◎ OS PUB Apache 2.0 ← All specifications

P149 — AIEP — Evidence Confidence Decay Model

Publication Date: 2026-03-27 Status: Open Source Prior Art Disclosure Licence: Apache License 2.0 Author/Organisation: Phatfella Ltd Schema: AIEP_OS_SPEC_TEMPLATE v1.0.1 — https://aiep.dev/schemas/aiep-os-spec-template/v1.0.1

Framework Context

[0001] This disclosure operates within an Architected Instruction and Evidence Protocol (AIEP) environment as defined in United Kingdom patent application number GB2519711.2, filed 20 November 2025, the entire contents of which are incorporated herein by reference.

[0002] The present disclosure defines the mathematical model governing how the assessed confidence contribution of an evidence artefact to a reasoning chain degrades over time, with domain-specific decay curves, positive recalibration triggers, and a confidence floor below which an artefact ceases to contribute to active reasoning chains.

Field of the Disclosure

[0003] This disclosure relates to time-dependent confidence decay models for evidence artefacts in governed artificial intelligence reasoning systems.

[0004] More particularly, the disclosure concerns a scheduled recalibration engine that adjusts per-artefact confidence contributions in active reasoning chains as those artefacts age, receive external validation or contradiction signals, or accumulate citation reputation events.

Background

[0005] The AIEP evidence trust scoring framework (P124) assigns a trust_score to evidence artefacts at admission time. This score reflects source credibility, jurisdiction fit, and content quality at the moment of ingestion but does not subsequently change in response to the passage of time or to downstream reputation signals such as citations, corrections, or retractions from other nodes.

[0006] Time is a reliable proxy for evidence reliability in many domains. A clinical trial result from ten years ago has likely been superseded, replicated, or contradicted. An economic statistic from last quarter is more reliable than one from five years ago for the same query. Treating all admitted evidence as contributing a static confidence weight produces reasoning chains whose confidence scores do not reflect the current reliability of their evidence base.

[0007] No existing AIEP specification defines a post-admission confidence decay model, a recalibration event taxonomy, or a confidence floor threshold below which artefact contributions are suspended.

Summary of the Disclosure

[0008] The Evidence Confidence Decay Model defines a confidence_contribution field on each EvidenceIndexEntry as a time-varying function of trust_score, age_seconds, and reputation_delta. At admission, confidence_contribution = trust_score. Thereafter, confidence_contribution is recalculated by the ConfidenceRecalibrationEngine at each scheduled recalibration event or on receipt of a reputation signal.

[0009] Temporal Decay Component: The temporal decay contribution of an artefact decreases as:

temporal_decay = exp(−decay_rate × age_seconds)

where decay_rate is a source-class-specific parameter (schema-defined, in units of seconds⁻¹). Decay curves are slower for stable domains (academic, regulatory, legislative) and faster for volatile domains (news, financial, sensor). The temporal decay factor is multiplied against the base trust_score to produce a time_adjusted_confidence.

[0010] Reputation Delta Component: reputation_delta is an additive adjustment to time_adjusted_confidence accumulated from reputation events:

Positive signals (+): inbound citation from a high-trust source; corroboration signal from an independent AIEP node; external expert endorsement recorded as a structured ReputationEvent.
Negative signals (−): retraction signal; contradiction with a higher-trust artefact; failed re-acquisition; correction notice from the originating publisher.

Each reputation event produces a ReputationEventRecord appended to the artefact’s reputation log and increments or decrements reputation_delta by the configured event weight.

[0011] Composite Confidence Contribution:

confidence_contribution = clamp(
    (time_adjusted_confidence + reputation_delta) × freshness_score,
    confidence_floor,
    1.0
)

where freshness_score is from P147 and confidence_floor is the minimum contribution value (default: 0.05) below which the artefact is flagged as LOW_CONFIDENCE and excluded from new reasoning chains.

[0012] Confidence Floor Suspension: When confidence_contribution < confidence_floor, the artefact enters a LOW_CONFIDENCE state. Existing reasoning chains that cite the artefact receive a LOW_CONFIDENCE_EVIDENCE_WARNING. The artefact remains in the index but is not eligible for citation in new reasoning chain queries until a recalibration event raises confidence_contribution above the floor.

[0013] Recalibration Schedule: The ConfidenceRecalibrationEngine is a scheduled background worker that recalculates confidence_contribution for all artefacts on a configurable schedule (default: daily). On-demand recalibration is triggered by any incoming ReputationEventRecord. Recalculation results are appended to the Reasoning Ledger as ConfidenceRecalibrationRecord entries.

Technical Effect

[0014] Exponential temporal decay — rather than linear decay or binary staleness flags — produces a smooth confidence gradient reflecting that evidence reliability typically declines gradually rather than dropping abruptly at a fixed age threshold.

[0015] Reputation delta accumulation allows evidence artefacts to gain confidence from subsequent corroboration or lose confidence from contradictions, without requiring manual re-evaluation: the reputation event log provides a governed, auditable trail of post-admission confidence adjustments.

[0016] The confidence floor suspension mechanism excludes low-confidence artefacts from new chains without deleting them, preserving the audit trail for any existing chains that cited them before their confidence declined.

Claims

An evidence confidence decay model for a governed AI reasoning system, the model comprising: a time_adjusted_confidence computed as the product of the artefact’s base trust score and an exponential decay factor parameterised by a source-class-specific decay rate; a reputation_delta accumulated from positive and negative reputation events recorded as ReputationEventRecords; and a composite confidence_contribution computed from time_adjusted_confidence, reputation_delta, and freshness score, clamped to a configurable confidence floor.
The model of claim 1, wherein reputation events include inbound citation from a high-trust source, corroboration signals, retraction signals, contradiction detection, failed re-acquisition, and publisher correction notices, each producing a ReputationEventRecord and an additive or subtractive increment to the reputation_delta.
The model of claim 1, wherein artefacts whose confidence_contribution falls below the confidence_floor are placed in a LOW_CONFIDENCE state and excluded from citation in new reasoning chains, while remaining in the index with a LOW_CONFIDENCE_EVIDENCE_WARNING issued to existing reasoning chains that cite them.
The model of claim 1, wherein a ConfidenceRecalibrationEngine recalculates confidence_contribution on a configured daily schedule and on receipt of any ReputationEventRecord, appending ConfidenceRecalibrationRecords to the Reasoning Ledger.

Brief Description of the Drawing

FIG. 1 — Confidence contribution over time for three source classes showing different exponential decay rates and reputation delta injection events.

FIG. 2 — Composite confidence_contribution formula components: time_adjusted_confidence, reputation_delta, freshness_score, and confidence_floor clamp.

Abstract

An evidence confidence decay model defines how confidence contributions of AIEP evidence artefacts to reasoning chains change after admission. A time_adjusted_confidence applies exponential decay parameterised per source class. A reputation_delta accumulates positive signals (citation, corroboration) and negative signals (retraction, contradiction, failure). A composite confidence_contribution combines both components with the freshness score and is clamped to a configurable floor. Artefacts below the confidence floor are suspended from new reasoning chain citations. A ConfidenceRecalibrationEngine runs on a scheduled basis and on receipt of reputation events, recording all recalibration outcomes in the Reasoning Ledger.

Licence

Apache License 2.0 — https://www.apache.org/licenses/LICENSE-2.0