocols.¶
Where a consequence is progressive, a first phase produces a bounded real effect. Authenticated evidence of that effect is evaluated, and only then is a continuation condition created for the next phase (Figure 4). Evidence is a structural dependency, not a log: the next sink verifies or consumes the continuation condition. This is the idea developed in the cited companion draft on receipt-gated IEV; it is included here as an optional profile and not redefined. Unknown or indeterminate evidence holds the act; it never defaults to pass.¶
Candidate Act A, Envelope_MAX
|
v
+--------------------+ scope1 <= Envelope_MAX
| Validation Phase 1 |--------------------------+
+---------+----------+ |
v |
+--------------------+ |
| Finality Sink FS1 | |
+---------+----------+ |
v |
[ BOUNDED REAL EFFECT E1 ] (not a simulation) |
| |
v |
+--------------------------+ |
| Effect Evidence R1 | destination- |
| (authenticated, bound to | origin, not |
| act, phase, nonce) | self-reported |
+------------+-------------+ |
v |
+--------------------------+ |
| Interim Effectuation | expected X1 vs |
| Validation (IEV) | observed O1 |
+------+-----------+-------+ |
PASS | | FAIL / UNKNOWN |
v v |
+-------------+ +-----------------------+ |
| Continuation| | HOLD / SAFE STATE / | |
| condition | | RECONCILE (no blind | |
| G2 created | | retry) | |
+------+------+ +-----------------------+ |
v |
+--------------------+ verifies/consumes G2 |
| Finality Sink FS2 |<-------------------------+
+--------------------+ scope2 <= Envelope_MAX - spent
When several agents produce work products, each may pass local validation while their composition does not. The profile materializes the composition at a shared commit boundary, evaluates cross-branch reads and writes, negative constraints and cumulative limits, and authorizes the accepted aggregate or a validated subset (Figure 5).¶
+---------+ +---------+ +---------+
| Agent A | | Agent B | | Agent C |
| WP-A ok | | WP-B ok | | WP-C ok |
+----+----+ +----+----+ +----+----+
| locally valid, each one |
+------------+-----------------+
v
+-------------------------+
| Composition snapshot | materialized merge result,
| (shared commit boundary)| not three separate diffs
+------------+------------+
v
+-------------------------+
| Aggregate validator | cross-branch reads/writes,
| | negative constraints,
+------+-------------+----+ cumulative limits
reject| |accept (whole or valid subset)
v v
+-----------+ +-------------------+
| withhold | | scoped capability |
| aggregate | | for accepted state|
+-----------+ +---------+---------+
v
+-------------------+
| Merge/commit sink |
+-------------------+
Conversion between model APIs, MCP adapters and provider-hosted or client-managed tool execution can drop restrictions or change effect semantics. The profile keeps a canonical authority contract outside the model context, a feature and exception matrix, and an effective tool-location map. For provider-hosted tools that an application-side gateway cannot intercept, enforcement moves to the end-target sink (Figure 6). This document does not assert which specific providers expose which hosted-tool modes; that must be verified per provider and version.¶
+----------------------------------------------+
| Canonical authority contract (outside model |
| context): prohibitions, scope, tool-location |
| map, feature/exception matrix |
+--------------------+-------------------------+
| consulted at every hop
Agent runtime v
+----------+ +-------------+ +-----------------------+
| model |-->| protocol |-->| tool-call dispatch |
| output | | bridge / | +-----+-----------+-----+
+----------+ | MCP adapter | | |
+-------------+ CLIENT-MANAGED PROVIDER-HOSTED
| |
v v
+-------------------+ +-------------------------+
| app-side gateway | | provider executes tool; |
| can intercept | | app gateway cannot |
+---------+---------+ | intercept the call |
| +------------+------------+
v |
+-------------------+ |
| tool server / MCP | |
+---------+---------+ |
+-------------+------------+
v
+-------------------------------+
| END-TARGET FINALITY SINK |
| (independent of both paths) |
+-------------------------------+
Rule: where a path cannot be intercepted, authority must be
enforced at the end target, or the path must be disabled or
declared uncovered.
A software decision that is compiled into wildcarded, prioritized, cached or stateful hardware rules may have a different real effect from the decision. The profile compares the effective installed behavior (read back from the device where supported) with the permitted effect and fences stale fast-path authority on policy change (Figure 7). Compared with existing practice (P4 runtime verification, match-action verification, consistent updates), the distinguishing element to examine is the binding of installed-effect equivalence to the software authorization and to an authority epoch (Appendix D).¶
Slow path (software) Fast path (hardware)
+---------------------------+ +---------------------------+
| Slow-Path Authorization | | Effective Hardware Action |
| Witness W | | Descriptor H |
| + Offload Effect Contract | | (wildcards, priority, |
| permitted behavior | | caches, state, actions) |
+-------------+-------------+ +-------------+-------------+
| compile / install |
+--------------+-------------------+
v
+-----------------------------+
| Offload-equivalence check: |
| effects(H) subset-of |
| permitted(W, Contract)? |
| + Fast-Path Authority Epoch |
+--------+-----------+--------+
FAIL | | PASS
v v
+-------------+ +----------------------+
| do not | | enable offload; read |
| enable; | | back installed rules |
| remove rule | | and compare |
+-------------+ +-----------+----------+
v
+--------------------------+
| epoch change / revocation|
| => stale fast-path rules |
| fenced (barrier) |
+--------------------------+
A dashboard, policy document or compliant sample trace does not show that every effect-producing route is controlled. The profile builds a route inventory, relates required controls to actual enforcement points, and obtains evidence of rejection at the sink for each route, using non-consequential probes where safe (Figure 8). Routes that cannot be probed safely or covered are declared uncovered, and authority for the associated consequence classes is lowered accordingly.¶
+-------------------+ +----------------------------------+
| Route inventory | | Governance coverage contract |
| APIs, credentials,| | required control per effect class|
| hardware, network | +----------------+-----------------+
| paths, tools | |
+---------+---------+ |
+-------------------+-------------+
v
+-------------------------------+
| Governance-to-effect graph |
+---------------+---------------+
v
+-------------------------------+
| Coverage probe: attempt a |
| non-consequential test effect |
| on each route (where safe) |
+--------+---------------+------+
sink | | effect got through
rejects | | without control
v v
+--------------+ +--------------------+
| route COVERED| | route UNCOVERED: |
| (evidence) | | disable, restrict, |
+--------------+ | or lower authority |
+--------------------+
Logs and dashboards are not coverage evidence; sink rejection is.
A permitted stored procedure can reach triggers, definer-rights code, autonomous transactions and external functions beyond its entry point (E31). A downstream consumer of an outbox or change-data-capture event should not treat an uncommitted intent, a stale outbox row or an unrelated transaction's event as authority; release is bound to a source-commit certificate and current effect state, with reconciliation after uncertain acknowledgments (E32). Figure 9 shows the flow. The transactional outbox and CDC are established patterns ([OUTBOX]); the profile concerns binding consumer effect authority to the source commit.¶
+--------------------+ txn +-----------------------------+
| Agent / app |------>| Source database |
| proposes write | | (stored programs, triggers) |
+--------------------+ +--------------+--------------+
|
transitive effect closure of the | COMMIT
entry point is reconstructed v
+-----------------------------+
| Commit log / outbox / CDC |
| source-commit certificate |
+--------------+--------------+
uncommitted intent X no authority |
stale outbox row X no authority v
unrelated txn event X no authority +----------------------------+
| Downstream consumer |
| derives child operation |
+-------------+--------------+
v
+----------------------------+
| Effect Finality Sink |
| checks: commit certificate,|
| policy epoch, child scope |
+-------------+--------------+
v
+----------------------------+
| External effect + ack |
| uncertain ack => reconcile |
+----------------------------+
At the non-real-time management plane, an AI-generated configuration or slice change is reconstructed into its effective multi-cell and slice state, validated, and applied as a limited real change; attributable impact evidence bounds expansion (Figure 10). Slice capacity lending (E38) additionally requires conservation (lending creates no capacity), preservation of the lender's entitlement floor and an enforceable reclaim bound. Real-time loops are out of scope and use pre-validated local state (Section 7). Relevant established mechanisms include 3GPP network slicing and the O-RAN architecture ([TS23501], [ORAN-ARCH]).¶
NON-REAL-TIME / MANAGEMENT PLANE (in scope)
+--------------+ +----------------+ +---------------------+
| rApp / AI |-->| SMO / Non-RT |-->| Candidate config |
| optimizer | | RIC policy | | or slice change |
+--------------+ +----------------+ +----------+----------+
v
+-------------------------------+
| Reconstruct effective state: |
| multi-cell, neighbors, slices,|
| affected-service footprint |
+---------------+---------------+
v
+-------------------------------+
| Validator + Finality Sink at |
| the configuration-effect point|
+---------------+---------------+
v
+-------------------------------+
| Limited real change (canary) |
+---------------+---------------+
attributable impact evidence|
v
+-------------------------------+
| Causal impairment evaluator |
| expand only if bounded impact |
+-------------------------------+
OUT OF SCOPE: PHY / MAC / HARQ and other real-time loops. Those use
prevalidated local state; validation occurs at lease or policy
creation, renewal, change and revocation, not per transport block.
