Why Rules and Prospectuses Drift Apart
The BIS supervisory screening procedure starts with a problem that ordinary text matching cannot solve: a regulatory rule and a financial instrument’s prospectus may describe the same obligation in different language without having the same meaning. This chapter explains how to identify the comparison target, why AT1 loss-absorption terms make it consequential, and how the procedure records a possible divergence without declaring a breach. That distinction lets a reviewer use the model’s output as evidence to examine rather than as a conclusion to accept.
A rule sets the condition; a prospectus sets the contract
A regulation and a prospectus are not two versions of the same document. They are two parts of an institutional arrangement. The regulatory package states the conditions an instrument must satisfy under the applicable capital framework. The prospectus states the contractual terms on which the instrument is issued. It explains to investors and other parties how the instrument pays, ranks, and absorbs losses. The contract must implement the rule, but implementation is not the same thing as repetition.
That distinction creates a normal and a problematic kind of difference. A prospectus may add detail that the regulation leaves implicit, specify how a permitted discretion will operate, or translate a general requirement into contractual language. Those differences can be legitimate. The problem begins when the added language narrows the circumstances in which a rule operates, extends an obligation beyond its scope, leaves the relationship unclear, omits an important condition, or prescribes a consequence that conflicts with the rule. Finding that boundary is a supervisory task, not a dictionary exercise.
AT1 instruments show why the boundary can remain dormant until a bank is under pressure. These perpetual securities may trade in normal times much like subordinated debt, paying discretionary, non-cumulative coupons. Their capital function becomes visible in stress. The issuer can conserve capital by cancelling coupons, and the instrument can absorb losses through a write-down or conversion into common equity after a contractual trigger is breached or an authority determines that the bank has reached the point of non-viability.
Four families of terms shape that loss-absorption mechanism. A mechanical trigger sets the capital threshold at which the process begins. A point-of-non-viability, or PONV, provision describes the circumstances in which an authority can require loss absorption. Coupon terms determine when payments can be cancelled without putting the issuer into default. Hierarchy determines whether common equity or AT1 bears losses first. The terms are related, but they do different work. A change to a threshold is not the same kind of change as a change to sequencing, and a condition attached to public support is not merely another way to describe a coupon.
Consider the practical consequence of a qualifier. If a rule permits loss absorption when a specified supervisory condition occurs, a prospectus that adds an extra condition can reduce the situations in which the instrument responds. The added words may have no visible effect while the bank remains healthy. They matter when an authority, investors, and the issuer must determine whether the contractual trigger has occurred. The reverse problem is also possible: a prospectus may describe an obligation more broadly than the rule, imposing a burden or exposure in circumstances the rule does not cover. In both cases, the important question is how the contractual wording changes the rule’s operation.
The applicable comparison is therefore contextual. A reviewer needs to know which regulatory package governs the instrument, which prospectus terms address the same requirement, and how surrounding provisions affect the apparent match. A phrase can look equivalent when isolated and become materially different when its condition, timing, or consequence is read with the rest of the clause. Conversely, a phrase can look different because the contract has made a general rule more specific without changing its effect.
This is why the target is not “similarity” in the everyday sense. A similarity search can find shared words. A supervisory comparison has to identify corresponding requirements and ask whether the contractual implementation preserves their scope, conditions, and consequences. The former is useful for retrieval; the latter is the beginning of interpretation.
The four AT1 term families also produce four different supervisory questions. For a mechanical trigger, the issue is whether the contractual threshold or condition starts loss absorption at the same point. For PONV, it is whether the prospectus preserves the circumstances in which an authority can require the measure. For coupons, it is whether cancellation remains available under the same conditions and without creating a default. For hierarchy, it is whether the order in which common equity and AT1 absorb losses has been preserved. A model that treats all four as generic “capital wording” would erase the mechanism that gives each difference its significance.
The comparison can therefore reveal a direction of change without determining its desirability or legal effect. A prospectus that is more detailed may make a procedure easier for investors to understand while also introducing a condition that is absent from the rule. A clause that uses broader language may appear more protective in one setting and create a different contractual exposure in another. The relevant question is always tied to the particular requirement and its consequences. That is why a candidate needs the surrounding text and the applicable regulatory context, not just a highlighted sentence.
The procedure turns a reading task into a comparison record
The BIS procedure gives the LLM a defined comparison rather than asking it for a general opinion about a prospectus. For each comparison, an analyst supplies the full regulatory package, the prospectus, and a structured prompt. The model identifies provisions that address the same requirement, quotes the relevant passages, and flags wording in the prospectus that may alter the requirement’s scope, conditions, or consequences. The result is a record in which a possible issue remains attached to the text that prompted it.
The sequence has an important asymmetry. The analyst supplies the governing materials; the LLM performs the initial alignment; and the record preserves the paired passages for a later reviewer. The model is not asked to infer a free-floating compliance score. Its output should make it possible to answer a narrower question: which contractual language appears to address this requirement, and in what way might the wording change it?
