The Decision Cannot Be Averaged

The Problem with Averaging

Conventional AI systems score decisions by averaging across dimensions. A candidate solution gets 0.9 on relevance, 0.85 on accuracy, 0.75 on completeness. Average: 0.83. Looks good. Ship it.

But what if that 0.75 on completeness means a critical piece of information is missing? What if the decision domain is one where incomplete information leads to wrong conclusions?

The average hides the weakness. The weakness determines the outcome.

When Averaging Fails

Averaging works when dimensions are compensatory: strength in one area can offset weakness in another. In many consumer applications, this is fine. A movie recommendation that is slightly off is not catastrophic.

But high-stakes decisions are often non-compensatory. In clinical diagnosis, a treatment that scores well on efficacy but poorly on safety contraindications is not a good treatment. In legal compliance, a contract that is mostly compliant but fails on one material term is not compliant at all.

The mathematical operation of averaging cannot express the logical requirement of alignment: all critical dimensions must be acceptable, not just the average.

Alignment-Based Architecture

The alternative is alignment-based scoring. Instead of averaging dimensions, the system identifies which dimensions are critical and requires each to meet its threshold independently.

This is the weakest-link principle: the strength of the decision is determined by its weakest critical dimension, not by the average of all dimensions.

OrbisFramework implements this architecture through configurable decision gates. Each gate can specify which dimensions are critical, what thresholds apply, and what happens when alignment fails: human review, rejection, or escalation.

Implications for Enterprise AI

Most enterprise AI platforms default to averaging. Their scoring systems, their ranking algorithms, their decision logic: all based on weighted averages.

This works until it does not. The first time a high-stakes decision fails because a critical weakness was hidden in an acceptable average, the cost becomes clear.

Organizations deploying AI for high-stakes decisions need infrastructure that supports alignment-based architecture natively, not as an afterthought.

Further Reading

This white paper summarizes research that will be expanded in Bradley W. Petersen's doctoral dissertation at Daniels College of Business, University of Denver, expected 2027.

For the platform implementation of these concepts, see the OrbisFramework decision gate documentation in the Platform Overview.