Multiple pending authorizations, captures and agent-issued instructions can collectively exceed a mandate. The profile maintains a conservative exposure ledger, reserves capacity before admitting an operation, and keeps unresolved exposure held until authoritative evidence resolves it (Figure 11). Existing network and issuer controls remain authoritative; the profile is complementary. For cross-border payments the nominal instruction, settlement stages and beneficiary-net outcome are distinguished (E34, with ISO 20022 tracking as established context [ISO20022]). For offline digital cash, E35 addresses delegated spending authority and post-compromise fork reconciliation and does not claim that secure-element protection is ineffective under intact-hardware assumptions.¶
+---------------------------+
| Protected Exposure Mandate| principal, agents, aggregate limit,
+-------------+-------------+ beneficiary/merchant restrictions
v
+---------------------------+ +--------------------------+
| Exposure ledger |<---->| reserved | authorized | |
| conservative accounting | | captured | released | |
+-------------+-------------+ | UNRESOLVED (kept held) |
v +--------------------------+
+---------------------------+
| Reserve capacity, then | capacity must exist BEFORE the
| issue transaction-bound | operation is admitted
| capability |
+-------------+-------------+
v
+---------------------------+ authorization --> capture -->
| Payment finality sink | clearing --> settlement
+-------------+-------------+
v
+---------------------------+ timeout or lost ack is NOT failure
| Outcome evidence: |------------+
| settlement / beneficiary | v
| net amount, FX, fees | +--------------------------+
+-------------+-------------+ | quarantine; query |
v | authoritative status; |
+---------------------------+ | retry/compensate only |
| Reconcile ledger |<--| under current authority |
+---------------------------+ +--------------------------+
Figure 12 shows the state machine with the attack and recovery paths. Every transition not drawn defaults to DENIED or HELD.¶
stage
[PROPOSED] ------------> [NON_EFFECTIVE]
| reconstruct
v
[RECONSTRUCTED]--substitution/mismatch--+
| validate |
v v
[VALIDATING]--stale epoch/revoked-->[DENIED]
| all predicates true ^
v |
[AUTHORIZED]--expiry-->[EXPIRED] |
| present at sink |
v |
replay / consumed -->[SINK_CHECK]---mismatch/no cap--------+
capability = reject | admit + consume (atomic)
v
[EFFECTUATING]
| |
effect evidence | | validator or evidence
v | unavailable
[CONFIRMED] v
[HELD] <---- [INDETERMINATE] (timeout or
| ^ lost ack)
+-- reconcile--+
| authoritative evidence
+--> [CONFIRMED] or [NOT_EFFECTED]
Every arrow not drawn defaults to DENIED or HELD (fail closed).
Retry from INDETERMINATE requires fresh authority and
reconciliation.
| State | Entered when | Leaves to |
|---|---|---|
| PROPOSED | Source submits act | NON_EFFECTIVE |
| NON_EFFECTIVE | Act staged; no consequence | RECONSTRUCTED |
| RECONSTRUCTED | Actual Pending Operation derived | VALIDATING; DENIED (mismatch) |
| VALIDATING | Predicates being evaluated | AUTHORIZED; DENIED; HELD |
| AUTHORIZED | Capability issued | SINK_CHECK; EXPIRED |
| SINK_CHECK | Capability presented at sink | EFFECTUATING; DENIED |
| EFFECTUATING | Capability consumed with effect | CONFIRMED; INDETERMINATE; HELD |
| INDETERMINATE | Timeout, lost ack, partial effect | HELD; CONFIRMED; NOT_EFFECTED |
| HELD | Validator or evidence unavailable, or quarantine | VALIDATING (fresh); DENIED |
| CONFIRMED / NOT_EFFECTED / DENIED / EXPIRED | Terminal for this act identity | (new act required) |
Retry from INDETERMINATE is not a transition back to AUTHORIZED. It requires a fresh validation under current authority after reconciliation, and it carries the same logical effect identity so that a duplicate is detectable (Section 6, REQ-12).¶
Because no wire format is defined, interoperability here means preserving the binding between act, effect, authority and epoch across components from different vendors. Candidate carriers for a Scoped Execution Capability include CWT/COSE and JWT ([RFC8392], [RFC9052], [RFC7519]); sender-constraint mechanisms exist ([RFC9449], [RFC8705]); attestation evidence may use RATS roles and EAT ([RFC9334], [RFC9711]). Existing deployments can often remain unchanged above and below the sink, provided the effect-holding credential is held by the sink rather than by the source. Where a component cannot be integrated, the route MUST be declared uncovered rather than assumed protected (REQ-16).¶
Legacy components that may remain unchanged include identity providers, existing RBAC/ABAC policy engines, transaction logs and message brokers; the architecture adds a binding and a check at the effect point. Whether any particular legacy component can host a sink is a deployment question this document does not answer.¶
Operations HOLD; no default allow. Fast paths continue under pre-validated bounded state until its expiry, then fail to the declared safe state.¶
The epoch is treated as not current for consequence classes that require it.¶
Recovery classifies the act from sink and destination state before any retry (REQ-13).¶
Outcome is INDETERMINATE; competing operations are held; reconciliation against the authoritative outcome criterion precedes retry or compensation (REQ-12).¶
Treated as unresolved; escalated to human or multi-party resolution as the profile specifies; never defaulted to success.¶
Unconsumed capability is rejected at the sink; in-flight effects follow the profile's fencing rule (for example E12 for queued segments, E3 for desired-state controllers).¶
Contain, hold further phases, move to safe state, then reconcile; compensation requires fresh authority.¶
Where lending or delegation was made (E38, E35), reclamation or fork reconciliation proceeds under the profile's deadline and evidence rules.¶
The adversary may control or influence the operation source and its inputs, may observe and replay messages, may race operations, may obtain credentials for routes that bypass the intended path, and may attempt to disable the validator. The adversary is not assumed to control both a genuine validator and a genuine sink simultaneously; if it does, the architecture offers no protection. The architecture does not promise that it prevents every attack; each threat below lists residual risk.¶
Controls or influences the model output; can propose arbitrary Candidate Acts and describe them misleadingly.¶
Operation source to protected enforcement domain (B1).¶
No act becomes effective unless the reconstructed operation is validated against current authority.¶
Validation over the reconstructed operation (REQ-01, REQ-08); sink re-derivation (REQ-03); quotas on proposals.¶
an operation within valid scope but harmful in intent; mandate limits bound the harm but do not remove it.¶
A compromised agent submits an act whose description says 'read' but whose effect is 'write'; the sink MUST reject.¶
Places instructions in retrieved content, tool results or messages that the model treats as authority.¶
Model context versus authority record.¶
Instructions found in data never create or widen authority.¶
Authority taken from a record outside the model context (REQ-02); quoted or forwarded content not operative (E24); tool output treated as data.¶
injection may still cause an in-scope but unwanted operation.¶
Inject 'ignore prior limits' via tool output; subsequent act beyond the original scope MUST be denied.¶
Uses a delegation chain, handoff or summarization to obtain a wider scope than the originator held.¶
Delegation boundaries and context transformations.¶
Each hop narrows or preserves scope.¶
REQ-02; authority record survives compaction (E15); human capsule minimizes disclosure (E17); aggregate and transitive checks (REQ-08).¶
legitimately broad originating authority.¶
Chain A to B to C where C asks for broader scope; MUST be denied; summarization removes a prohibition; later effect MUST still be denied.¶
Changes target, configuration, dependency topology or policy after validation.¶
Time gap between B1 decision and B2/B3 effect.¶
Effect occurs only under the state that was validated or a state that is at least as restrictive.¶
Short validity; re-derivation at the sink (REQ-03); epoch binding (REQ-04); live-topology revalidation (E28).¶
change after the sink's check but before the effect in systems without atomic commit.¶
Change a dependency edge after validation; the sink MUST reject or re-validate.¶
Copies, replays or retains a capability; restores a replica that lost consumption state.¶
Capability store and sink-local state.¶
A capability produces at most its designated effect.¶
Non-bearer binding; consumption record (REQ-05, REQ-13); epoch binding (REQ-06).¶
loss of consumption state without detection; mitigated by fencing and state commitments.¶
Present the same capability to a failed-over replica; the second use MUST be rejected.¶
Changes the resolved destination, account or principal behind a nominal name.¶
Name resolution and identity mapping at B2.¶
Effect binds to the destination and principal validated.¶
REQ-03; bind to resolved identity not nominal name.¶
legitimate re-assignment between sink check and effect.¶
Reassign an account after validation; the sink MUST reject.¶
Replaces or controls the validator, or forges its decisions.¶
B1.¶
Only a genuine validator in a genuine domain can issue authority; the sink does not rely solely on validator output.¶
Attestation of the validator ([RFC9334]); sink-side re-derivation; multi-validator quorum for high-consequence classes; bounded capability scope.¶
a compromised validator and a compromised sink together defeat the control.¶
Substitute a validator with an unattested build; sinks MUST refuse its capabilities.¶
Forges or reuses a receipt, or writes a self-reported 'done'.¶
Evidence path from destination to validator.¶
Only evidence bound to the act, phase and nonce and originating outside the source advances an act.¶
REQ-07; single-use receipts; binding to act and phase; destination-origin evidence.¶
a compromised destination.¶
Replay a prior receipt for a new phase; MUST be rejected.¶
Suppresses or delays evidence, or induces contradictions among sources.¶
Evidence collection.¶
Absence or conflict is never treated as success or as failure.¶
REQ-07, REQ-12; hold and quarantine; reconciliation against authoritative criteria (E14, E19).¶
prolonged unresolved outcome ties up exposure.¶
Withhold evidence; the act MUST remain HELD and no retry MUST be admitted.¶
Induces a privileged tool or service to act on the attacker's behalf.¶
Privileged tool versus act scope.¶
A tool's privilege never exceeds the scope of the act on whose behalf it runs.¶
Credential isolation; capability bound to act and tool; REQ-15.¶
tools that cannot be wrapped.¶
Have a privileged tool accept an act-scoped request that asks for broader privilege; MUST be denied.¶
Reaches the target through a route the sink does not mediate.¶
Route inventory versus enforcement points.¶
Every covered route reaches the effect only through a sink.¶
REQ-09, REQ-16; route discovery and rejection evidence (E25).¶
undiscovered routes.¶
Add a credential with direct target access; the coverage test MUST flag it or the sink MUST reject its use.¶
Uses DMA, residual translations, offloaded rules or accelerator queues to cause an effect outside software checks.¶
Software authorization versus installed hardware behavior.¶
Installed behavior is within the authorized effect and prior-domain access paths are closed before reassignment.¶
