Equivalent cases should not produce inexplicably different decisions

R003DECISION SYSTEMS · FINANCIAL SERVICESNikolai Mathias FurusethFenori Technologies, August 2026

Two compliance cases can look different and still be materially equivalent. A payment narrative rephrased, the same facts arriving in a different order, an analyst’s summary written more tersely: none of that changes what the institution’s policy should conclude. They should not receive different treatment simply because one was phrased differently.

At the same time, two cases that look similar should not receive the same treatment when a material risk factor has changed. A counterparty newly connected to a flagged network is not the same case as its lookalike from last quarter, however similar the paperwork appears.

That creates two important tests for any AI-assisted compliance system, and most evaluation practice today applies neither.

Invariance and sensitivity

Invariance: immaterial changes should not change the outcome. If the same underlying facts, expressed differently, produce different dispositions, the system is responding to phrasing rather than to policy, and every disposition it has ever produced becomes harder to defend.

Sensitivity: material changes should change the outcome, or at least the analysis. A system that treats a materially different case identically because it pattern-matches the surface is failing in the more dangerous direction: it is missing risk while appearing consistent.

Hallucination testing is not decision testing

Most evaluation of AI in compliance today asks whether the model invented a fact, cited a nonexistent source, or drifted from its prompt. Those checks matter, but they operate at the level of the model’s output. The institution’s exposure sits one level higher: at the decision.

The deeper question is whether the institution’s policy, evidence requirements and authority model are being applied consistently across cases. A model can be perfectly factual and still produce dispositions that vary with wording, evidence order, model version or time of day. Every one of those inconsistencies is a finding waiting to be written, by an internal auditor, a model-risk reviewer or a supervisor.

What decision-level testing looks like

This is measurable, and it has been measured. SCORE, presented at the NAACL 2025 Industry Track, held the underlying question fixed and varied only its surface: option order, prompt phrasing, few-shot seeds. Accuracy moved by up to ten percentage points on MMLU-Pro and 6.1 points on AGIEval. Errica and colleagues, at NAACL 2025, formalised the same property as separate sensitivity and consistency measurements, on the argument that a single accuracy figure cannot express it.

The literature does not agree that this is entirely a defect of the system under test. Hua and colleagues, at EMNLP 2025, argue that a meaningful share of measured prompt sensitivity is an artifact of how evaluations are built rather than instability in the model. Both readings lead an institution to the same place. If the measurement itself is contested, a vendor’s benchmark number is not evidence about your cases, and the institution has to construct its own.

Testing the decision rather than the output means constructing case families deliberately: equivalent cases with different wording; cases with irrelevant details added; cases where exactly one material factor changes; the same case executed repeatedly, and across model versions; cases with missing evidence; cases with conflicting evidence; cases evaluated under different policy versions; and cases that must route to a human regardless of what any model concludes.

For each family, the expected behavior is defined in advance, not by the model, but by the institution’s policy. The system passes when equivalence produces consistency, materiality produces change, missing evidence produces escalation rather than confident completion, and human-authority conditions produce a human every time.

Consistency is an architecture property

A system cannot be tested this way as an afterthought. It has to be built for it: policy expressed as something executable rather than prompt prose; evidence retrieval recorded, so what the decision saw is knowable; model and policy versions pinned to each decision; and the decision itself retained as a complete record: inputs, reasoning, both sides of the argument where the system argues both sides, and the human signer where one is required.

Princeton’s Towards a Science of AI Agent Reliability (ICML 2026) separates reliability into consistency, robustness, predictability and safety, and reports that across twelve frontier models reliability improved at roughly half the rate of accuracy on a general agentic benchmark, and roughly a seventh of the rate on customer service. Capability and dependability are not the same curve, and an institution buying the first should not assume it has bought the second.

A consequential decision should be testable, challengeable and reconstructable. Not simply generated.