Enforcement or Instrument: the base rate nobody quotes, and why public data cannot settle it

How to read this. [measured] marks a figure from a named primary source with its retrieval date; every count below is reproducible with the tool named at the end. [I] marks our own argument. The paper reaches no conclusion about anyone's motives, and does not contain one.

What this paper is not. It is not a claim that any investigation was proper or improper. We have no access to a sealed docket, no view of a charging decision, and no opinion we could support about an individual case. The paper is about a measurement gap, and it was written because we noticed we could not answer a question we had opinions about.

The argument is conducted without rates

[measured] In the four days to 2026-10-13 our own news corpus recorded, from NPR and the BBC: federal agents searching the Washington home of a sitting member of Congress, with the Justice Department confirming a criminal investigation; a Justice Department antitrust investigation of television networks, which the BBC's headline connected to coverage of the President; and a report that the Department's civil-rights mission has been reshaped in its work and its workforce. [measured] From outside the United States in the same week: an instruction from Israel's Prime Minister to identify and prosecute the interviewees in a documentary film.

[I] Each is a fact. None of them is a rate, and the difference is the whole problem. "A legislator's home was searched" tells you an event occurred. Whether that event is unusual — whether it happens twice a decade or twice a year, and whether the frequency has changed — is a question about a distribution, and no participant in the public argument quotes one. Both sides are therefore arguing about the interpretation of a single observation, which is a conversation that cannot terminate.

[I] We include the Israeli item deliberately. A structural claim about enforcement power that only ever reaches for American examples is a political claim wearing a structural costume. If the mechanism is real it shows up in more than one system.

So we computed one

[measured] CourtListener, operated by the Free Law Project, exposes a public API over the RECAP archive of federal court records. Queried on 2026-10-14 it reports 60,739,259 records for a common control term and 452,401 records matching the phrase "search warrant" across all years. Counting by filing year, with a control series over the same corpus and the same years:

[measured] From 2021 to 2025 the raw count fell 30.9%. Over the identical period the control series rose 22.0%. The share therefore fell from 2.601% to 1.474%, a relative decline of 43.3%. The 2026 row covers 287 days; annualised it implies about 18,116 records, 6.9% below 2025, so the decline flattens rather than accelerating.

[I] The control matters more than the headline. A raw decline of 31% is the kind of figure that gets quoted on its own, and it would have been wrong to quote: the obvious objection is that the archive shrank, and the obvious objection is false, because the denominator moved the other way. Normalising makes the decline larger, not smaller. That is the one thing in this paper that behaved better than expected.

And the number does not mean what it looks like

[I] Here is where a paper that wanted a conclusion would stop. We are not able to.

[measured] RECAP contains what was docketed and unsealed. Search-warrant applications are routinely filed under seal, many are never unsealed, and a sealed proceeding does not appear in this archive at all. So the series measures the visibility of warrant proceedings, not their occurrence.

[I] That makes a 43% fall in visible share consistent with at least four different worlds, and the data cannot distinguish them:

1. Fewer warrants. Activity genuinely declined.
2. More sealing. The same activity, less of it visible. This is the possibility that should worry a reader most, because it is the one under which a declining series is evidence of reduced accountability while reading as reassurance.
3. Changed docketing conventions. The same activity, recorded differently, so the phrase stops matching.
4. Changed archive contribution. RECAP grows when its users fetch documents. A shift in who uses it, or in what they fetch, moves this series without anything changing in any courthouse.

[I] Note the asymmetry, because it is the reason we will not report the decline as a finding about enforcement. The first two explanations point in opposite directions — one says the state is doing less of something, the other says the public can see less of it — and they produce the same number. A measurement that cannot separate "less happening" from "less visible" is not a measurement of what happened.

Why this is hard in principle, not just in practice

[I] A reader might assume better data would settle it. The literature says the problem survives better data, and the clearest result comes from a setting with no politics in it at all.

[measured] Gerardino, Litschig and Pomeranz (American Economic Journal: Applied Economics, 2024) found that audits of Chilean public procurement reduced the use of competitive auctions, because auctions "underwent more than twice as many checks and led to twice as many detected infractions" than comparable direct contracts. Detected infractions scale with how auditable a procedure is, not with how much is being stolen.

[I] Transposed, that is the central difficulty. Investigations concentrate where investigation is tractable — where records exist, where jurisdiction is clean, where a statute is clear. Tractability is not randomly distributed across people or institutions. So a pattern of enforcement that correlates with some political characteristic is exactly what you would observe whether or not anyone intended it, because political characteristics correlate with being the kind of institution that keeps records. Counting cases cannot distinguish targeting from auditability. Any honest analysis has to model the selection, and modelling the selection requires knowing the denominator — which the preceding section says we do not have.