E6 prior-path closure; E10 equivalence and readback; fast-path epoch.¶
hardware that cannot report installed state.¶
Install a rule broader than authorized; enablement MUST be refused. Reassign a device with an open DMA mapping; transition MUST be withheld.¶
Restores a checkpoint or fails over so that two workers hold effect authority.¶
Worker identity and ownership epoch.¶
At most one worker holds effect-commit authority for a logical computation.¶
E5 source fence, destination admission, effect watermark; REQ-05, REQ-13.¶
loss of fencing state.¶
Restore a checkpoint while the source still runs; the source's publish MUST be rejected after cutover.¶
Exploits translation that drops restrictions or changes effect semantics.¶
Protocol bridges.¶
Translated operation preserves authority semantics or fails closed.¶
REQ-14, REQ-15; feature and exception matrix; end-target enforcement.¶
unknown unsupported features.¶
Translate a prohibition into a protocol lacking it; result MUST be denial or declared uncovered.¶
Races concurrent agents or retries to exceed a mandate or duplicate a payment.¶
Exposure ledger and payment sink.¶
Reserved plus committed plus unresolved exposure never exceeds the mandate.¶
E33 reservation before admission; E14 retry quarantine; conservative unresolved accounting.¶
counterparty behavior outside the ledger.¶
Run N concurrent agents; the sum MUST not exceed the mandate; lose an ack; retry MUST be held.¶
Extracts keys or clones state from an offline secure element.¶
Secure element assumptions.¶
Conflicting transfers become detectable and loss is bounded by allocation limits.¶
E35 bounded child allocations, lineage and recipient evidence, fork reconciliation.¶
undiscovered conflicting transfers; privacy cost of reconciliation.¶
Clone state and spend twice offline; reconciliation MUST identify the fork and the bounded exposure.¶
Floods or disables validators or epoch sources to stall operations or force unsafe defaults.¶
Availability of B1.¶
Unavailability yields HOLD/safe state, never implicit allow.¶
Redundant validators; local pre-validated bounded state for fast paths; bounded hold and safe-state rules; rate limits.¶
availability loss is a real cost; fail-closed trades availability for safety.¶
Disable the validator; consequential operations MUST be held and no default allow MUST occur.¶
Causes or exploits a partially completed physical or irreversible effect.¶
Effect domain.¶
Partial effect is classified and contained before any continuation.¶
Bounded real effect; hold; authoritative outcome criteria; safe-state; compensation only under current authority (REQ-11, REQ-12).¶
some effects cannot be undone.¶
Interrupt a multi-step actuation; further steps MUST be held pending classification.¶
This document follows the guidance of [RFC3552]. The threat analysis in Section 12 is part of these Security Considerations. Additional points:¶
This section follows [RFC6973]. The architecture creates records that can contain personal or confidential data and can enable correlation:¶
This document has no IANA actions. A separate individual draft by the same author, draft-das-ef-registries, proposes registries for the execution-finality family; nothing in this document depends on that proposal.¶
This section credits established work and states candidate differences without asserting originality. Patentability arguments (motivation to combine, reasonable expectation of success, technical interdependence) are kept in Appendix D and are not part of the proposed architecture.¶
Established function: Attestation evidence, verifier appraisal and relying-party policy.¶
Overlap and candidate difference: E9 concerns correspondence between remediation, evidence and policy per consequence class, and E10/E6 bind attested or installed state to specific effects. It uses RATS roles and does not replace them.¶
Established function: Delegated authorization, sender constraint, fine-grained authorization details.¶
Overlap and candidate difference: The candidate additions are re-derivation of the effective operation at the resource boundary, binding to epochs and consumption, and route-coverage evidence. A RAR authorization detail could carry part of the scope.¶
Established function: Claim and signature formats.¶
Overlap and candidate difference: Candidate carriers only; no format is selected.¶
Established function: Prevent stale holders from writing after lease change.¶
Overlap and candidate difference: E5 and E3 use epoch fencing for effect-commit authority of a computation and for reconciliation generations. The difference to examine is the combination with checkpoint cutover, destination admission and an effect watermark.¶
Established function: Atomic state change plus reliable event publication; CDC.¶
Overlap and candidate difference: E32 binds downstream effect authority to the source commit and to effect state; E31 addresses transitive stored-program effects. Established patterns are credited; the examined element is the authority binding.¶
Established function: Per-hop identity and policy.¶
Overlap and candidate difference: E36 traces parent authority across subrequests, queues and background workers and enforces at outbound effect paths. Mesh policy is closely related and must be compared.¶
Established function: P4Runtime, match-action verification, consistent updates.¶
Overlap and candidate difference: E10 compares installed effective behavior with a software-authorized effect and fences stale fast-path authority. Existing verification work overlaps and must be compared.¶
Established function: Access steering, mobility hysteresis, slice isolation and resource management.¶
Overlap and candidate difference: E7 evaluates cumulative connectivity trajectory; E38 evaluates slice lending with conservation and enforceable reclaim. 3GPP mechanisms are the primary comparison.¶
Established function: Offline double-spend prevention under intact-hardware assumptions.¶
Overlap and candidate difference: E35 considers delegated allocations and post-compromise fork reconciliation. It does not claim that existing designs lack double-spend protection under their own assumptions.¶
Established function: Principle of continuous verification; role and attribute decisions.¶
Overlap and candidate difference: Consistent with and layered over; the added element is verification at the actual effect boundary.¶
This section is non-normative. Per Source C and the applicant's description, the June 2026 mothership (PCT/IB2026/055615) already describes the broad protected execution-finality architecture, including Candidate Acts, non-effective staging, protected evidence and receipts, cryptographic validation, Finality Sinks, and accelerator and network egress control in many domains. The later Indian provisional (202631136555, 15 November 2026) is a specialized disclosure that examines narrower implementations and combinations.¶
Limit of this comparison: the editor did not have the mothership JSON text. Statements about what the mothership contains are taken from the applicant's comparison and are not independently verified. This document does not claim that any feature below is absent from the original text; a feature not appearing under an exact phrase may still be disclosed expressly or inherently. Distinguishing a disclosure refinement from independently established novelty or priority is a matter for examination against the as-filed text and relevant prior art.¶
Technical predicate: Source fence, destination admission and effect watermark give a single effect-commit owner.¶
Actual effect governed: Externally authoritative compute output (publish or model-state commit) during migration or failover.¶
Claimed functional interdependence: Checkpoint transfer, fencing and the sink's admission check must cooperate; copying state alone confers no authority.¶
Technical predicate: Evidence that old-domain DMA, interrupt, debug and queued-I/O paths are closed.¶
Actual effect governed: Reassignment of a peripheral, memory region or accelerator to another trust domain.¶
Claimed functional interdependence: Successor authority is withheld until path-quiescence evidence from the controller and the translation hardware is joined.¶
Technical predicate: Cumulative effect of transition history compared with a temporal envelope.¶
Actual effect governed: Discretionary access transitions of a device.¶
Claimed functional interdependence: A protected transition record and the next transition's validation cooperate.¶
Technical predicate: Composed error versus decision margin; permit, request higher precision, or block.¶
Actual effect governed: Release of a numerical decision to a consequential consumer.¶
Claimed functional interdependence: Model/precision state, error composition and the consequence's margin must be evaluated together.¶
Technical predicate: Patch state, fresh evidence, reference values and policy match per consequence class.¶
Actual effect governed: Admission of workload effects after a security remediation.¶
Claimed functional interdependence: Remediation controller, attestation evidence controller and class-admission matrix cooperate.¶
Technical predicate: Installed fast-path match/action effects are a subset of the permitted effect, with an authority epoch.¶
Actual effect governed: Hardware-accelerated processing of authorized flows.¶
Claimed functional interdependence: Slow-path witness, offload contract, effective hardware descriptor and barrier cooperate.¶
Technical predicate: Audience and route evaluated across concurrent and in-flight output.¶
Actual effect governed: Human-perceptible presentation of information.¶
Claimed functional interdependence: Presentation controller, route state and queue fencing cooperate.¶
Technical predicate: 'No event' admitted only with sensing opportunity and coverage.¶
Actual effect governed: Actions inferred from absence of an event.¶
Claimed functional interdependence: Sensor health, coverage and time-gap evidence feed the admission decision.¶
Technical predicate: Authority register independent of model context binds later calls.¶
Actual effect governed: Tool calls after context summarization or handoff.¶
Claimed functional interdependence: Register, dispatcher and credential isolation cooperate.¶
Technical predicate: Acceptance criteria and oracle isolated from the modifying agent; deployed bytes bound to approval.¶
Actual effect governed: Release of agent-produced software.¶
Claimed functional interdependence: Oracle isolation and artifact binding cooperate at the promotion gate.¶
Technical predicate: Canonical authority contract, feature/exception matrix and tool-location map.¶
Actual effect governed: Tool execution across model APIs and hosted or client-executed tools.¶
Claimed functional interdependence: Bridge, contract and end-target enforcement cooperate.¶
Technical predicate: Materialized composition validated against cross-branch constraints.¶
Actual effect governed: Commit of combined agent work.¶
Claimed functional interdependence: Composition snapshot and aggregate validator cooperate at the merge boundary.¶
Technical predicate: Route inventory and sink-rejection evidence.¶
Actual effect governed: Assurance that governance controls are enforceable at effect routes.¶
Claimed functional interdependence: Inventory, contract, probes and sink evidence cooperate.¶
Technical predicate: Mandatory row, column, masking and recipient obligations survive the actual route.¶
