P105 — AIEP — AGI Governed Uncertainty Quantification Extension

applicant: Neil Grassby inventor: Neil Grassby status: review — to file — URGENT classification: Patent Application — Confidential priority: Claims priority from GB2519711.2 filed 20 November 2025

Field of the Invention

[0001] This invention concerns governed epistemic uncertainty quantification within AIEP reasoning substrates — specifically the deterministic mechanism by which a substrate produces a cryptographically committed UncertaintyRecord for each output, quantifying uncertainty arising not from evidence absence (covered by P16 NegativeProofRecord) but from contested, contradicted, or substrate-boundary evidence conditions.

Background

[0002] P16 provides NegativeProofRecord — cryptographic commitment that evidence for a claim does not exist in the corpus. This covers absence. It does not cover the condition where evidence exists but is contested, where admitted evidence contradicts other admitted evidence without resolution, or where a claim lies at the boundary of current substrate knowledge without sufficient weight on either side.

[0003] EU AI Act Article 13 requires high-risk AI systems to provide transparency enabling users to interpret outputs. Epistemic uncertainty quantification — not just confidence scores, but structured governed uncertainty with cryptographic commitment — is the architectural implementation of Article 13 at substrate level.

Summary of the Invention

[0004] An UncertaintyRecord is generated for each substrate output comprising four canonical uncertainty fields:

• EvidenceContestationScore — normalised measure of active ContradictionRecords (P112) bearing on the output’s evidence base
• BranchDivergenceScore — normalised measure of current challenger branch weight relative to dominant branch weight
• SubstrateBoundaryFlag — boolean indicating the output domain approaches the boundary of the substrate’s evidence corpus
• ResolutionPathIndicator — structured indication of what evidence type would resolve current uncertainty

[0005] An UncertaintyHash is computed as H(EvidenceContestationScore ‖ BranchDivergenceScore ‖ SubstrateBoundaryFlag ‖ ResolutionPathIndicator ‖ OutputHash ‖ SchemaVersionId).

[0006] UncertaintyRecord is a mandatory output field at Tier 2 and above. At Tier 0 it is null. At Tier 1 it carries EvidenceContestationScore only.

Claims

1. A governed uncertainty quantification system within an AIEP substrate, configured to: compute a four-field UncertaintyRecord for each substrate output comprising EvidenceContestationScore, BranchDivergenceScore, SubstrateBoundaryFlag, and ResolutionPathIndicator; compute UncertaintyHash binding the UncertaintyRecord to the output hash; and include UncertaintyRecord in substrate output at Tier 2 and above.

2. The system of claim 1 wherein EvidenceContestationScore is computed from active ContradictionRecords bearing on the output’s evidence base.

3. The system of claim 1 wherein BranchDivergenceScore is computed from the weight ratio of challenger branches to the dominant branch.

4. A computing system comprising processors executing the method of claim 1.

Drawings

FIG. 1 — Architecture diagram (see filed application for figures)

Figure 1 — UncertaintyRecord Four-Field Structure

Substrate Output Event
       │
       ▼
┌─────────────────────────────────────┐
│          UncertaintyRecord          │
│  evidence_contestation_score ◄──────┼── ContradictionRecords
│  branch_divergence_score     ◄──────┼── Challenger branch weights
│  substrate_boundary_flag     ◄──────┼── Corpus edge proximity
│  resolution_path_indicator   ◄──────┼── Required evidence descriptor
│  UncertaintyHash = H(all fields)    │
└─────────────────────────────────────┘

Figure 2 — EvidenceContestationScore and BranchDivergenceScore Derivation

┌──────────────────────┐   ┌─────────────────────────────┐
│  Active              │   │  Branch Weight Distribution │
│  ContradictionRecords│   │  Dominant   Challenger      │
│  bearing on output   │   │  ─────────  ────────────    │
│  evidence base       │   │  W_dom      W_chal_sum      │
└──────────┬───────────┘   └──────────────┬──────────────┘
           │                              │
           ▼                              ▼
EvidenceContestationScore        BranchDivergenceScore
= f(contradiction_count,         = W_chal_sum / W_dom
    weight_of_contradictions)

Figure 3 — Tier-Based Output Inclusion

EvidenceContestationScore < LOW_THRESHOLD ──► Include without flag
EvidenceContestationScore >= LOW_THRESHOLD ──► Include with uncertainty flag
SubstrateBoundaryFlag = TRUE               ──► Include with boundary warning
ResolutionPathIndicator present            ──► Display resolution path

Figure 4 — Fail-Closed Gate on Missing UncertaintyHash

Substrate Output Generation
          │
          ▼
┌─────────────────────────┐
│  UncertaintyHash        │
│  present in output      │
│  record?                │
└──────┬──────────────────┘
       │ YES              NO
       ▼                  ▼
    PROCEED           SUPPRESS OUTPUT
                      NegativeProofRecord
                      committed

Abstract

A governed uncertainty quantification system is disclosed extending P16 NegativeProofRecord to cover contested, contradicted, and boundary-condition uncertainty. A four-field UncertaintyRecord is generated for each substrate output and bound by UncertaintyHash. EvidenceContestationScore reflects active contradictions. BranchDivergenceScore reflects challenger branch weight. SubstrateBoundaryFlag indicates proximity to corpus edge. ResolutionPathIndicator identifies what evidence would resolve uncertainty. Directly implements EU AI Act Article 13 transparency requirements at substrate level.