P207 — AIEP — Outcome Learning and Model Update 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 outcome-driven model update mechanism of the Phase-2 AIEP cognition architecture, enabling the reasoning system to improve its predictive accuracy by comparing prior predictions to observed outcomes.
Field of the Invention
[0002] The present invention relates to outcome learning systems and predictive model improvement architectures for evidence-bound artificial intelligence.
[0003] More particularly, the invention relates to a system that compares predictions and hypothesis simulation outcomes against subsequently admitted evidence artefacts, evaluates prediction quality, and updates the reasoning model’s causal rule set to improve future prediction accuracy — all within an evidence-bound, auditable framework.
Background
[0004] Reasoning systems that predict future world states must have a feedback mechanism to identify when predictions diverge from observed reality. Without such a mechanism, the system cannot improve its predictive models and continues using causal rules that produce inaccurate forecasts.
[0005] Model update procedures in existing AI systems are typically implemented through gradient-based weight updates, which are not interpretable, not auditable, and cannot be traced to specific evidence artefacts. An evidence-bound architecture requires model updates that are traceable, reversible, and governed.
Summary of the Invention
[0006] The invention provides an Outcome Learning and Model Update Engine (OLMUE) that: retrieves outstanding predictions and hypothesis simulation outcomes from the reasoning history; compares them against newly admitted evidence artefacts; computes a prediction accuracy delta for each outstanding prediction; and generates causal rule update proposals based on systematic prediction failures.
[0007] Rule update proposals are evaluated by the governance gate (P215) before application. Applied rule updates are recorded in the causal rule audit log with the evidence artefact reference that triggered the update, enabling complete traceability of model evolution.
[0008] The OLMUE therefore provides a governed, evidence-bound learning loop that improves the reasoning system’s world model without introducing unaudited parameter changes.
ASCII Architecture
New Evidence Artefact Admitted
|
v
+---------------------------------------+
| OLMUE: Prediction Comparison Engine |
| |
| predicted_state <-> observed_state |
| accuracy_delta computed |
+-------------------+-------------------+
|
systematic failure detected?
|
v
+---------------------------------------+
| Rule Update Proposal Generator |
+-------------------+-------------------+
|
v
Governance Gate (P215)
|
v
+---------------------------------------+
| Causal Rule Update Applied |
| Causal Rule Audit Log updated |
+---------------------------------------+
Definitions
[0009] PredictionRecord: A stored record of a prediction or hypothesis simulation outcome produced by the reasoning system during a prior reasoning session. Each PredictionRecord contains: a prediction identifier; the predicted entity state or causal relationship; the world state snapshot hash at prediction time; and the evidence hashes supporting the prediction.
[0010] PredictionAccuracyDelta: A numerical measure of the divergence between a PredictionRecord’s predicted_state and the actual evidence-grounded state for the same entity at the equivalent temporal position. Positive deltas indicate over-prediction; negative deltas indicate under-prediction.
[0011] SystematicFailurePattern: A pattern identified by the OLMUE in which multiple PredictionRecords referencing the same causal rule or entity type consistently produce PredictionAccuracyDeltas beyond a configurable systematic-failure threshold.
[0012] RuleUpdateProposal: A structured proposal generated by the OLMUE to modify one or more causal rules in the shared world model in response to an identified SystematicFailurePattern. Each proposal contains: the target rule identifier; the proposed modification; the evidence artefact references that triggered the proposal; and a predicted accuracy improvement estimate.
[0013] CausalRuleAuditLog: An append-only log recording all applied rule updates, including: the RuleUpdateProposal hash; the governance approval record; the causal rule version before and after update; and the triggering evidence artefact hashes.
Detailed Description
[0014] Outstanding Prediction Retrieval. On each evidence admission cycle, the OLMUE retrieves PredictionRecords whose evaluation horizon has elapsed — that is, whose predicted temporal position is now covered by admitted evidence. For each such record, the OLMUE queries the ESR (P201) for the current entity state and computes the PredictionAccuracyDelta.
[0015] Accuracy Delta Computation. The PredictionAccuracyDelta for a categorical entity state is 0.0 if the predicted category matches the actual category, and 1.0 otherwise. For numerical entity attributes, the delta is the normalised absolute difference between predicted and actual values. Composite entity states produce a weighted average delta across attributes.
[0016] Systematic Failure Detection. The OLMUE aggregates PredictionAccuracyDeltas over a rolling window defined by the systematic_failure_window configuration parameter. When the median delta for PredictionRecords referencing the same causal rule exceeds the systematic_failure_threshold, the OLMUE classifies this as a SystematicFailurePattern and initiates a RuleUpdateProposal.
[0017] Rule Update Proposal Generation. The OLMUE generates a RuleUpdateProposal by: identifying the causal rule most associated with the systematic failure pattern; formulating a candidate modification that would have improved prediction accuracy on the failing records; and estimating the accuracy improvement the modification would produce on historic predictions. The proposal includes all triggering evidence artefact hashes as justification.
