Enforcement discretion

Why does one written rule produce very different treatment depending on which official applies it?

One statute. One inspectorate. Two officials, both conscientious, both trained on the same manual. One writes the notice; the other gives a warning. Change the beat, the shift or the district and the pattern is systematic, not random.

The assumption worth putting under pressure is that a rule determines outcomes. It plainly determines something. But if the same rule yields different treatment at a rate this stable, the rule cannot be the only input — and it is worth asking what happens before the rule is ever consulted.

A rule is a function: given certain facts, produce a certain outcome. Notice how much has to happen before that function is called.

First, someone has to be there. Enforcement capacity is a small fraction of the conduct a rule covers — think of speeding, parking, tax filings, building codes, food hygiene. Full enforcement is not merely expensive; it is arithmetically impossible. So non-enforcement is the resting state of every enforcement system, and every act of enforcement is a selection out of a very large pool.

Second, something decides which cases enter the pool at all. Patrol routes, inspection schedules, whether a regime is complaint-driven or proactive, targets, and simple workload all shape that. This is where most geographic variation actually enters — not in how officials read the rule, but in which conduct they are standing in front of.

Third, and only now, the rule. But the rule's predicates are rarely things you can see. "Reckless." "Suspicious." "Unfit for habitation." "Wilful." The official has to infer an unobservable from noisy cues, and the inference runs off a prior built from the cases that particular official has personally seen. Two officials with different case histories will read the same cue differently and both be reasoning correctly. Nothing has gone wrong yet, and the outcomes already diverge.

Now the part that explains why the divergence persists rather than being trained out — the load-bearing claim here. Enforcement generates a record: a ticket, a charge, a notice, a file with a person on the other end who can contest it, reviewable by a court, a supervisor, an appeal, a complaint. Non-enforcement generates nothing. No document, no aggrieved party, no data point — what Herman Goldstein called the low-visibility decision. So the feedback loop capable of correcting error is wired to exactly one of the two branches. Errors in the enforcing direction get caught and pushed back; errors in the lenient direction are invisible and free to drift. An official whose threshold sits too low learns; one whose threshold sits too high never finds out. Variation is not merely produced, it is protected.

That yields a test that could go either way. If asymmetric visibility sustains the variation, a domain where non-action is also recorded should show variation collapse. Automated detection is the cleanest case: a speed or red-light camera observes every passing vehicle, so the selection stage disappears and inference reduces to a measurement — residual variation should then reflect camera placement, not any enforcer's judgement. The prediction runs the other way too: require officials to log and justify declinations, publish the rates, and variation should compress though nobody's rule changed. Find inter-enforcer variation persisting undiminished under exhaustive automated detection, or a well-run declination-logging regime that leaves the spread untouched, and this mechanism is not the operative one.

Four neighbours border this piece, and the boundaries matter. bright-line-rules.md is the closest and shares the discretion theme, but it is about rule design — the trade between a sharp rule and a flexible standard — and it treats the applier as a parameter you design around. This piece takes the rule as given and asks why a single rule still varies. The two meet at a specific disagreement with the folk view: sharpening a rule does not remove discretion, it relocates it. The sharpest rule anyone writes is a numeric speed limit, and enforcement of speed limits varies enormously between officers and places, because the variation moved upstream into whether you were stopped at all.

everyday-legal-compliance.md asks why citizens obey; this asks why officials differ. exclusionary-rule.md is a remedy aimed at almost exactly the incentive problem described here — an official who gains from a breach and loses nothing — but for unlawful action, the visible branch; the invisible branch has no comparable remedy.

Three folk accounts fail. "Some officials are biased or lazy" cannot be the general mechanism: the same variation appears in domains with no protected-group content whatever — parking, drainage, food hygiene grading — and among officials whose stated attitudes do not differ. Bias is a direction that can ride on this mechanism, not the mechanism itself, and treating it as the mechanism predicts that screening people better would fix it, which it does not. "The rule is too vague, draft it better" is answered above. "Discretion is an abuse and should be abolished" fails on the arithmetic of capacity: since full enforcement is impossible, the real choice is between structured and unstructured discretion, not between discretion and none.

One caveat: the balance among the three sources — selection, inference, asymmetric feedback — almost certainly differs by domain, and I have not weighted them. In proactive policing selection may dominate; in technical inspection the inferential step probably does.

• How mandatory-arrest and mandatory-charging policies redistribute discretion rather than eliminate it.
• What structured discretion looks like in practice: recorded reasons, published declination criteria, audited sampling of cases not pursued.

Every term the collection defines is gathered in the glossary.