Home Institute The Institute Eva M. Erpenbach Research The Research Architecture Institute Essays Research Dialogues Guest Essays External Contributions Fellowship The Fellowship Fellows Fellowship Contributions Partners Project Partnerships Project Partners Young Research Contact Legal Notice
DEEN
← Institute Essays
Institute Essays / 02

The Control Worked. The Finding Was Wrong.

A deterministic control can execute flawlessly on a mistaken classification. Governance must examine the evidence required before a finding can activate a consequence.

Eva M. Erpenbach
Founder & Scientific Director, Erpenbach Institute

A familiar governance response to probabilistic systems is to place deterministic controls around them. Once a trigger is accepted, enforcement becomes stable, reproducible and auditable.

Consider a company with a zero-tolerance rule: if a transaction meets condition X, it must be stopped. X is established by a classification system. Once the system assigns the transaction to X, the control blocks it automatically and records the entire sequence. A mistaken classification can therefore enter a process that works exactly as designed.

This structure has immediate corporate-governance relevance. Provision 29 of the UK Corporate Governance Code applies to financial years beginning on or after 1 January 2026 and requires boards to review the effectiveness of material financial, operational, reporting and compliance controls and provide a declaration of their effectiveness at the balance sheet date. The first reporting under the new provision is expected in 2027.

The FRC guidance places significant judgement with the board. The board identifies which controls are material and forms its own view of effectiveness from the evidence obtained, reaching a reasonable conclusion on that basis. For controls fed by probabilistic findings, the quality of the trigger becomes relevant to what that evidence can support. The declaration turns that judgement into an external statement about the state of the control environment at a specific date.

Where certainty enters the system

Once a trigger has been accepted, a deterministic control can apply the same consequence every time and leave a complete record of what followed. That certainty ends at the trigger.

Upstream, X may come from a model output, risk score, matching system or classifier. Sanctions screening provides a familiar example. A mandatory block may follow automatically from a match whose identification depends on names, transliterations, identifiers, incomplete records and similarity thresholds. The resulting audit trail can show flawless execution while the factual basis for a particular match remains weak.

The consequence attached to the same analytical signal can vary substantially. A review queue, a payment freeze and the termination of a customer relationship place progressively greater weight on the finding that activates them. The evidentiary condition attached to each trigger can reflect those differences, allowing one analytical system to support several controls without treating every level of uncertainty as sufficient for every consequence.

Standards of proof

Law has a familiar way of separating consequence from factual sufficiency. A legal rule sets the consequence attached to an established fact, while a standard of proof sets the degree of confidence required before that fact is accepted.

In civil proceedings, the balance of probabilities provides one familiar standard. Other proceedings and legal systems apply different standards and different requirements for factual sufficiency.

A material control can use the same structural logic. Its enforcement rule specifies what happens after X is established, and its trigger design specifies what counts as sufficient evidence for X. That second specification can take several forms: a confidence threshold, corroborating information, a second source, a human review requirement or an escalation rule for equivocal evidence.

These conditions can also be made testable. Assurance can examine cases close to a threshold and verify whether required corroborating information was present. A sample can distinguish straightforward cases from borderline findings that required additional evidence or review. Escalation rules and required human review can then be tested against the cases in which the design called for them.

Designing the trigger

Two questions capture the design problem. What happens once the triggering fact is established? What level of confidence is sufficient to establish it?

The second question matters most where the downstream consequence is automatic or difficult to reverse. A mistaken classification moves through a deterministic control with the same speed, discipline and auditability as a correct one. Consistency gives every accepted finding, including an error, the full force of the rule. The relevant governance object is therefore the rule that converts an assessment into an actionable finding.

Under Provision 29, boards need evidence supporting their conclusions on the effectiveness of material controls. How certain should the organisation have to be about X before a control with this consequence is allowed to treat X as established?