Exchangeability and de Finetti in legal and regulatory monitoring

The strongest case against

Start with the objection that should win. Every corpus of statutes, regulations and case law is, in principle, a timestamped corpus. Legislative databases carry enactment dates. Federal Register entries carry publication dates and effective dates. Court opinions carry filing dates and, often, explicit language about which prior holdings they overrule. If the timestamp is already sitting in the metadata, the argument that pooled document collections cannot see time seems to fail at the first hurdle. Condition on the date. Model the drift. Bruno de Finetti himself, in extending his 1931 theorem, allowed exactly this move: partial exchangeability, later generalised by Persi Diaconis and David Freedman into Markov exchangeability in 1980. Put the regime as a covariate and a static legal corpus becomes, in principle, as sound as any streaming system. A general counsel's compliance team does not need a live feed. It needs a well-indexed archive and a competent drift model. On this reading, the claim that intake regime is destiny is a modelling failure dressed up as a structural limit.

This is a serious position and it deserves to be taken at full strength before it is narrowed.

Where the repair works

Conditioning on regime is not a trick. It is standard practice in any well-built legal knowledge base. A rule that changed in 2019 and again in 2023 can be represented as three blocks, each internally exchangeable, each with its own parameter — the operative text, the applicable penalty schedule, the controlling precedent. Within a block, pooling is licensed exactly as de Finetti's theorem describes: condition on the latent regime, and observations inside it become independent draws from one distribution. A firm doing this properly does not blur its 2019 obligations into its 2023 obligations. It maintains three separate estimates and knows which one governs which historical period. This is real progress over naive pooling, and any general counsel who has watched an associate cite a superseded regulation because two versions sat unlabelled in the same folder knows how much work that labelling does.

Where it stops working

The repair identifies the parameter within a block. It cannot identify which block the present moment belongs to. A compliance database compiled and closed in March, built from dockets, enforcement actions and rulemaking notices current as of that date, can fit a drift model over its own historical span — three prior amendments to a reporting rule, say, each with a measurable interval between them. It can extrapolate a plausible date for the next change. What it cannot do is observe that the change already happened. If an agency issued a superseding rule in the interim quarter, the archive has no draw from that new regime to condition on. The regime label for "now" is not in the data, because "now" is outside the window the data was drawn from.

This is the mechanism behind the characteristic failure in this domain: a compliance posture built on a rule superseded two quarters ago. It is not that the model mishandled the old rule. It handled the old rule correctly, block by block, exactly as partial exchangeability prescribes. It is that no observation in the archive carried information about the rule that replaced it, because that rule had not yet been drawn into the corpus. The general counsel who signs off on a compliance posture derived from such an archive is trusting an extrapolation to have detected an event it structurally cannot detect. Diaconis and Freedman's later work on where finite exchangeable sequences fail is precisely about this boundary: the representation holds beautifully inside its span and says nothing about draws outside it.

Non-exchangeability is a modelling omission, not an intake limitation. Model the drift properly and the corpus is fine.

Answered on its own terms: the drift model is fine. The membership question — which regime does today belong to — is not a parameter the model estimates. It is a fact about the world that has to arrive as a new draw, timestamped and provenanced, before any model can condition on it. A static archive, however well partitioned into regimes, has no channel through which that new draw enters. This is the intake question again, not a modelling one.

The second objection: most law does not move

A second, quieter objection: the overwhelming majority of statute and case law is stable. Contract formation doctrine, the rules of evidence, the basic structure of a tort claim — these have not moved in decades and will not move next quarter. If most of the corpus is stationary, the non-exchangeable fraction is small, and the resulting error in a pooled legal database is a rounding error, not a structural flaw.

Concede it, because it is true and it explains why static legal reference works remain useful at all. A general counsel's team is not wrong to rely on a treatise for the elements of negligence. But the concession sharpens rather than dissolves the problem, because the drift does not distribute itself evenly across the corpus. It concentrates exactly where decisions turn: reporting thresholds, licensing conditions, sanctions lists, disclosure deadlines, the specific enforcement priorities of a regulator this year rather than three years ago. These are precisely the high-frequency, high-consequence categories that regulatory monitoring exists to track, and they are precisely the categories where a rule superseded two quarters ago causes real exposure. The stationary 90 percent of the corpus is not where compliance failures originate. And there is no way to tell, from inside a closed archive, which 10 percent has just moved without a fresh draw to compare against the old one — which is again a question about what keeps arriving, not about what is already held.

What continuous monitoring actually buys, and does not

A third objection deserves airing because it keeps the claim honest. A live feed of dockets, enforcement actions and rulemaking notices is not itself exchangeable. It is autocorrelated — one enforcement action often triggers a cluster of related filings — and selection-biased, since not every jurisdiction publishes with equal promptness, and it is open to being gamed, since parties time filings strategically. Replacing a static, wrongly-pooled archive with a noisy, ongoing stream is not obviously an upgrade; it is trading one violated assumption for another.

Accurate, and worth stating plainly rather than talking around. The stream does not deliver exchangeability. Nothing delivers exchangeability to legal data; the world genuinely changes, and pretending otherwise in either direction — pooling blindly or distrusting everything but the last filing — is unjustified. What continuous, provenanced intake delivers is the ability to run a change-point test at all. A rulemaking docket that keeps arriving lets a monitoring function apply something like a CUSUM statistic against the historical rate of amendments to a given rule, flagging the point at which a new filing breaks the pattern the archive had assumed. That test requires an ordered sequence still accumulating past the point where the archive was closed. A frozen corpus, however well partitioned into regimes, has no such sequence. It has a boundary, and beyond the boundary it has silence.

This is the narrower claim, and it is the one that holds. De Finetti's theorem does not say recent filings outrank old ones — most doctrine is stable and the historical record is what makes drift detectable as drift rather than noise. It says pooling requires a symmetry judgement with empirical content, and that judgement can only be tested against draws that keep coming, carrying their timestamp and their source, past the moment any static archive was sealed. A general counsel relying on such an archive is not wrong to trust the regimes it correctly identifies. The exposure sits at the edge, in the quarter the archive cannot see because nothing has arrived from it yet.