A Candidate Divergence Is Not a Regulatory Judgment

The BIS screening queue now contains three different kinds of information: a relationship observed between two texts, a recurrence signal that helps order attention, and an expert decision about what the relationship means. This checkpoint connects those layers and turns them into a usable judgment model. The reader should be able to take any candidate record and separate four questions—whether the texts differ, whether the difference is permitted, whether it is material, and what action follows—without asking the LLM to answer questions that belong to supervisory or legal authority.

The queue answers one question, not four

A candidate record can be highly informative without being a conclusion. At its narrowest, it says that a provision in the regulatory package and a provision in the prospectus appear to address the same requirement, and that their wording may differ in scope, conditions, or consequences. The divergence category gives the difference a direction: perhaps the prospectus adds a condition, extends a requirement, omits an element, leaves the relationship unclear, or prescribes an inconsistent consequence. Recurrence then indicates how consistently equivalent candidates appeared across the independent runs.

That record can answer the first question—did the texts appear to differ? Even there, the answer remains reviewable rather than final. An expert must check that the passages actually concern the same requirement, that the comparison used the applicable regulatory package and prospectus, and that definitions, exceptions, and surrounding clauses have not changed the apparent relationship. The model’s pairing is a well-organised claim about the documents, not a fact that becomes immune from inspection because it was repeated.

The second question is different: is the difference permitted? That question depends on the authority and scope of the applicable rule. A prospectus may add detail that the rule allows, or a condition may be supplied by another contractual provision. A rule may also have changed after an instrument was issued, producing a difference that cannot be understood by looking at present wording alone. Determining permission may require jurisdiction, effective date, cross-references, and the legal significance of the instrument’s contractual terms. A language model can make the relevant passages easier to find, but the supplied texts do not give it authority to settle their legal relationship.

The third question is whether the difference is material. Materiality is not another name for textual distance. A long change in explanatory language may have little practical effect, while a short qualification to a trigger may matter greatly when a bank is in distress. Materiality depends on what the clause controls, who may be affected, and what decision or loss allocation could follow. It is possible for experts to agree that a difference exists and still disagree about its importance in the relevant supervisory setting.

The fourth question is what action follows. A permitted variation may need no intervention. An unclear relationship may warrant clarification or a broader review. A material difference may require escalation to legal or policy specialists, a supervisory response, or a change in how the instrument is monitored. The appropriate action also depends on the authority’s mandate and procedure. The queue can direct attention to the decision; it cannot manufacture the power to take it.