Actual effect governed: Disclosure of governed data through alternative engines or shares.¶
Claimed functional interdependence: Policy semantics per engine and result-release validation cooperate.¶
Technical predicate: Consumer effect authority bound to a source-commit certificate.¶
Actual effect governed: External operations derived from committed transactions.¶
Claimed functional interdependence: Source-transaction controller, certificate and consumer sink cooperate.¶
Technical predicate: Conservative aggregate exposure with reservation before admission.¶
Actual effect governed: Payment obligations from one or more agents.¶
Claimed functional interdependence: Mandate, ledger, reservation and capture reconciliation cooperate.¶
Technical predicate: Delegated allocations, lineage and recipient evidence for fork attribution and bounded loss.¶
Actual effect governed: Offline monetary transfers.¶
Claimed functional interdependence: Allocation continuity, transfer evidence and fork reconciliation cooperate.¶
Technical predicate: Attributable impact evidence bounds expansion of a limited real change.¶
Actual effect governed: Progressive deployment of RAN optimization.¶
Claimed functional interdependence: Canary change, impact attribution and expansion gate cooperate.¶
Technical predicate: Lending conserves capacity, preserves floors and is reclaimable by a deadline.¶
Actual effect governed: Temporary slice resource loans.¶
Claimed functional interdependence: Entitlement table, lease state, reclaim state and scheduler enforcement cooperate.¶
This section is non-normative. It explains engineering relevance only. This document is not an EU compliance standard, and nothing here states that the architecture satisfies any legal obligation. Obligations vary by actor, product, sector and application date; readers should consult the legal texts and competent authorities.¶
The Regulation addresses risk management, record-keeping, human oversight and robustness for certain AI systems. A mechanism that makes a high-consequence action conditional on current, evidence-backed authority at the point of effect, with records of what was authorized and what occurred, is of engineering relevance to these themes. It is not shown here to satisfy any article.¶
Entities in covered sectors manage cybersecurity risk and incident handling. Bounded blast radius, route coverage evidence and fail-closed holds relate to risk-management measures at an engineering level.¶
Financial entities manage ICT risk and third-party risk. Exposure ledgers, reconciliation of unresolved outcomes and control over agent-initiated operations relate to operational resilience themes (E14, E33, E34).¶
Products with digital elements require security by design and vulnerability handling. The remediation-attestation correspondence profile (E9) and artifact-promotion gate (E16) are relevant to engineering of update and release flows.¶
Jurisdiction-bound, region-bound authority and evidence held by European operators can be modelled as policy predicates at a sink. The architecture does not itself guarantee data residency or provider independence.¶
AI-native orchestration and network automation research may benefit from management-plane authorization at configuration-effect boundaries (E2, E37, E38). No claim is made about any research programme's work or standards contribution.¶
Causal impairment control (E37), slice lending with reclaim (E38) and cumulative connectivity limits (E7) address service continuity at the management and device-policy level.¶
Protected hand-off of setpoints, deployments and remote commands can be gated at the actuation or deployment boundary, with local interlocks retained and explicit time budgets.¶
Exposure control for agent payments (E33), beneficiary-net outcomes (E34) and offline post-compromise reconciliation (E35) are research-level extensions. Existing secure-element and scheme controls remain authoritative; the Eurosystem's own research should be consulted ([ECB-DE]).¶
Attestation-to-effect linkage (E9), trust-domain transitions (E6) and offload equivalence (E10) relate to confidential computing deployments and require the relevant hardware interfaces.¶
Autonomous change and distributed actuation can be controlled by effect-specific gates. The architecture does not claim universal safety or certification.¶
Radio scheduling, HARQ, protection relays and similar safety-control loops operate at time scales where per-event authorization is generally infeasible; they rely on pre-validated local state and on certified safety functions. The proposals in this document address the slower boundaries where AI-generated or autonomous decisions become configuration, authorization, settlement or deployment effects. Where both are present, validation occurs when the bounded local state is created, renewed, changed or revoked (Section 7).¶
Sangam Das, "EUROPE'S NEXT AI SECURITY CHALLENGE: WHO CONTROLS THE FINAL ACTION? From EU AI Act and NIS2 to 6G, the Digital Euro, Industrial AI and Critical Infrastructure", published 16 November 2026, DOI 10.5281/zenodo.24274259 ([ZENODO-EU]). It is a non-peer-reviewed public research publication by the same author. Its relevance to this document is that it discusses, from a European perspective, who controls the final action in AI and autonomous systems, which is the question this document turns into an architecture. The editor could not retrieve the record when preparing this text, so this description is limited to the title and author-supplied context. Neither this document nor that publication establishes regulatory compliance or technical verification.¶
This section is non-normative and AI-assisted. Each entry separates verified public context from proposed integration. Verified context is a coarse public description that readers must re-verify against current vendor documentation. Proposed integration points are the author's engineering hypotheses and are not adoption, endorsement or infringement statements. No vendor has been shown to implement the proposed combinations.¶
Public SDK with function tools, handoffs and guardrails.¶
Place the sink in the tool dispatcher: before a function tool with a consequential effect runs, the dispatcher re-derives the operation, checks the authority record held outside the prompt, and rejects calls whose target or scope differs. It matters because handoffs and summarization can otherwise lose prohibitions (E15, E20).¶
Agent code and model unchanged; real credentials held by the dispatcher or a proxy, not the model process. Adds one local check per consequential call; no latency measurement is claimed.¶
Public tool-use interfaces; MCP is an open protocol for tool and context servers.¶
Place enforcement in the MCP client/host dispatch and at the MCP server's end target: the host looks up the canonical authority contract for each tool call, the server's sink verifies a scoped capability against the actual operation and rejects widened scope. It matters for cross-protocol semantic loss (E20) and confused-deputy risk (T10).¶
MCP servers continue to expose tools; the capability is an additional parameter or header; downstream credentials stay server-side.¶
Public managed agent service with tool integration and managed identities.¶
Insert the sink in the connector that holds the downstream credential: the managed identity is used only by the sink after capability verification, so an agent cannot call the target directly. It matters for alternate-credential bypass (T11).¶
Existing RBAC and identity unchanged; the agent loses direct access to the target credential.¶
Public agent framework and agent-to-agent protocol.¶
In multi-agent runs, validate the composed work product at the shared commit step (E21), and preserve the authority contract across agent-to-agent hand-offs (E15, E20).¶
Framework unchanged; commit step gains a validator call.¶
Public managed agent runtime with gateway and policy features.¶
Where a gateway already enforces policy, add end-target sink checks for effect classes the gateway cannot see (provider-hosted or direct-credential paths), and gather rejection evidence for route coverage (E25).¶
Gateway unchanged; adds coverage evidence and an end-target check.¶
Public automation platform with rApps in the non-real-time domain.¶
Gate the point where an rApp's output becomes an applied configuration: reconstruct multi-cell and slice effect, validate, apply as a limited change, and expand only on attributable impact evidence (E2, E37). It matters because a locally successful change can harm neighbors.¶
rApps and SMO unchanged; adds an admission step; PHY/MAC loops untouched.¶
Public SMO and RAN automation offering.¶
Same placement as above for SMO policy and slice control; for slice lending, the slice controller enforces entitlement floors and reclaim deadlines (E38).¶
Existing controllers unchanged; lease state added.¶
O-RAN defines Non-RT/Near-RT RIC and open interfaces; 3GPP defines slicing.¶
Admission and effect verification at A1/O1-style management actions; no per-transport-block validation. Cumulative connectivity policy at the device (E7).¶
Standard interfaces unchanged; new checks sit in management or device policy.¶
Public agents operating on SAP business processes.¶
Insert a late-bound validator at the posting or release boundary: recompute tax, FX, pricing and allocation effects from live state and compare with the approval before the posting commits (E26). It matters because the approved document and the posted effect can differ.¶
ERP logic remains authoritative; the check calls the same computation in preview mode where available.¶
Public database with stored programs, triggers and agent integration.¶
Close the effect set of a stored program before a privileged mutation: enumerate triggers, definer-rights code and autonomous transactions reachable from the entry point and require authority for each material effect (E31).¶
Database security remains authoritative; enforcement in a supported hook or proxy.¶
Public agent platform on CRM data and flows.¶
Gate customer-facing 'completed' statements on authoritative fulfillment evidence (E27) and validate cohort fan-out (E29).¶
Flows unchanged; communication dispatch gains an evidence check.¶
Public workflow and ITOM features.¶
Revalidate the live dependency topology at the remediation commit (E28) and quarantine unresolved outcomes (E14).¶
CMDB unchanged; commit step adds a check.¶
Public governance and orchestration products.¶
Governance dashboards can export route inventory; add sink-rejection evidence per route to turn dashboards into coverage proof (E25).¶
Existing governance product unchanged.¶
Public catalog with row and column controls.¶
Test that obligations survive each query route, including shared or alternative engines, and validate the actual result released (E30); reconcile activation populations (E29).¶
Catalog policies unchanged; adds route-level equivalence tests.¶
Public GPU partitioning and DPU offload.¶
Single-writer output cutover on failover (E5); offload readback versus authorized flow (E10).¶
Drivers unchanged; requires device interfaces for installed-state readback, which must be verified.¶
Public TDX and attestation services.¶
Join attestation evidence with patch state and policy per consequence class before admitting effects (E9) and close prior-domain paths on reassignment (E6).¶
Attestation service unchanged.¶
Public confidential compute architecture with realm management.¶