“Corresponding” does not mean that the two passages have the same length or use the same vocabulary. It means that both purport to govern the same requirement or event. The regulatory package may express that requirement across several provisions, while the prospectus may gather the contractual consequence into one clause or distribute it across definitions and cross-references. The analyst’s choice of the applicable package and prospectus is therefore part of the comparison’s foundation. If the wrong jurisdiction, instrument, or version is supplied, a precise comparison can still answer the wrong question.
The structured prompt matters for a similar reason. It gives the model a repeatable request to identify the corresponding provisions, quote them, and examine changes in scope, conditions, or consequences. That structure does not make the output deterministic or legally authoritative. It does, however, make the expected record more inspectable than an open-ended request such as “review this prospectus.” The reviewer can see whether the model followed the intended comparison and where it departed from it.
The BIS procedure organises that question into five types of candidate divergence. Each type describes the relationship between the two texts, not the legal result.
Narrowing occurs when the prospectus adds a condition or limitation that the rule does not impose. The contractual promise may consequently apply in fewer circumstances than the regulatory requirement appears to contemplate. A narrowing is a signal to examine whether the added condition is authorised or consequential; it is not, by itself, proof that the contract is unlawful.
Expansion occurs when the prospectus extends a requirement beyond the circumstances covered by the rule. This could make a contractual obligation operate more broadly than its regulatory source. Whether that matters depends on the provision, the parties’ rights and duties, and the applicable framework. The label identifies direction, not severity.
Omission occurs when the prospectus leaves out a condition, trigger, procedural step, or other element of the rule. An omission may be obvious in a short clause or hidden in the absence of a qualification that appears elsewhere in the regulatory package. The reviewer must determine whether the missing element is operative, supplied by another provision, or immaterial in context.
Ambiguity occurs when the prospectus wording leaves its relationship to the rule unclear. Ambiguity is not the same as contradiction. Two provisions may point in different directions without expressly prescribing incompatible conditions. The uncertainty itself can be important if it becomes a dispute over timing, authority, or consequence, but the model cannot decide how an adjudicator would resolve it.
Contradiction occurs when the texts prescribe inconsistent conditions, sequencing, or consequences. A rule may require one order of loss absorption while a prospectus preserves another, or the two texts may attach incompatible conditions to the same event. Contradiction is the strongest-sounding label, but it still describes the textual relationship. It does not answer whether a later rule superseded an earlier term, whether the contract complied when issued, or what remedy is available.
These categories are useful because they give expert review a common vocabulary. They also impose discipline on the model’s language. “Narrowing” tells a reviewer to look for an added constraint. “Omission” directs attention to a missing element. Neither word should be silently expanded into “breach,” “invalid,” or “must be remedied.” The category is a route into the relevant evidence, not a substitute for the authority that interprets it.
The paired quotations serve the same discipline. An unsupported summary can hide which words were compared, omit a limiting sentence, or make the model’s interpretation sound more settled than the source text allows. When the record shows the relevant passage from the regulatory package beside the corresponding passage from the prospectus, an expert can challenge the match itself. They can ask whether the provisions really address the same requirement, whether the model missed a cross-reference, whether another clause supplies the omitted condition, or whether the apparent contradiction depends on a different effective date or jurisdiction.
That is also why the procedure is not finished when it has assigned a label. A candidate can be valuable even when an expert rejects it: the record may reveal an unclear drafting choice, a missing source passage, or a need to inspect the full document. Conversely, a plausible label can be dangerous if it encourages a reviewer to stop reading. The quality of the procedure lies in how well it directs attention while keeping the underlying evidence visible.
A usable comparison record supports at least three kinds of follow-up. First, an expert can verify the model’s basic reading by checking whether the passages really concern the same requirement. Second, they can investigate the direction of the difference by reading definitions, exceptions, and connected provisions in the broader texts. Third, they can document a disposition: the difference may be permitted, may need clarification, may warrant escalation, or may justify further supervisory action. The model need not produce that disposition for its output to be useful. It needs to leave enough evidence for the responsible person to reach, contest, and record it.
This also explains why the categories should not be treated as a ranking of seriousness. A contradiction may be harmless if it reflects a superseded or inapplicable provision; an ambiguity may be highly consequential if it arises at a point where parties must decide whether loss absorption begins. The label tells the reviewer what to examine. Its importance emerges from the rule, the contract, the facts, and the decision that could follow.
The central mechanism can now be stated precisely. The regulatory package supplies the requirement, the prospectus supplies the contractual implementation, and the LLM proposes a correspondence plus a description of the possible difference. The description may concern narrowing, expansion, omission, ambiguity, or contradiction. The supervisory and legal task begins after that description: determine whether the difference is permitted, material, actionable, or unresolved.
A single model response cannot make that process reliable, because language-model outputs can vary even when the inputs do not. The comparison record is therefore only the first layer. The next layer must determine how repeated outputs should be consolidated and ranked, and how much confidence—if any—such recurrence deserves.