[measured] A further pair of results bound what any fix can expect. Olken (2009), measuring Indonesian road construction physically and then surveying the villagers beside those roads, found the elasticity of perceived corruption to actually-missing expenditure was about 0.08. Across 26 countries the individual-level correlation between perceiving corruption and having experienced it is 0.09 (Donchev and Ujhelyi, 2014).

[I] Perception is almost uncoupled from measurement in this domain. That cuts against everyone: it means public confidence that enforcement is being abused is weak evidence that it is, and public confidence that it is not is equally weak evidence that it is not. Neither side's certainty is informative, which is an uncomfortable thing to say to both.

The fix already exists, for a narrower instrument

[I] The useful discovery in writing this was that American law already solves precisely this reporting problem — for wiretaps, and not for warrants.

[measured] 18 U.S.C. § 2519 requires that in January of each year, any judge who issued an interception order or an extension, or who denied approval of one, report to the Administrative Office of the United States Courts: that an order was applied for; the kind applied for; whether it "was granted as applied for, was modified, or was denied"; the period authorised and any extensions; and the offence specified. [measured] Subsection (3) requires that in June of each year the Director of the Administrative Office "transmit to the Congress a full and complete report concerning the number of applications for orders authorizing or approving the interception of wire, oral, or electronic communications…and the number of orders and extensions granted or denied." [measured] The resulting Wiretap Reports are published by the federal judiciary.

[I] Read what that statute gets right, because it is a template rather than an analogy. It is reported by the judge, not the prosecutor, so the reporting party is not the party with an interest in the number. It counts denials, which is the single most informative figure and the one no voluntary disclosure ever includes — a grant rate close to certainty means something very different from one with visible refusals, and you cannot compute either without the denominator. It is aggregate, so it survives sealing: nothing about an individual investigation is revealed by a district-level count of applications made and refused. And it goes to Congress, annually, by operation of law rather than by request.

[I] Search warrants have no equivalent. There is no annual public count of federal warrant applications, no count of refusals, and no breakdown by offence or district. The consequence is the one documented above: the public can observe warrants only when somebody chooses to unseal one, which makes the visible series a measure of disclosure policy.

[I] The proposal, then, is unexciting and already drafted by Congress: extend the § 2519 model to search-warrant applications. Judge-reported, aggregate, including denials, by district and offence class, annually. It reveals no sealed investigation. It would make every claim in the current argument — including the ones we would like to be true — checkable within a year.

What each explanation predicts

[I] The point of a measurement proposal is that it can fail. If warrant reporting existed on the § 2519 model, the four competing worlds separate:

[I] And the targeting question becomes tractable in the same move. Enforcement used as an instrument predicts a grant rate and an offence mix that shift with administrations; auditability predicts a stable grant rate with volume tracking record-keeping. Neither prediction is available today, which is why the argument does not end.

Reproducing this, and what we will not claim

[measured] Every count here is reproducible with tools/courtlistener_research.py, which queries the public API, paces itself at 13 seconds because the anonymous limit is five requests a minute, and writes the raw series to disk so a figure can be re-derived from the same bytes. The two series behind the table are search_warrant_series.json and control_court_series.json, retrieved 2026-10-14. No token was used and none is needed.

[I] What we are not claiming, stated plainly because the subject invites readers to supply a conclusion we did not write:

• We do not claim enforcement has been used as a political instrument. The data cannot support that, and we would have to withdraw it.
• We do not claim it has not been. The data cannot support that either, and the selection-effect section says public confidence in either direction is weak evidence.
• We do not claim the 43% decline is a decline in enforcement activity. The four readings above stand and we cannot choose between them.
• We have read no sealed docket and no charging file, and nothing here bears on any individual case or person.

[I] This is the third paper this month to end in the same place, which is starting to look like a pattern rather than a coincidence. Below n log n found a published complexity result whose practical content was a speedup of one part in 10⁵⁴. What Holds When the Money Moves found illicit-flow estimates spanning a factor of a thousand with no replicable denominator, and a United Nations indicator still at Tier II. In researching the European researcher-data rules this week we found that the mechanism built to make platforms auditable publishes nothing about its own operation: no register of who was granted access or refused. Now a question about the use of state power turns on proceedings that are sealed by default.

[I] In all four, the obstacle is not that the answer is hidden by someone. It is that nobody arranged to count the thing, and the absence then gets read as whichever finding the reader arrived with. The cheapest civic technology available is a published denominator, and it keeps being the thing that does not exist.