Domain transition controller withholds successor authority until DMA and interrupt paths are closed (E6).¶
Depends on SMMU and RME interfaces.¶
Public SoCs and mobile platforms with sensor hubs and output routing.¶
Wake-time revalidation of escrowed intents (E13); audience-bound release across display and casting routes (E11, E12); connectivity trajectory limits (E7).¶
Requires OS vendor interfaces; no adoption is implied.¶
Public Eurosystem research on offline payments with secure hardware.¶
Research extension: delegated child allocations and post-compromise fork reconciliation (E35).¶
Existing secure-element controls remain authoritative.¶
Public agent-payment initiatives and payment networks.¶
Exposure ledger and reservation at authorization-to-capture (E33); beneficiary-net checks and reconciliation at cross-border flows (E34); retry quarantine after lost acknowledgments (E14).¶
Network controls remain authoritative; participants control only their own phases.¶
Public industrial edge platforms and automation suites.¶
Gate deployments and remote commands at the edge runtime and at the actuation boundary; retain local interlocks; use digital-twin results as evidence, not as authority.¶
Safety functions unchanged; validator unavailability must map to the system's safe state.¶
Foundational reference: PCT/IB2026/055615 [PCT-055615], filed 4 June 2026, published 16 July 2026 as WO 2026/150382 A2.¶
Later Indian provisional application: 202631136555 [IN-136555], dated 15 November 2026 (as reported by the applicant from the filing acknowledgment; the supplied PDF is not a certified registry record).¶
Public research cross-reference: https://doi.org/10.5281/zenodo.24274259 [ZENODO-EU]¶
Related public disclosures by the same author (informative): Technical Disclosure Commons records [TDC-12542], [TDC-12667], [TDC-12668], [TDC-12853], [TDC-12931], [TDC-12961]; Zenodo records [ZEN-21861642], [ZEN-22082995], [ZEN-23040181], [ZEN-23657985]; and the earlier Internet-Drafts cited in Section 1.3, including [I-D.das-execution-finality-protocol-layer]. Titles are as listed in the applicant's index and should be verified.¶
Attribution: the underlying technical invention is attributed to the human inventor, Sangam Kumar Das. AI-assisted drafting, summarization, indexing, translation or literature analysis does not establish independent inventorship.¶
Limits: citation of a patent document does not establish patent validity, priority, essentiality or licensing terms. This document contains no licensing commitment and makes no statement about royalties. Contributors are bound by BCP 78 and BCP 79 ([RFC5378], [RFC8179]). The author will make the IPR disclosure for this document at the appropriate time, through the official IETF process and in accordance with IETF guidelines; text in this document is not a substitute for that formal disclosure. An earlier IPR declaration relating to related DAS Internet-Drafts is listed at [IETF-IPR-7610]; whether it covers this document has not been established. Readers should consult the IETF IPR disclosure database for the current status of any declaration relating to this author's drafts.¶
Source integrity: the Indian filing text supplied to the editor contains 427 physically present numbered claims (1-273 and 276-429). Claims 274 and 275 are absent and claim 273 ends mid-sentence (at 'current Group Audience'). This document does not reconstruct the missing text. Verify against the official filed copy (Appendix C).¶
Non-endorsement: this is an individual technical proposal and not an endorsed IETF standard, EU regulatory approval, telecommunications standard, financial certification or implementation by a named company. Company and product references indicate potential technical integration points and are not evidence of adoption, sponsorship, copying, infringement or commercial affiliation. Because AI-generated material may contain errors, readers should verify all patent numbers, dates, claim quotations, product capabilities, standards references and technical assertions against primary sources.¶
This appendix is non-normative. It uses the claim mapping of the Indian provisional's consolidated claim section as supplied to the editor. For each profile it gives the technical objective, the distinguishing enforcement predicate (taken from the 'wherein' clause of the independent claim), the effectuation boundary, primary failure modes, the claim group, interoperability relevance and a classification. The classification (N narrows an earlier principle; C combines earlier mechanisms; X appears to introduce additional subject matter requiring independent examination) is an editorial reading of the applicant's comparison and is not verified against the mothership text. No independent novelty is asserted merely because a profile contains newly named components. Claim numbers must be verified against the official filed copy: claims 274 and 275 are absent and claim 273 is truncated in the supplied PDF.¶
| ID | Claims | Subject | Class |
|---|---|---|---|
| E1 | 1-11 | Protected cryptographic key-lifecycle finality | HSM and KMS command admission; PKCS#11/KMIP-style management interfaces; key-state fencing. |
| E2 | 12-22 | AI-native radio-network state authorization | O-RAN SMO/Non-RT RIC policy admission; O1/A1-style configuration paths; vendor automation platforms. |
| E3 | 23-33 | Declarative reconciliation and non-resurrection | Kubernetes controllers, GitOps revision pinning, admission control, lease/epoch fencing. |
| E4 | 34-44 | Causal downstream-effect closure | Event meshes, workflow engines, message brokers, saga orchestrators; ERP and CRM event consumers. |
| E5 | 45-55 | Accelerator execution-authority cutover | Distributed training/inference fleets; GPU partitioning; checkpoint promotion; fencing-token style ownership epochs. |
| E6 | 56-66 | Trust-domain transition and prior-access closure | Arm CCA/RME with SMMU, Intel TDX, AMD SEV-SNP, SoC domain controllers. |
| E7 | 67-77 | Cumulative connectivity trajectory authorization | Device connectivity managers; multi-access steering policies; mobility hysteresis and dwell-time management. |
| E8 | 78-88 | Consequence-specific numerical-error reliability | Mixed-precision and quantized inference stacks; safety-relevant decision gates; accelerator runtimes. |
| E9 | 89-99 | Remediation and current attestation correspondence | RATS verifier/relying-party roles (RFC 9334); confidential-computing attestation services. |
| E10 | 100-110 | Software-to-hardware offload effect equivalence | SmartNIC/DPU/IPU offload; P4 pipelines; flow-rule tables; programmable switch fast paths. |
| E11 | 111-121 | Multimodal audience and presentation exposure | Device output routers; casting and screen-sharing stacks; notification delivery. |
| E12A | 122-132 | Negative-observation and sensor-coverage evidence | Home/building automation hubs; physical-security and ambient-sensing platforms. |
| E12 | 133-143 | In-flight media release and route changes | Streaming and notification queues; casting revocation; secure audio/video output. |
| E13 | 144-154 | Low-power intent escrow and wake-epoch continuity | Sensor hubs and always-on coprocessors; mobile and embedded SoC power domains. |
| E14 | 155-165 | Uncertain execution outcomes and retry safety | Idempotency keys, durable workflows, outbox reconciliation, payment status inquiry. |
| E15 | 166-176 | Authority continuity through agent context compaction | Agent SDKs and runtimes with tool dispatch, memory and handoff features. |
| E16 | 177-187 | Independent verification of autonomous software repairs | CI/CD pipelines, artifact signing, provenance attestations, protected branches. |
| E17 | 188-198 | Human delegation and minimal-disclosure authority | Service-desk, approval and case-management workflows with human escalation. |
| E18 | 199-209 | Spoken intent and interruption-epoch validation | Voice assistants, realtime voice agents, in-car and smart-home voice control. |
| E19 | 210-220 | Cross-source factual evidence consistency | Retrieval-augmented systems; enterprise search; temporal and snapshot data stores. |
| E20 | 221-231 | Cross-protocol tool-semantic authority | MCP, vendor agent APIs, agent-to-agent protocols, tool gateways. |
| E21 | 232-242 | Aggregate multi-agent work-product finality | Multi-agent orchestration; collaborative revision and merge systems. |
| E22 | 243-253 | Document-field spatial provenance | Document AI, invoice processing, accounts-payable automation. |
| E23 | 254-264 | Visual-target and persistent referent continuity | Multimodal assistants, robotics, spatial computing, GUI-automation agents. |
| E24 | 265-273* | Cross-channel communication-origin authority | Enterprise messaging and collaboration channels, business messaging APIs. |
| E25 | 276-286 | Governance enforcement-coverage reconciliation | Agent-governance platforms, API/credential inventories, effect-route test harnesses. |
| E26 | 287-297 | Late-bound ERP and accounting effects | ERP and accounting systems, procure-to-pay, financial close. |
| E27 | 298-308 | Customer commitments versus actual fulfillment | Customer-service agents, case-management and order-fulfillment systems. |
| E28 | 309-319 | Live-topology IT remediation | IT operations management, CMDB, cloud change and remediation tooling. |
| E29 | 320-330 | Analytical cohort identity and activation | Data platforms with activation to marketing, CRM and account actions. |
| E30 | 331-341 | Cross-engine data-policy semantics | Lakehouse and warehouse catalogs, data-sharing and multi-engine query paths. |
| E31 | 342-352 | Transitive database stored-program effects | Relational databases with stored programs, triggers and external-call features. |
| E32 | 353-363 | Source-commit-certified downstream effects | Transactional outbox, change-data-capture, brokers and event buses. |
| E33 | 364-374 | Agentic payment exposure reservation | Tokenized agent payments, issuer and gateway limit controls, PSP APIs. |
| E34 | 375-385 | Cross-border FX and beneficiary outcomes | Cross-border payment messaging and tracking; remittance networks. |
| E35 | 386-396 | Offline digital cash post-compromise fork evidence | Offline digital-currency research and wallets with secure elements. |
| E36 | 397-407 | Edge subrequest and outbound-effect authority | Edge and serverless runtimes, queues, service bindings, egress gateways. |
| E37 | 408-418 | Causal radio-service impairment control | RAN automation and rollout management, canary and closed-loop optimization. |
| E38 | 419-429 | 6G network-slice capacity lending and reclamation | Network-slice orchestration and RAN slice controllers. |
Gate key creation, activation, rotation, derivation, export, use and destruction on the effective key/partition transition, not on the requesting credential.¶
Authorization of the originating request, or possession of administrative credentials, does not authorize a materially different or no-longer-authorized lifecycle transition.¶
HSM/KMS command-admission point; key-state transition inside the cryptographic module.¶
Residual key-use authority after rotation or destruction; derived keys outliving a revoked parent; replayed administrative command; split-brain partition state.¶
Claims 1-11 of the consolidated claim section.¶
HSM and KMS command admission; PKCS#11/KMIP-style management interfaces; key-state fencing.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Reconstruct the aggregate multi-cell, neighbor and slice state an AI-generated radio configuration would produce and authorize that effective state.¶
