P206 — AIEP — Action Execution and Intervention Engine
Applicant: Neil Grassby Classification: Patent Application — Confidential Priority: Claims priority from GB2519711.2 filed 20 November 2025 Architecture Layer: AIEP AGI Cognition Layer — Phase 2
Framework Context
[0001] This specification operates within an AIEP environment as defined in GB2519711.2 and GB2519798.9. The present specification defines the governed action execution layer of the Phase-2 AIEP cognition architecture, enabling AI reasoning systems to take external interventions within a governance-constrained, evidence-bound execution framework.
Field of the Invention
[0002] The present invention relates to governed action execution engines for evidence-bound artificial intelligence systems.
[0003] More particularly, the invention relates to a system for planning, authorising, executing, and recording external interventions by an AI reasoning system, wherein all action proposals are evaluated against governance constraints before execution and all executed actions produce evidence artefacts that are admitted to the AIEP evidence ledger.
Background
[0004] AI systems that interact with external environments — sending messages, triggering processes, modifying system state — must do so within a governance framework that prevents unsafe, unauthorised, or irreversible actions. Without a formal action governance layer, AI systems may take interventions that violate policy constraints, lack audit trails, or cannot be reconstructed for compliance review.
[0005] Action governance in existing systems is typically implemented as post-hoc filtering (rejecting outputs after generation) rather than pre-execution policy evaluation. Post-hoc filtering cannot prevent partial execution of multi-step action sequences.
Summary of the Invention
[0006] The invention provides an Action Execution and Intervention Engine (AEIE) that receives action proposals from the reasoning layer, evaluates each proposal against governance policy (P215) before execution, executes approved actions through registered actuation adapters, and records execution receipts as AIEP evidence artefacts.
[0007] Action proposals that fail the governance gate are rejected without execution and produce a negative proof record. Action proposals that pass the governance gate execute through the actuation adapter and produce an execution receipt artefact containing: the action identifier; the actuation timestamp; the action parameters; and the SHA-256 hash of the action output.
[0008] The AEIE enforces a reversibility classification on all actions — classifying each as reversible, partially reversible, or irreversible — and applies elevated governance thresholds to irreversible actions.
ASCII Architecture
Reasoning Layer Action Proposal
|
v
+--------------------------------------+
| Action Governance Gate (P215) |
+------------------+-------------------+
|
approved / rejected
|
+--------+--------+
| |
v v
Execute via Adapter Negative Proof
| Record (P129)
v
+-------------------------------------+
| Execution Receipt Artefact |
| - action_id |
| - parameters_hash |
| - executed_at |
| - outcome_hash |
| - reversibility_class |
+-------------------------------------+
|
v
AIEP Evidence Ledger
(admitted as evidence artefact)
Definitions
[0009] ActionProposal: A structured record from the reasoning layer specifying an intended external intervention. Each ActionProposal contains: an action type; a target identifier; a parameter set; a reversibility pre-classification; and one or more justification evidence hashes.
[0010] ActuationAdapter: A registered module implementing the execution interface for a specific action type or target system. Adapters are registered in the ActuationRegistry and must implement a standard execution contract including pre-execution validation and execution receipt production.
[0011] ExecutionReceiptArtefact: A canonical AIEP evidence artefact produced by the AEIE after a successful action execution. Each receipt contains: action identifier; parameter hash; executed_at timestamp; outcome hash; reversibility class; and the governance approval record hash.
[0012] ReversibilityClass: A classification applied to each action as: REVERSIBLE (outcome can be fully undone), PARTIALLY_REVERSIBLE (elements can be undone but with residual effects), or IRREVERSIBLE (outcome cannot be undone). Governance thresholds differ by reversibility class.
[0013] NegativeProofRecord: An evidence artefact produced when an ActionProposal is rejected by the governance gate, recording the proposed action, the rejection reason, and the policy version in force at rejection time.
Detailed Description
[0014] ActionProposal Submission. The reasoning layer submits an ActionProposal to the AEIE. The AEIE applies a pre-screening check validating that: the action type is registered in the ActuationRegistry; the target identifier is well-formed; and the parameter set conforms to the action type schema. Proposals failing pre-screening are rejected without governance gate evaluation.
[0015] Reversibility Pre-Classification. Before governance evaluation, the AEIE applies reversibility pre-classification. The actuation adapter for the action type exposes a classify_reversibility(action_params) method that returns the ReversibilityClass for the proposed action. IRREVERSIBLE actions require a higher-authority approval group and a mandatory reflection interval before execution proceeds.
[0016] Governance Gate Evaluation. The ActionProposal, enriched with its ReversibilityClass, is submitted to the Safety Constraint and Governance Enforcement Engine (P215). P215 evaluates the proposal against the active governance policy. Possible outcomes: APPROVE — execution proceeds; REJECT — NegativeProofRecord is created; ESCALATE — proposal is elevated to a higher-authority authorisation tier.
[0017] Execution Protocol. On governance approval, the AEIE invokes the registered ActuationAdapter’s execute method with the action parameters. The adapter performs the external intervention and returns an execution outcome including the outcome status and an outcome hash (SHA-256 over the execution outputs). The AEIE wraps the outcome as an ExecutionReceiptArtefact.
