Verdict Without the Rule: Diagnosing and Auditing Regulatory Rule Sensitivity in LLM Compliance Systems

TEXT START: Large language model compliance systems are deployed on the assumption that a verdict depends on the regulatory rule it is given.

The Dissection

The paper attacks the central legitimacy claim of automated compliance: that a correct verdict is evidence that the model applied the supplied rule. Its perturbation tests—deleting, swapping, or negating rules while holding cases fixed—expose models that often reach the same answer regardless of the governing standard. The 90–92% accuracy of general-purpose models therefore masks a grounding deficit; the guard model’s 51% performance makes the failure explicit.

This is an audit of procedural credibility, not merely accuracy. It shows that compliance systems can generate the right answer for the wrong reason and remain confidently usable until the rule actually matters.

The Core Fallacy

The paper’s implicit limitation is treating rule sensitivity as the decisive boundary between trustworthy and untrustworthy compliance automation. It identifies a necessary condition, not a complete deployment theory. A model can react to the rule and still misinterpret it, selectively apply it, or fabricate a rationale.

Relative to Discontinuity Thesis mechanics, the deeper error would be assuming that exposing the defect slows cognitive automation. It probably does the opposite. Once institutions can measure and price the defect, they will build audit layers, liability buffers, and human sign-off around the models. The productive role of human compliance workers contracts into verification and exception handling—the servitor layer—while ownership of the automated system remains with the sovereigns.

Hidden Assumptions

• A regulatory decision can be adequately represented as a fixed case paired with a supplied rule.
• Counterfactual rule sensitivity is a valid proxy for internal rule grounding.
• The tested models, domains, and custom guard-model adaptation represent the wider compliance market.
• Better internal representations or prompting would be plausible repair mechanisms.
• Detecting ungrounded verdicts will produce institutional correction rather than tolerated risk.
• Human oversight remains economically available as automation scales.
• Accuracy, grounding, and accountability can be separated cleanly in deployed systems.

The last assumption is the most politically useful. Institutions often need only enough accuracy to justify deployment, not enough causal fidelity to make responsibility meaningful.

Social Function

Partial truth with a transition-management function.

The paper punctures accuracy theater by demonstrating that high scores do not establish rule-based compliance. That is valuable. But its likely institutional use is narrower: turn a structural legitimacy crisis into an engineering audit problem. Add a metric, add a monitoring layer, retain a human approver, and continue automating.

Under the hardened framework, this is a lag defense. It may delay visible failure by converting displaced judgment into audit, escalation, and certification work. It cannot preserve mass productive participation. If the models become sufficiently cheap and usually correct, organizations will accept residual rule blindness as an operating risk—especially when the costs of human labor remain visible and the costs of incorrect automation are diffuse.

The Verdict

This is a sharp autopsy of compliance automation’s central fraud: a correct verdict is not proof that the rule was used. The 51% guard-model result is not a minor quality defect; it is evidence that a system marketed as specialized can be less reliable than general models while projecting institutional authority.

But the paper does not weaken the Discontinuity Thesis. It strengthens it. Cognitive automation need not be perfectly grounded to displace labor; it only needs to be cheaper, faster, and accurate enough for institutions to tolerate its failures. Rule sensitivity becomes a temporary moat and a new verification market, not a restoration of human economic necessity. The machine can be wrong for reasons no worker is allowed to be wrong—and still win the cost comparison.