P240 — AIEP — Compute-Aware Epistemic Prioritisation Engine

Applicant: Neil Grassby Classification: Patent Application — Confidential Priority: Claims priority from GB2519711.2 filed 20 November 2025 Architecture Layer: AIEP Phase 2 Support Layer

Framework Context

[0001] This specification operates within an AIEP environment as defined in GB2519711.2 and GB2519798.9. The present specification defines a mechanism for adjusting epistemic prioritisation — the selection of which aspects of a question or world state to investigate — in response to available compute budget.

Field of the Invention

[0002] The present invention relates to compute-aware epistemic resource allocation for evidence-bound AI reasoning.

Background

[0003] Comprehensive evidence-bound reasoning may require extensive world state queries, hypothesis simulations, and multi-agent deliberation. In resource-constrained circumstances — high load, reduced budget, emergency conditions — the system must reason effectively within tighter compute bounds. An epistemic prioritisation mechanism selects the highest-value reasoning activities for the available budget.

Summary of the Invention

[0004] The invention provides a Compute-Aware Epistemic Prioritisation Engine (CAEPE) that, at the start of each reasoning session, receives the current compute budget from the Resource Allocation Engine (P213) and constructs a prioritised epistemic plan: a ranked list of investigation steps ordered by expected epistemic gain per compute unit.

[0005] Expected epistemic gain is estimated using knowledge utility scores (P232) for candidate evidence activation, historical information gain records from the Meta-Reasoning Quality Engine (P214), and goal relevance weights from the active goal set. Under full budget, all investigation steps proceed. Under constrained budget, lower-gain steps are deferred.

ASCII Architecture

Reasoning Session Start
         |
         v
  Current compute budget (P213)
         +
  Goal specification
         |
         v
+------------------------------------------+
| Compute-Aware Epistemic Prioritisation   |
|   Engine (CAEPE)                         |
|                                          |
|  Candidate investigation step catalog   |
|  Expected gain scoring (P232, P214)     |
|  Gain / compute_cost ranking            |
|  Epistemic plan construction            |
+-------------------+----------------------+
                    |
                    v
  Prioritised Epistemic Plan
  → Reasoning session executes in plan order
  → Low-gain steps deferred if budget depletes

Detailed Description

[0006] Investigation Step Catalog. At session start, the CAEPE enumerates candidate investigation steps: evidence artefact activations, CWSG subgraph traversals, hypothesis simulations, federated knowledge queries. Each step has an associated compute cost estimate.

[0007] Expected Gain Scoring. For each candidate step, expected epistemic gain is estimated by: looking up knowledge utility scores (P232) for relevant artefacts; consulting historical information gain records from P214 for similar steps; and applying a goal relevance multiplier based on the step’s alignment with the active goal set.

[0008] Gain/Cost Ranking. Steps are ranked by expected_gain / compute_cost. This ratio represents epistemic return per unit of compute.

[0009] Budget-Adaptive Execution. The reasoning session executes investigation steps in ranked order. As compute units are consumed, the remaining budget is tracked. If budget falls below the deferred threshold, remaining unstarted steps are deferred to subsequent sessions with a deferral record.

[0010] Deferral Record. Deferred investigation steps are stored with their priority score and session context, enabling the next session to continue efficiently from where the current session stopped.

Technical Effect

[0011] The invention provides evidence-grounded, budget-adaptive investigation scheduling for AI reasoning sessions that maximises epistemic return per unit of compute under constrained resource conditions. By ranking investigation steps by expected epistemic gain per compute unit and executing them in ranked order, the engine ensures that the most valuable knowledge acquisition work is completed first within each budget-bounded session. By recording deferred steps with priority scores and session context, the engine enables efficient continuation in subsequent sessions without redundant re-planning.

Claims

1. A computer-implemented method for compute-aware epistemic prioritisation, the method comprising: (a) generating a set of candidate investigation steps for the current reasoning session; (b) estimating expected epistemic gain for each step from knowledge utility scores and historical information gain records for similar steps; (c) applying a goal relevance multiplier based on each step’s alignment with the active GoalVector; (d) computing an expected-gain-to-compute-cost ratio for each step and ranking steps by this ratio; (e) executing investigation steps in ranked order, tracking compute budget consumption; and (f) on budget depletion, deferring remaining unstarted steps with their priority scores and session context for continuation in the next session.

2. The method of claim 1, wherein the deferred threshold is configurable, allowing the engine to stop accepting new steps before full budget exhaustion to preserve budget for the active session’s conclusion steps.

3. The method of claim 1, wherein deferred step records are stored in the Long-Term Reasoning Memory Engine (P208) with priority scores, enabling the next session to resume without re-ranking.

4. The method of claim 1, wherein historical information gain records used for gain estimation are derived from step execution evidence artefacts admitted in prior sessions.

5. The method of claim 1, wherein the goal relevance multiplier is updated dynamically if the active GoalVector changes during the session.

6. A Compute-Aware Epistemic Prioritisation Engine comprising: one or more processors; memory storing a step candidate queue, gain/cost ranking index, and deferred step store; wherein the processors are configured to execute the method of claim 1.

7. A non-transitory computer-readable medium storing instructions that, when executed by a processor, implement the method of claim 1.

Abstract

A compute-aware epistemic prioritisation engine for evidence-bound artificial intelligence ranks investigation steps in each reasoning session by expected epistemic gain per compute unit, executing them in ranked order within the available compute budget. Unstarted steps are deferred at budget depletion with priority scores and session context preserved for efficient resumption in subsequent sessions. Gain estimates are derived from knowledge utility scores and historical information gain records, with goal relevance multipliers applied from the active GoalVector.