Whitepaper
Evidence Fusion for High-Consequence AI
A cross-domain method for combining signals, model outputs, human context, risk, and uncertainty into one decision layer.
Abstract
High-consequence AI systems rarely depend on one signal. They need to combine statistical signals, model outputs, human context, threat intelligence, business impact, policy, and uncertainty into a decision that can be reviewed.
Publication context
This paper is part of the Evening Star AI publication series for usable AI judgment: short, decision-focused work for builders, security teams, leaders, and operators. It follows the institute's core pattern: observe context, reveal change, reason about impact, preserve uncertainty, and help humans move under governance.
Thesis
High-consequence decisions are multi-evidence decisions. A vulnerability is not urgent only because of a score. A market signal is not actionable only because of a pattern. A security alert is not real only because of one detector. The decision emerges from evidence fusion.
Evidence fusion is the method layer beneath Evening Star's decision systems. It combines heterogeneous signals into a structured judgment while preserving uncertainty, provenance, and the ability for a human to challenge the result.
Inputs
The evidence layer should accept statistical indicators, detector outputs, model classifications, human annotations, asset metadata, threat intelligence, business impact, control state, freshness, and confidence. Each input should carry provenance and reliability. A verified telemetry signal should not be treated the same as an inferred business label.
The goal is not to force all evidence into one opaque score. The goal is to produce a decision surface where signal agreement, conflict, and gaps are visible.
Fusion method
A practical fusion model has five steps: normalize evidence, weight by reliability and relevance, detect agreement and contradiction, assign confidence and uncertainty, and map the result to action. The output should show which evidence drove the decision and which missing evidence could change it.
This works across domains. In vulnerability intelligence, fusion combines exploitability, exposure, controls, and business context. In anomaly intelligence, it combines baselines, drift, detector agreement, and operator notes. In AI security, it combines policy, input risk, tool path, and sensitive-data context.
Decision layer
The result should be a recommendation with reasons, confidence, and next action. For high-consequence systems, the fusion layer should not automatically collapse uncertainty into certainty. It should expose uncertainty as a first-class output. Evidence fusion is how applied AI earns trust. It acknowledges that reality is noisy and context-dependent, then gives humans a better structured way to decide.