Testimony and epistemic dependence in legal and regulatory monitoring

The rule that had already died

A general counsel signs off on a compliance memorandum in March. The memorandum cites a rule from the Federal Register, correctly quoted, correctly paginated, correctly reasoned from. The trouble is that the agency issued a superseding rulemaking notice the previous October, folded a comment-period revision into it in December, and the docket has been sitting in "final rule, effective" status for ten weeks. Nobody in the compliance function saw it. The memorandum goes out. Training materials get built on it. Vendor contracts reference it. Two quarters later an enforcement action lands, citing the current rule, and the general counsel is explaining to the board why the posture was built on a corpse.

This is not a research failure. Nobody was lazy. The legal team read the primary source, checked the citation, applied competent reasoning to the text in front of them. The text in front of them was simply no longer the law. What failed was not analysis but intake — the mechanism by which the firm's beliefs about the regulatory world were supposed to stay attached to the regulatory world as it actually stood, moment to moment. The memorandum was testimony from a source that had, unbeknownst to anyone, stopped speaking the truth in October and started speaking a different truth that nobody relayed.

Testimony, and why lawyers already know this

Almost nothing a compliance function knows about the law is known first-hand. Nobody in the general counsel's office watched the rulemaking hearing, sat on the enforcement panel, or wrote the statute. What the office holds is testimony — assertions by absent institutions: an agency, a court, a docket clerk, a Federal Register editor — accepted on their say-so because checking each one from first principles is not merely impractical but conceptually incoherent. You cannot personally re-derive an enforcement priority. You take the docket's word for it. This is epistemic dependence, and it is not a weakness peculiar to legal practice. It is the ordinary condition of finite knowers, lawyers included. The philosophical question, going back to David Hume's insistence that testimony only warrants belief once experience has shown the testifier reliable, and Thomas Reid's 1785 reply that human beings carry a native disposition to trust and be trusted, is not whether to depend on testimony. It is what kind of dependence survives contact with a testifier that keeps talking after you stopped listening.

That is exactly the compliance function's problem. The general counsel did not fail to trust a source. The general counsel trusted a source that had since revised itself, and had no mechanism for finding out.

Three ways to relate to an informant

There are only three configurations available to a system that depends on testimony rather than direct perception of every fact for itself, and legal monitoring has, at different points, tried all three.

The first is the closed corpus: ingest a body of statutes, regulations and case law once, strip it of its update trail, and reason over the frozen result. This is what happens whenever a firm builds a compliance playbook from a point-in-time review and treats it as durable. It is testimony with the informants silenced — the docket cannot be asked "has this changed" because the copy in the binder has no channel back to the docket. The March memorandum was, functionally, output from a closed corpus, even though it was written by a human rather than a machine. The failure mode is identical either way: correct history, wrong present.

The second is direct perception within a bounded scene: monitor the live docket for one matter, one agency, one jurisdiction, exhaustively and in real time, but only that scene. This solves the specific-matter version of the problem and creates a new one. A firm that watches its one active enforcement docket with total fidelity can still be blindsided by a rulemaking notice in an adjacent agency that touches the same conduct, because that notice was outside the aperture. Direct perception is certain about what it is pointed at and blind to everything else, for exactly as long as someone keeps pointing it.

The third configuration keeps the chain open across all of it: dockets, rulemaking notices, enforcement actions and case law arriving continuously, each claim tagged with its source, its date, and its revision history, so that a rule superseded in October is marked superseded in October, not merely absent from an update nobody ran. This is not a bigger corpus. It is a different relation to the informants — one in which the agency, the court and the docket clerk are still permitted to speak again, and the compliance function is still listening.

What the reductionist gets right, and what it misses

Continuous monitoring gives you more testimony, not better testimony. You have no independent grounds to trust most of these streams any more than you trusted the frozen memorandum. Scale without verification is not an epistemic advance.

The objection is correct as stated, and it is the strongest one this domain produces, because compliance functions are routinely sold "more feeds" as if volume were the fix. It is not. Ten thousand unattributed alerts about regulatory activity are no more trustworthy than one stale memorandum, and arguably less useful, because now someone has to triage the noise. The reductionist demand — independent grounds for trusting a source, not blind uptake — is exactly right.

But independent grounds do not have to exist before the fact. They can be built, and only an open chain can build them. A regulatory-intelligence stream that has flagged 1,200 rulemaking notices over three years, of which 1,187 were later confirmed as materially significant by outside counsel review, has an empirical track record a frozen citation never accumulates, because the frozen citation stopped generating evidence about itself the day it was filed away. Track record is the independent ground reductionism asks for. It is only obtainable if the source keeps speaking and someone keeps scoring it against outcomes. That scoring is precisely what closing the chain — freezing the corpus, filing the memorandum — makes impossible.

The circularity objection, and where chains are allowed to stop

Every provenance chain bottoms out somewhere unauditable — a docket clerk's transcription, a court reporter's summary, a Federal Register editor's discretion about what counts as "final." Perfect traceability is a fiction, so the advantage claimed for open monitoring is one of degree, and degrees are not a difference in kind.

This is true and does not do the work required of it. No compliance system will ever trace a rule to bedrock certainty; somewhere there is a human transcription nobody re-verifies. The claim is narrower than perfect provenance. It is that the chain is recorded rather than erased. A monitoring system that knows its reading of a rule descends from a docket entry filed at 4:52pm on a specific date, superseding a notice from October, holds a materially different epistemic position from a memorandum that has forgotten the October notice existed at all. Knowing exactly where the trail goes cold — which informant is the last one you can check — is itself compliance-relevant knowledge. The general counsel in the opening scenario did not need a perfect chain. A chain that stopped at "October, superseded, unread" would have been enough to prevent the failure. The frozen memorandum did not have a shallow chain. It had none.

The terminal rung

Beyond a configuration in which every regulatory source is still live, tagged by origin and date, and revisable when a later filing supersedes an earlier one, there is nothing further to add on this axis. More jurisdictions, more agencies, more history depth — these are matters of coverage, not of kind. The Large Language Model corresponds to the frozen memorandum: comprehensive, well-reasoned, and silently out of date the moment any source moves. The Large World Model corresponds to the single watched docket: perfectly current about one scene, blind outside it. The Large Universe Model corresponds to what regulatory monitoring is actually trying, imperfectly, to become — every stream still running, every claim carrying its own expiry condition, so that the rule superseded in October is known to be superseded in October, and the general counsel finds out before the board does.