[0018] Governance Review. The RuleUpdateProposal is submitted to the Safety Constraint and Governance Enforcement Engine (P215). Governance evaluation verifies that: the proposed rule modification is consistent with constitutional constraints; the accuracy improvement estimate is positive; and the modification does not introduce conflicts with other active causal rules. Approved proposals are applied to the shared causal rule set.
[0019] Rule Audit Log. All applied rule updates are recorded in the CausalRuleAuditLog with: the RuleUpdateProposal hash; the governance approval record hash; the causal rule version identifiers before and after the update; and the ISO 8601 application timestamp. The audit log enables complete traceability of model evolution from prediction failures to governance-approved updates.
Technical Effect
[0020] The invention provides a governed, evidence-bound learning loop enabling the AIEP reasoning system to improve its predictive world model from observed prediction failures. By systematically collecting prediction accuracy deltas, identifying causal rule failure patterns, and routing all rule modifications through the governance gate with full audit logging, the system achieves learning behaviour that is interpretable, reversible, and traceable — attributes absent from gradient-based model update approaches.
Claims
-
A computer-implemented method for evidence-bound outcome learning and model updating in an AI reasoning architecture, the method comprising:
(a) retrieving outstanding prediction records whose evaluation horizon has elapsed and comparing each predicted entity state against the evidence-grounded actual state, computing a prediction accuracy delta;
(b) aggregating prediction accuracy deltas over a rolling window for each causal rule referenced by the prediction records;
(c) classifying a systematic failure pattern when the median accuracy delta for a causal rule exceeds a configurable systematic failure threshold;
(d) generating a rule update proposal identifying the failing causal rule, proposing a candidate modification, and estimating the accuracy improvement, with all triggering evidence artefact hashes included as justification; and
(e) submitting the rule update proposal to a governance enforcement engine, and on approval, applying the modification to the shared causal rule set and recording the applied update in an append-only causal rule audit log.
-
The method of claim 1, wherein prediction accuracy deltas for categorical entity states are binary, and deltas for numerical attributes are normalised absolute differences.
-
The method of claim 1, wherein each entry in the causal rule audit log includes the rule update proposal hash, the governance approval record hash, the causal rule version identifiers before and after the update, and the application timestamp.
-
The method of claim 1, wherein rejected rule update proposals produce a negative proof record retained in the evidence ledger, preserving evidence that the systematic failure was detected.
-
The method of claim 1, wherein the method further computes an accuracy improvement estimate as the predicted delta reduction the proposed rule modification would have achieved on the historic failing predictions.
-
An outcome learning and model update engine comprising one or more processors, 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 outcome learning and model update engine is disclosed for evidence-bound AI reasoning architectures. Prediction records from prior reasoning sessions are compared against subsequently admitted evidence to compute prediction accuracy deltas. Systematic failure patterns are detected when median deltas for specific causal rules exceed a configurable threshold. Rule update proposals are generated from systematic failures and submitted to the governance enforcement engine; approved updates are applied to the shared causal rule set and recorded in an append-only causal rule audit log with full traceability from triggering evidence to governance-approved modification. All rule updates are reversible and auditable. v +---------------------------------------+ | Causal Rule Update Applied | | (recorded in Rule Audit Log) | +---------------------------------------+
---
## Detailed Description
[0009] **Prediction Registry.** The OLMUE maintains a Prediction Registry storing all outstanding predictions: hypothesis simulation outcomes (P204), goal-state predictions (P210), and action outcome predictions (P206). Each entry includes the predicted world state, the temporal horizon, and the evidence artefacts on which the prediction was based.
[0010] **Comparison Protocol.** When a new evidence artefact is admitted, the OLMUE retrieves all predictions whose horizon falls at or before the artefact's world_time. For each retrieved prediction, the OLMUE compares the predicted entity states to the actual entity states in the CWSG as updated by the new artefact.
[0011] **Accuracy Delta.** The accuracy delta is computed as the fraction of predicted state changes that were confirmed by the new evidence. A delta below the configurable threshold (default: 0.6) triggers a systematic failure flag for the relevant causal rules.
[0012] **Rule Update Proposal.** When systematic failures are detected for a causal rule, the OLMUE generates a rule update proposal specifying: the rule identifier; the proposed change (parameter adjustment, applicability condition, or rule suspension); and the supporting evidence artefact hashes.
[0013] **Governance Gate.** All rule update proposals pass through the safety governance gate (P215) before application. The gate evaluates whether the proposed change falls within permitted model evolution constraints defined in the governance policy.
[0014] **Audit Trail.** Applied rule updates are recorded in a Rule Audit Log maintaining complete history of causal rule evolution, enabling inspection of any prior model state by replaying the audit log to a specified sequence number.
---
## Claims
1. An outcome learning engine for an evidence-bound reasoning architecture that updates causal rules based on comparison of predictions with admitted evidence.
2. The system of claim 1 wherein all rule updates are proposed through a governance gate before application.
3. The system of claim 1 wherein rule update history is maintained in an append-only audit log.
4. The system of claim 1 wherein prediction accuracy is computed against evidence-grounded actual world state.
5. The system of claim 1 wherein the complete history of model evolution is reconstructible from the audit log.