[0018] Receipt Admission. The ExecutionReceiptArtefact is admitted to the AIEP evidence ledger as an evidence artefact of type execution_receipt. This admission ensures that all governed AI interventions in the external environment are recorded in the same provenance-bound substrate as all other evidence, enabling complete reasoning-action-outcome audit trails.
[0019] Reversibility Tracking. For REVERSIBLE and PARTIALLY_REVERSIBLE actions, the AEIE maintains a reversal index associating the ExecutionReceiptArtefact with the actuation adapter’s undo method. Reversal requests submitted to the AEIE include the execution receipt identifier; the AEIE invokes the adapter’s undo method and produces a reversal receipt artefact.
Technical Effect
[0020] The invention provides a governed, evidence-bound external intervention capability for AI reasoning systems. By evaluating all action proposals through the governance gate before execution, reversibility-classifying each action, and admitting execution receipts to the evidence ledger, the system ensures that every AI intervention in the external environment is: pre-evaluated against safety policy; classified by reversibility with elevated thresholds for irreversible actions; auditable through the evidence ledger; and traceable from reasoning justification to execution outcome.
Claims
-
A computer-implemented method for governed execution of external interventions by an AI reasoning system within an evidence-bound architecture, the method comprising:
(a) receiving an action proposal from a reasoning layer specifying an action type, target, parameters, and one or more justification evidence hashes;
(b) pre-classifying the reversibility of the proposed action as reversible, partially reversible, or irreversible using the registered actuation adapter’s reversibility classification method;
(c) submitting the action proposal with its reversibility classification to a governance enforcement engine and receiving an approve, reject, or escalate decision;
(d) on approval, executing the action through the registered actuation adapter and collecting an execution outcome comprising an outcome status and an outcome hash; and
(e) producing an execution receipt artefact from the execution outcome and admitting it to the AIEP evidence ledger as a provenance-bound record of the intervention.
-
The method of claim 1, wherein rejected action proposals produce a negative proof record admitted to the evidence ledger, recording the proposed action, rejection reason, and governance policy version in force at rejection.
-
The method of claim 1, wherein irreversible actions require a higher-authority approval group evaluation and a mandatory reflection interval before execution proceeds.
-
The method of claim 1, wherein the outcome hash in the execution receipt artefact is computed as a SHA-256 hash over the execution outputs, providing a tamper-evident record of the intervention result.
-
The method of claim 1, wherein reversible and partially reversible actions maintain a reversal index associating the execution receipt with the adapter reversal method, enabling governed reversal requests.
-
An action execution and intervention engine comprising one or more processors and an actuation adapter registry, configured to perform the method of any of claims 1 to 5.
-
A non-transitory computer-readable medium storing instructions which, when executed by one or more processors, perform the method of any of claims 1 to 5.
Abstract
A computer-implemented action execution and intervention engine is disclosed for evidence-bound AI reasoning architectures. External interventions proposed by the reasoning layer are pre-classified by reversibility and evaluated through the governance enforcement engine before execution. Irreversible actions require elevated authorisation. Approved actions execute through registered actuation adapters; execution receipt artefacts are admitted to the AIEP evidence ledger, creating a provenance-bound record of every governed AI intervention. Rejected proposals produce negative proof records. Reversal support is maintained for reversible actions through a reversal index. | - executed_at | | - outcome_hash | | - reversibility_class | +-------------------------------------+ | v AIEP Evidence Ledger
---
## Detailed Description
[0009] **Action Proposal Format.** An action proposal contains: `action_type` (classified action category e.g. `message.send`, `file.write`, `process.trigger`); `target_system` (actuation adapter identifier); `parameters` (action-specific parameter object); `proposed_by` (reasoning engine identifier); `reversibility_class` (`reversible | partial | irreversible`); and `justification_trace_id` (identifier of the reasoning trace that generated this proposal).
[0010] **Governance Gate.** The governance gate evaluates the action proposal against the active governance policy (P215) by: checking whether the action type is permitted for the requesting reasoning engine; verifying that the target system is within the authorised target set; applying elevated scrutiny rules to irreversible actions (requiring additional justification evidence); and checking whether the action falls within the parameter bounds defined by the policy.
[0011] **Actuation Adapter Registry.** Each external system that the AEIE may act upon is represented by a registered actuation adapter implementing: `execute(parameters) -> ExecutionResult`; `validate_parameters(parameters) -> ValidationResult`; and `classify_reversibility(parameters) -> ReversibilityClass`.
[0012] **Execution Receipt.** Upon successful execution, the AEIE produces an execution receipt artefact and admits it to the AIEP evidence ledger, creating a permanent record of the action that can be used in subsequent reasoning, audit, and compliance evaluation.
---
## Claims
1. A governed action execution engine for an evidence-bound reasoning architecture, wherein action proposals are evaluated against governance policy before execution.
2. The system of claim 1 wherein execution receipts are admitted as evidence artefacts to the AIEP evidence ledger.
3. The system of claim 1 wherein irreversible actions are subject to elevated governance thresholds.
4. The system of claim 1 wherein rejected action proposals produce negative proof records.
5. The system of claim 1 wherein actuation adapters are registered per target system and implement a canonical execution interface.