Acceptance of a controller instruction does not authorize a different effective network configuration or an aggregate resource effect exceeding the validated scope.¶
Configuration-effect point on the management plane (SMO/Non-RT RIC to network element); not PHY/MAC loops.¶
Neighbor-cell side effects; slice footprint underestimated; stale policy epoch; optimizer request differs from installed parameters.¶
Claims 12-22 of the consolidated claim section.¶
O-RAN SMO/Non-RT RIC policy admission; O1/A1-style configuration paths; vendor automation platforms.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Bind reconciliation authority to a current generation so a controller cannot reinstall a revoked configuration.¶
Persistence of a desired-state declaration, repetition of a reconciliation loop, or restoration of a controller does not recreate authority after the generation authority is retired.¶
Reconciliation-admission step before the control-plane change is committed.¶
Obsolete desired-state snapshot reapplied after rollback; restored controller using retired generation; GitOps revision drift.¶
Claims 23-33 of the consolidated claim section.¶
Kubernetes controllers, GitOps revision pinning, admission control, lease/epoch fencing.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Materialize the causal tree of child effects (workflows, queues, APIs, stored programs) and require current operation-specific authority at each externally effective child.¶
Authorization of the Root Candidate Act does not confer unrestricted authority on downstream operations generated by causal expansion.¶
Each downstream sink that can make a derived effect externally effective.¶
Unmodelled child branch; confused-deputy child using broader credentials; child outliving revocation of root authority.¶
Claims 34-44 of the consolidated claim section.¶
Event meshes, workflow engines, message brokers, saga orchestrators; ERP and CRM event consumers.¶
Editorial class C: this profile combines earlier mechanisms into an additional implementation arrangement.¶
Separate checkpoint possession from authority to publish results during migration or failover; single effect-commit owner via source fence, destination admission and effect watermark.¶
Possession or restoration of computational state does not confer effect-commit authority; source and destination are prevented from committing conflicting effects.¶
Protected output publish or model-state commit interface (not every tensor operation).¶
Zombie source worker publishing after failover; destination double-commit; stale checkpoint promoted; watermark regression.¶
Claims 45-55 of the consolidated claim section.¶
Distributed training/inference fleets; GPU partitioning; checkpoint promotion; fencing-token style ownership epochs.¶
Editorial class C: this profile combines earlier mechanisms into an additional implementation arrangement.¶
Prove old-domain DMA, interrupt, debug and queued-I/O paths are closed before successor authority for a resource is granted.¶
A declared ownership change, permission update or domain reassignment does not establish completion while a prior-authority path can still produce a prohibited effect.¶
Resource-assignment controller and protected DMA-mapping/transition interlock.¶
Residual IOMMU/SMMU translation; in-flight DMA; interrupt remap not retargeted; debug path left enabled; queued I/O completing in new domain.¶
Claims 56-66 of the consolidated claim section.¶
Arm CCA/RME with SMMU, Intel TDX, AMD SEV-SNP, SoC domain controllers.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Evaluate the history of access transitions (Wi-Fi, cellular, satellite) against a cumulative temporal envelope rather than each switch alone.¶
Independent authorization of individual connectivity operations does not authorize a sequence whose cumulative governed effects exceed the temporal envelope.¶
Terminal connectivity policy controller, outside high-frequency radio loops.¶
Oscillating switches each individually permitted; history record tampered or reset; service-continuity obligation violated by sequence.¶
Claims 67-77 of the consolidated claim section.¶
Device connectivity managers; multi-access steering policies; mobility hysteresis and dwell-time management.¶
Editorial class X: this profile appears to introduce additional technical subject matter requiring independent prior-art and priority-support examination.¶
Compose quantization, rounding and accumulation error with the decision margin of a specific consequential decision; permit, request higher precision, or block.¶
Approval of a model, processor, precision or configuration does not authorize a consequential operation when the effective numerical result fails, or cannot be shown to satisfy, its consequence-specific conditions.¶
Release of a numerical decision to a consequential downstream consumer.¶
Error bound underestimated; precision downgrade between approval and use; margin computed on a different model build.¶
Claims 78-88 of the consolidated claim section.¶
Mixed-precision and quantized inference stacks; safety-relevant decision gates; accelerator runtimes.¶
Editorial class X: this profile appears to introduce additional technical subject matter requiring independent prior-art and priority-support examination.¶
Match patch state, fresh attestation evidence, reference values and relying-party policy per consequence class, while allowing defined safe continuity.¶
Installation or reported activation of a remediation does not establish attestation correspondence; earlier attestation acceptance does not authorize an effect that requires a newer condition.¶
Attested workload admission and effect-specific trust-state fence.¶
Stale appraisal after patch; reference values lag the patch; grace-period abuse; policy epoch change not reflected in cached verdict.¶
Claims 89-99 of the consolidated claim section.¶
RATS verifier/relying-party roles (RFC 9334); confidential-computing attestation services.¶
Editorial class C: this profile combines earlier mechanisms into an additional implementation arrangement.¶
Compare the effective installed hardware behavior (wildcards, priority, caches, state) with the effect permitted by the software decision; fence stale fast-path authority.¶
Installation of an accelerated rule, or software authorization of the corresponding operation, does not authorize repeated hardware effectuation when installed behavior exceeds the permitted scope or current authority.¶
Offload rule install/readback and protected fast-path enablement barrier.¶
Wildcard rule broader than the authorized flow; rule precedence changes effect; cached rule surviving revocation; readback differs from installed state.¶
Claims 100-110 of the consolidated claim section.¶
SmartNIC/DPU/IPU offload; P4 pipelines; flow-rule tables; programmable switch fast paths.¶
Editorial class X: this profile appears to introduce additional technical subject matter requiring independent prior-art and priority-support examination.¶
Evaluate concurrent audience, modality and residual output routes before information becomes perceptible.¶
Authorization to generate, access, transmit or present information through one route does not authorize exposure through another route with different audience, modality, privacy or residual-output conditions.¶
Output routing and endpoint release (display, speaker, casting).¶
Cast/mirroring route added; bystander-visible surface; residual buffer on the old route.¶
Claims 111-121 of the consolidated claim section.¶
Device output routers; casting and screen-sharing stacks; notification delivery.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Admit an absence-of-event inference only if sensing opportunity, health, coverage, buffering and time-gap conditions were met.¶
Absence of a received event is not sufficient authorization when a mandatory opportunity to observe, or evidence about that opportunity, has not been established.¶
Protected evidence qualification before physical, access or notification actions.¶
Sensor offline treated as 'no event'; coverage hole; buffered events delayed past decision; single-source completeness assumption.¶
Claims 122-132 of the consolidated claim section.¶
Home/building automation hubs; physical-security and ambient-sensing platforms.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Re-evaluate or fence already queued or streaming segments when audience or route changes.¶
Authorization of an initial presentation segment does not authorize continued release of unreleased information after a material change in audience authorization or route.¶
Queue/output sink with epoch-bound release.¶
Buffered audio or video reaching a newly unauthorized audience; notification queue flushed after privacy-state change.¶
Claims 133-143 of the consolidated claim section.¶
Streaming and notification queues; casting revocation; secure audio/video output.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Hold pending intent prepared by an always-on processor and revalidate it against current identity, policy and host state at wake.¶
Permission for low-power observation or computation does not confer authority to execute an escrowed operation after a power, boot, session or domain change.¶
Wake boundary and protected intent queue.¶
Stale escrowed intent executed after user change or policy revocation; wake-epoch not advanced; escrow tampering.¶
Claims 144-154 of the consolidated claim section.¶
Sensor hubs and always-on coprocessors; mobile and embedded SoC power domains.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Separate logical-effect identity from attempt identity; quarantine unresolved completion; reconcile against authoritative endpoint evidence before retry or compensation.¶
Loss of an acknowledgment, timeout, worker failure or originator restart does not establish that an earlier operation failed or authorize an unrestricted duplicate attempt.¶
Protected retry controller and effect ledger at the consequential endpoint interface.¶
Blind retry duplicating a committed effect; compensation under stale authority; conflicting replacement operation admitted during quarantine.¶
Claims 155-165 of the consolidated claim section.¶
Idempotency keys, durable workflows, outbox reconciliation, payment status inquiry.¶
Editorial class X: this profile appears to introduce additional technical subject matter requiring independent prior-art and priority-support examination.¶
Keep an authority record independent of the model-visible context so summarization or handoff cannot drop a prohibition.¶
Omission, modification or reinterpretation of authority information during model-context transformation does not remove an operative prohibition or expand authority.¶
Tool dispatcher or credential-isolated tool proxy.¶
Summary omits a 'do not' constraint; handoff widens scope; memory replacement rewrites the task; prompt-injected authority claim.¶
Claims 166-176 of the consolidated claim section.¶
Agent SDKs and runtimes with tool dispatch, memory and handoff features.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Isolate acceptance criteria and verification oracle from the modifying agent; bind deployed bytes to the approved, verified artifact.¶
An agent's ability to modify code, generate tests or report success does not confer authority to redefine acceptance criteria or release an artifact lacking independent verification evidence.¶
CI/CD artifact-promotion and deployment gate.¶
