P231 — AIEP — Abstraction Extraction 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 the abstraction extraction mechanism of the Phase-2 AIEP knowledge architecture, enabling the system to identify recurring patterns in the evidence corpus and world state and extract them as reusable abstracted inference rules.

Field of the Invention

[0002] The present invention relates to abstraction extraction systems and pattern generalisation architectures for evidence-bound artificial intelligence.

[0003] More particularly, the invention relates to a system that identifies evidence-grounded recurring causal patterns across multiple independent instances in the world state model and extracts them as abstracted, parameterised inference rules with provenance records linking each abstraction to its supporting instances.

Background

[0004] Evidence-bound reasoning systems that reason only from specific evidence instances cannot transfer learning across domains without re-deriving reasoning from first principles on each application. Abstraction — identifying general patterns from specific instances — enables efficient knowledge transfer while maintaining provenance binding.

[0005] Abstraction in evidence-bound architectures must preserve the proof chain from each abstracted rule back to the specific evidence instances that support it. Without this, abstractions become indistinguishable from ungrounded parametric generalisations.

Summary of the Invention

[0006] The invention provides an Abstraction Extraction Engine (AEE) that: monitors the CWSG (P200) for recurring causal patterns across multiple entity instances; identifies candidate abstractions using a coverage threshold (minimum number of supporting instances); generates abstracted rule candidates as parameterised inference rules; validates candidates through the Hypothesis Simulation Engine (P204) to confirm their predictive validity; and commits validated abstractions to the abstraction library with full provenance chains.

[0007] Abstracted rules are granted provisional status only and are subject to ongoing staleness evaluation — if the evidence instances supporting an abstraction are contradicted by new evidence, the abstraction is demoted pending revalidation.

ASCII Architecture

CWSG Pattern Monitoring (P200)
    |
    v  (recurring pattern candidates)
+------------------------------------------+
| Abstraction Extraction Engine (AEE)      |
|                                          |
|  Pattern clustering                     |
|  Coverage threshold check               |
|  Rule candidate generation              |
|  HSE validation (P204)                  |
|  Provenance chain construction          |
+-------------------+----------------------+
                    |
                    v
  Abstraction Library
  (parameterised rules + provenance chains)
           |
           v
  Knowledge Distillation Integration (P217)

Detailed Description

[0008] Pattern Monitoring. The AEE continuously monitors the CWSG mutation log for causal edges matching recurring structural patterns. A pattern is a causal structure (e.g., A[type=X] -> event[type=Y] -> B[state=Z]) that appears multiple times across different entity instances.

[0009] Coverage Threshold. A candidate abstraction is eligible for extraction only when the pattern appears in at least N distinct entity instances (default N=5), each supported by independent evidence artefacts. This prevents abstraction from single-instance observations.

[0010] Rule Candidate Generation. A candidate rule is produced by parameterising the recurring pattern: replacing specific entity identifiers with type constraints, and specific attribute values with predicate expressions. The result is a formal inference rule applicable to any entity satisfying the type constraints.

[0011] HSE Validation. Each rule candidate is submitted to the Hypothesis Simulation Engine (P204) as a hypothesis. The simulation tests whether applying the rule to new entity instances produces outcomes consistent with subsequent evidence artefacts. Rules failing the fitness threshold are rejected.

[0012] Provenance Chain. For each committed abstraction, the provenance chain records: the specific entity instance identifiers from which the abstraction was derived; the evidence artefact hashes supporting each instance; and the HSE simulation records validating the rule.

Technical Effect

[0013] The invention provides evidence-grounded abstraction extraction that converts recurring causal patterns into validated, provenance-preserving inference rules. By requiring a minimum instance coverage threshold across independent evidence artefacts, the engine prevents abstraction from anecdotal or single-instance observations. By validating rule candidates through hypothesis simulation before commitment, the engine ensures that extracted abstractions have predictive validity. By maintaining provenance chains linking each abstraction to its source instances and by automatically demoting abstractions when source evidence is contradicted, the engine preserves continuous epistemic integrity of the abstracted knowledge layer.

Claims

1. A computer-implemented method for abstraction extraction, the method comprising: (a) monitoring the CWSG and evidence corpus for recurring causal patterns appearing in at least N distinct entity instances each supported by independent evidence artefacts; (b) generating rule candidates by parameterising recurring patterns, replacing entity-specific identifiers with type constraints and attribute values with predicate expressions to produce formal inference rules; (c) submitting rule candidates to the Hypothesis Simulation Engine for fitness evaluation, with rules failing the fitness threshold rejected; (d) committing validated rules to the abstraction store with provenance chains recording the source entity instance identifiers, supporting evidence artefact hashes, and simulation validation records; and (e) evaluating committed abstractions on each evidence admission cycle and demoting abstractions whose source instances have been contradicted by new evidence.

2. The method of claim 1, wherein the minimum instance coverage threshold N is configurable via the active governance policy.

3. The method of claim 1, wherein rule candidates that fail hypothesis simulation fitness evaluation are stored as rejected candidates with failure reason, enabling retrospective analysis of abstraction attempts.

4. The method of claim 1, wherein abstraction demotion sets a demotion flag without deleting the rule record, preserving the abstraction history for audit.

5. The method of claim 1, wherein validated abstractions are forwarded to the Knowledge Distillation Engine as distillation candidates with their provenance chains.

6. An Abstraction Extraction Engine comprising: one or more processors; memory storing a pattern occurrence index, rule candidate store, and committed abstraction 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

An abstraction extraction engine for evidence-bound artificial intelligence identifies recurring causal patterns across a minimum coverage threshold of independent entity instances, generates parameterised inference rules, and validates them through hypothesis simulation before committing to the abstraction store with full provenance chains. Committed abstractions are continuously evaluated against new evidence, with contradicted abstractions demoted to preserve epistemic integrity of the knowledge base.