Agent edits its own tests; artifact swapped after verification; oracle shares agent credentials.¶
Claims 177-187 of the consolidated claim section.¶
CI/CD pipelines, artifact signing, provenance attestations, protected branches.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Send a bounded task capsule to a human and independently validate the returned work and continuing authority before any external effect.¶
Transfer of task information to a human does not transfer the agent's full context, stored credentials, general privileges or unrestricted authority.¶
Human-in-the-loop handoff gateway and action gateway.¶
Over-disclosure to the human; returned artifact treated as authority; stale authority on re-entry.¶
Claims 188-198 of the consolidated claim section.¶
Service-desk, approval and case-management workflows with human escalation.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Bind the final operative spoken intent, after corrections and barge-in, to the actual tool operation through a turn/interruption epoch.¶
A provisional transcript, incomplete instruction, outdated interpretation or superseded proposal does not confer authority for the corresponding effect.¶
Speech-to-tool commit boundary.¶
Early partial transcript executed; interrupted command still dispatched; ASR correction not reflected.¶
Claims 199-209 of the consolidated claim section.¶
Voice assistants, realtime voice agents, in-car and smart-home voice control.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Identify material factual dependencies and reconcile source validity and snapshot coherence at effectuation.¶
Authenticity of individual retrieved records, or production of a model synthesis, does not establish a coherent and current factual basis for the actual operation.¶
Evidence-sensitive tool execution gate.¶
Mutually inconsistent authentic records; stale cache; two 'independent' sources sharing one origin.¶
Claims 210-220 of the consolidated claim section.¶
Retrieval-augmented systems; enterprise search; temporal and snapshot data stores.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Preserve instruction authority and tool restrictions across model-API conversion, MCP adapters and provider-hosted versus client-executed tools.¶
Syntactic acceptance of an adapted request does not establish preservation of instruction authority, confidentiality or tool restrictions across protocols.¶
Protocol bridge plus independent tool/end-target control.¶
Instruction hierarchy dropped in translation; hosted tool bypasses gateway; unsupported feature silently ignored.¶
Claims 221-231 of the consolidated claim section.¶
MCP, vendor agent APIs, agent-to-agent protocols, tool gateways.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Materialize the composition of agent work products and authorize the accepted aggregate or a validated subset.¶
Individual authorization or local validation of work products does not authorize the consequential state produced by their combination.¶
Shared merge/commit boundary.¶
Cross-branch read/write conflict; cumulative limit exceeded by sum; negative constraint violated only after merge.¶
Claims 232-242 of the consolidated claim section.¶
Multi-agent orchestration; collaborative revision and merge systems.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Bind an extracted value to its source row/cell, page geometry, payee and business-effect identity before payment or record change.¶
Recognition of correct strings or field values does not authorize an operation when values are not sufficiently associated with the corresponding labels, records or relationships.¶
Extraction-to-ERP/payment finality point.¶
Amount bound to wrong invoice row; payee/IBAN mismatch; table geometry shift; duplicate-page confusion.¶
Claims 243-253 of the consolidated claim section.¶
Document AI, invoice processing, accounts-payable automation.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Verify that the entity currently targeted is the one originally referred to ('this one') before physical, financial or access effects.¶
An earlier reference to a screen position, region, label, description or coordinate does not authorize execution against a different entity that later occupies it.¶
Perception-to-action target verification.¶
UI re-layout; moving or occluded object; coordinate reuse; ambiguous referent.¶
Claims 254-264 of the consolidated claim section.¶
Multimodal assistants, robotics, spatial computing, GUI-automation agents.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Separate message transmission from authenticated original-speaker intent and check the current audience before sending protected content. (*Claim 273 truncated in the supplied file; claims 274-275 absent.)¶
Receipt, quotation, forwarding or summarization of a message does not confer an alleged earlier sender's authority; earlier authorization to disclose to a group does not authorize disclosure to a changed audience.¶
Controlled send/recipient-release point.¶
Quoted instruction treated as command; account reassignment; group membership changed after approval.¶
Claims 265-273* of the consolidated claim section. The supplied file has nine physically present claims (265-273); claim 273 ends mid-sentence.¶
Enterprise messaging and collaboration channels, business messaging APIs.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Verify, route by route, that required governance controls are enforceable at actual effect paths; do not equate logging with prevention.¶
Registration of an agent, a nominal policy, or compliant sampled traces do not establish that the required control is enforceable across the covered execution routes.¶
Assurance controller with proof of sink rejection per discovered route.¶
Undiscovered route; credential bypass; dashboard green while a path is uncontrolled; coverage probe unsafe to run.¶
Claims 276-286 of the consolidated claim section.¶
Agent-governance platforms, API/credential inventories, effect-route test harnesses.¶
Editorial class X: this profile appears to introduce additional technical subject matter requiring independent prior-art and priority-support examination.¶
Recompute effective posting, tax, FX and approval consequences against live state immediately before posting.¶
Authorization of an initial instruction, document or simulated transaction does not authorize a later commitment with a materially different financial or operational consequence.¶
ERP posting/release boundary.¶
Tax/FX/pricing change between approval and posting; allocation rule changed; simulated result differs from actual.¶
Claims 287-297 of the consolidated claim section.¶
ERP and accounting systems, procure-to-pay, financial close.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Gate outward statements ('refunded', 'shipped') on independently qualified fulfillment evidence, separately from initiating the internal request.¶
Authorization or submission of the underlying operation does not authorize a statement that it is completed or settled without supporting evidence; approval of the statement does not authorize the operation.¶
Customer-communication dispatch with authoritative fulfillment check.¶
'Done' claimed on request acceptance; refund pending but announced; stale case status.¶
Claims 298-308 of the consolidated claim section.¶
Customer-service agents, case-management and order-fulfillment systems.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Reconstruct the live dependency topology at the remediation commit; stale remediation generations cannot resume.¶
A previously valid change request does not authorize the same nominal command when the current target, live dependencies or material impact differs from the approved scope.¶
Protected change/repair controller at the commit point.¶
CMDB drift; new dependency edge; criticality changed; resumed stale plan.¶
Claims 309-319 of the consolidated claim section.¶
IT operations management, CMDB, cloud change and remediation tooling.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Reconcile the actual downstream target population, suppression and consent per target before fan-out.¶
Permission to run an analytical query or obtain a policy-filtered population does not authorize an operation against an expanded, duplicated, unsubscribed, suppressed or ineligible target population.¶
Analytics-to-activation fan-out boundary.¶
Join or identity-resolution expansion; suppressed identity re-added; duplicate recipient; consent withdrawn after query.¶
Claims 320-330 of the consolidated claim section.¶
Data platforms with activation to marketing, CRM and account actions.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Test that mandatory row, column, masking, aggregation and recipient obligations survive the actual query route and result release.¶
Syntactic compatibility, execution by an alternative engine, or storage-level credentials do not establish preservation of governance semantics.¶
Query route and effective result-release validation.¶
Zero-copy share bypassing masking; engine-specific semantic difference; storage-layer credential used directly.¶
Claims 331-341 of the consolidated claim section.¶
Lakehouse and warehouse catalogs, data-sharing and multi-engine query paths.¶
Editorial class X: this profile appears to introduce additional technical subject matter requiring independent prior-art and priority-support examination.¶
Reconstruct the closure of material effects reachable through triggers, definer-rights code, autonomous transactions and external functions.¶
Permission to invoke an entry-point operation or stored program does not confer authority for every material effect transitively reachable through it.¶
Supported database/transaction enforcement point before privileged mutation.¶
Trigger chain; definer-rights escalation; autonomous transaction side effect; external function call.¶
Claims 342-352 of the consolidated claim section.¶
Relational databases with stored programs, triggers and external-call features.¶
Editorial class C: this profile combines earlier mechanisms into an additional implementation arrangement.¶
Bind downstream release to an authoritative source commit and current policy/effect state; reconcile after uncertain acknowledgments.¶
Commitment of the source transaction, or presence of a derived event, does not authorize an external operation produced by a downstream consumer or transformed event.¶
Transactional outbox/CDC and consumer effect boundary.¶
Uncommitted intent consumed; stale outbox row; unrelated transaction event; duplicate delivery.¶
Claims 353-363 of the consolidated claim section.¶
Transactional outbox, change-data-capture, brokers and event buses.¶
Editorial class C: this profile combines earlier mechanisms into an additional implementation arrangement.¶
Conserve aggregate reserved, authorized and captured exposure against the mandate from authorization through capture.¶
A payment request cannot consume authority already reserved or committed for another incompatible obligation; unresolved exposure is classified conservatively.¶
Authorization-to-capture and exposure-reservation boundary.¶
Concurrent agents exceeding mandate; double-counted representations of one obligation; unresolved exposure treated as released.¶
Claims 364-374 of the consolidated claim section.¶
Tokenized agent payments, issuer and gateway limit controls, PSP APIs.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Distinguish instruction acceptance, settlement stages and independently evidenced beneficiary-net outcome.¶
Authorization of a nominal originating amount does not authorize an intermediary transformation or final beneficiary-side consequence violating the mandate; a message acknowledgment does not establish beneficiary credit.¶
Beneficiary-outcome and exception/reconciliation boundary.¶
FX or fee drift; intermediary deductions; ack mistaken for credit; routing change.¶
Claims 375-385 of the consolidated claim section.¶
Cross-border payment messaging and tracking; remittance networks.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Preserve delegated spending-authority continuity and reconcile conflicting transfer evidence after protection assumptions fail.¶
Prevention of unauthorized transfers under intact hardware does not establish that cloning remains impossible after compromise; detection of conflicting evidence does not establish the full amount of undiscovered conflicting transfers.¶
Offline wallet effectuation point plus later reconciliation.¶
Secure-element compromise; cloned state; incomplete fork discovery; privacy loss in reconciliation.¶
Claims 386-396 of the consolidated claim section.¶
Offline digital-currency research and wallets with secure elements.¶
Editorial class X: this profile appears to introduce additional technical subject matter requiring independent prior-art and priority-support examination.¶
Preserve parent source/purpose restrictions across subrequests, queues, retries and background workers; enforce at outbound effect paths.¶
Authentication or authorization of an originating edge request does not authorize each transitive, deferred, cached or independently executed effect the workload generates.¶
Edge runtime egress and credential-isolated service connector.¶
Background worker using broader credentials; queued task outliving revocation; cached result reused.¶
Claims 397-407 of the consolidated claim section.¶
Edge and serverless runtimes, queues, service bindings, egress gateways.¶
Editorial class N: this profile narrows an earlier architectural principle.¶
Correlate attributable impact evidence of a limited real RAN change and expand only within bounded impairment.¶
An individually authorized radio-control action is not automatically granted unrestricted continued or expanded deployment because its parameters were valid or telemetry has not yet shown degradation.¶
Non-real-time RAN rollout gate with bounded canary feedback.¶
Local metrics miss neighbor-cell harm; confounded attribution; premature expansion.¶
Claims 408-418 of the consolidated claim section.¶
RAN automation and rollout management, canary and closed-loop optimization.¶
Editorial class C: this profile combines earlier mechanisms into an additional implementation arrangement.¶
Account for loaned resources, continuity and enforceable reclaim before activation and further borrowing.¶
Temporary non-utilization of a first slice's resource does not establish that it may be lent unless entitlement, conservation and deadline-bounded reclamation remain satisfiable.¶
Slice orchestration and capacity-admission controls; protected scheduler lease state.¶
Unreclaimable loan after mobility or reroute; entitlement floor violated; capacity created by lending.¶
Claims 419-429 of the consolidated claim section.¶
Network-slice orchestration and RAN slice controllers.¶
Editorial class C: this profile combines earlier mechanisms into an additional implementation arrangement.¶
This appendix is non-normative. It lists the inventor's terminology and conventional industry equivalents, keyword clusters mapped to sections and embodiments, substantive search phrases and Boolean strings. The index does not imply that any search result anticipates, or is infringed by, any patent claim, and index terms are not claims or verified classification codes.¶
| Inventor terminology | Industry search terminology | Embodiments |
|---|---|---|
| Candidate Act | Proposed operation; deferred operation; staged transaction; pending request | 1-38 |
| Non-Effective State | Pending non-committed state; staged/uncommitted state | 1-38 |
| Actual Pending Operation | Reconstructed effective operation; resolved operation; effective-operation derivation | 1-38 |
| Operation Reconstruction | Effective-operation reconstruction; action canonicalization; effect derivation | 1,2,4,10,31 |
| Finality Sink | Protected effectuation boundary; execution enforcement point; policy enforcement point at the effect | 1-38 |
| Finality Gate | Commit gate; release gate; admission gate | 1-38 |
| Source-Authority Envelope | Originating authorization context; delegated authority envelope | 4,15,17,20,24,36 |
| Protected Enforcement Domain | Trusted enforcement environment; protected policy domain | 1-38 |
| Independent Validator | Runtime policy revalidation service; out-of-band authorization verifier | 1-38 |
| Scoped Execution Capability | Operation-bound execution authorization; non-bearer capability; sender-constrained, act-bound token | 1-38 |
| Consequence-Bound Authorization | Effect-aware authorization; consequence-specific authorization | 1-38 |
| Authority Continuity | Delegated authority preservation; authority propagation integrity | 4,15,17,20,36 |
| Protected Policy Epoch | Policy version fence; policy generation counter | 3,10,25 |
| Revocation Epoch | Revocation generation; revocation freshness marker | 1,3,9 |
| Evidence-Carrying Receipt | Authenticated effect evidence; signed execution receipt | 14,27,32,34 |
| Effect Confirmation Receipt | Authenticated destination effect evidence; destination-origin confirmation | 14,27,34 |
| Interim Effectuation Validation (IEV) | Receipt-gated continuation validation; staged-effect checkpoint | 14,33 |
| Ledger-Anchored Validation Receipt | Commit-log-anchored attestation record; transparency-log-anchored receipt | 32,35 |
| Effective Operation Descriptor / Effective Hardware Action Descriptor | Installed rule semantics; match-action effective behavior | 10 |
| Material Dependency Graph | Live dependency topology; causal effect graph; transitive effect closure | 4,28,31 |
| Protected State Commitment | Sealed state digest; authenticated state commitment | 5,9,14 |
| Authoritative Outcome Evidence | Destination-of-record confirmation; system-of-record outcome | 14,27,34 |
| Stale Authority | Expired/superseded authority; time-of-check/time-of-use gap | 3,5,13,28 |
| Transitive Effect | Indirect side effect; derived effect; second-order effect | 4,31,36 |
| Aggregate Consequence | Cross-agent combined-state effect; cumulative effect | 7,21,33 |
| Bounded Real Effect | Canary effect; limited trial effect; staged partial effectuation | 37,Receipt-gated IEV |
| Effectuation-Boundary Enforcement | Enforcement at the point of effect; last-mile enforcement | 1-38 |
| Control-to-Effect Route Coverage | Verified enforcement-path completeness; governance coverage | 25 |
| Protected Offload Equivalence | Software-to-hardware policy correspondence; fast-path equivalence | 10 |
| Aggregate Effect Validation | Cross-agent combined-state authorization; merge-time policy validation | 21 |
| Execution Finality | Protected consequential operation commitment; point-of-no-return control | 1-38 |
140 cluster keywords plus 40 additional keywords and 31 inventor terms listed in the synonym table.¶
Relevant sections and embodiments: Sections 4, 6 (REQ-02, REQ-14, REQ-15), 8.3; Embodiments E15-E21, E24, E36.¶
Relevant sections and embodiments: Sections 6 (REQ-04 to REQ-06, REQ-10), 8.4; E1, E5, E6, E9.¶
Relevant sections and embodiments: Section 8.7; E2, E7, E37, E38.¶
Relevant sections and embodiments: Sections 8.5, 8.6; E4, E22, E26, E28, E31, E32.¶
Relevant sections and embodiments: Appendix A (E29, E30).¶
Relevant sections and embodiments: Section 8.8; E14, E33, E34, E35.¶
Relevant sections and embodiments: Sections 8.4, 18, 19; E5, E6, E8, E10, E13.¶
72 phrases, each usable as an exact-phrase or proximity query.¶
This appendix is non-normative.¶
PCT/IB2026/055615, filed 4 June 2026, international publication WO 2026/150382 A2 on 16 July 2026. The inventor describes it as approximately 8,598 pages and more than 300 distinct embodiments; these figures are inventor-reported and unverified.¶
Indian provisional 202631136555, dated 15 November 2026; approximately 2,405 pages; 39 embodiments (1-38 plus 12A); consolidated claim section. Number and date are from the applicant's filing acknowledgment.¶
Publication on 16 July 2026 (not 16 June) does not by itself establish an earlier priority date for every feature of a later filing. Entitlement requires comparison with the as-filed earlier text.¶
The supplied merged PDF has 427 physically present numbered claims (1-273, 276-429); 274 and 275 absent; claim 273 ends mid-sentence. The text is neither reconstructed nor supplied here.¶
The mothership JSON (Technical_Disclosure_1.json) was not available to the editor; its identity with the as-filed text is uncertified.¶
| PCT application | Filed | WO publication | Published |
|---|---|---|---|
| PCT/IB2026/053385 | 7 Apr 2026 | WO 2026/115520 | 4 Jun 2026 |
| PCT/IB2026/054453 | 5 May 2026 | WO 2026/150381 | 16 Jul 2026 |
| PCT/IB2026/055615 | 4 Jun 2026 | WO 2026/150382 | 16 Jul 2026 |
| PCT/IB2026/055760 | 7 Jun 2026 | WO 2026/150383 | 16 Jul 2026 |
| PCT/IB2026/055870 | 10 Jun 2026 | WO 2026/150384 | 16 Jul 2026 |
| PCT/IB2026/056058 | 13 Jun 2026 | WO 2026/154461 | 23 Jul 2026 |
| PCT/IB2026/056809 | 1 Jul 2026 | WO 2026/172334 | 20 Aug 2026 |
| PCT/IB2026/056941 | 6 Jul 2026 | WO 2026/172336 | 20 Aug 2026 |
The list above is non-exhaustive and is drawn from the applicant's index; it is not a declaration that all listed applications share a valid priority claim. Verify each in WIPO PATENTSCOPE.¶
This appendix is non-normative and holds patentability-oriented considerations that are deliberately kept out of the architecture. It does not assert novelty or inventive step.¶
(1) Written description and effective date: where is each limitation supported in the as-filed June PCT or in the later filing. (2) Novelty: does one earlier source disclose the combination as properly construed. (3) Inventive step: which closest references, what objective technical problem, whether the skilled person would (not merely could) combine them, and with what reasonable expectation of success.¶
For each candidate distinction the examiner should ask what problem the combination solves that its constituents do not. Examples: for E10, why combine software authorization with installed-rule readback; for E32, why bind consumer authority to a source-commit certificate rather than to message authenticity; for E35, why add fork evidence to secure-element prevention.¶
The profiles that depend on device or database interfaces (E6, E10, E31) depend on supported readback or hook interfaces. Whether such interfaces exist for a given product is a factual question.¶
Each profile claims that its elements cooperate (see Section 17). An examiner should test whether the elements are a mere aggregation or produce a combined technical effect (for example a single effect-commit owner, or a bounded exposure).¶
Fencing tokens and lease epochs (E3, E5); transactional outbox and CDC (E32); P4/match-action verification (E10); 3GPP ATSSS and slicing (E7, E38); secure-element offline cash (E35); policy pushdown and row/column security (E30); agent guardrails and MCP gateways (E15, E20, E25).¶
Stronger detail alone does not establish a new priority date, independent novelty or inventive step.¶