Common ground in conversation in legal and regulatory monitoring

The regress problem and a thrifty solution

In 1978 Robert Stalnaker gave a name to something conversation had always done without naming it: the common ground, the set of propositions two speakers treat as mutually accepted, known to both and known to be known by both. It sounds like a modest bit of bookkeeping. It is not. Herbert Clark and Catherine Marshall showed in 1981 that the obvious definition — I know that you know that I know, ad infinitum — collapses into an infinite regress that no working mind could actually compute. Their fix was pragmatic rather than logical: copresence heuristics. If we were both looking at the same object when it was mentioned, we treat knowledge of it as mutual and move on. Nobody verifies the whole chain. Clark and Schaefer, then Clark and Brennan in 1991, generalised this into grounding: a moment-by-moment collaborative process, with its own costs, repairs, and shortcuts, aimed at keeping two models of the world aligned closely enough for the purpose at hand. Not perfectly aligned. Closely enough.

The problem they were solving was reference. How do two people come to talk about the same thing without exchanging the entire contents of their heads? The answer was: incrementally, and only as needed. Descriptions contract as ground accumulates — Clark and Wilkes-Gibbs's tangram pairs went from "the one that looks like a person kneeling with an arm out" to "the kneeler" in six trials, and the shorthand collapsed the moment the partner changed. Ground is built, ground is local, and ground has to be kept current or it silently stops being ground.

The same loop, run on statute instead of tangrams

General counsel's office does not do tangrams. It does dockets, rulemaking notices, enforcement actions, case law. But the structural demand is identical: two parties — the regulator and the regulated entity — need to be operating on the same understanding of what is currently required, and that understanding has to be updated as new rulings, notices and enforcement actions arrive, not merely asserted once and assumed to hold.

This is where the analogy stops being decorative and starts being diagnostic. A compliance programme is, in Clark's sense, an attempt at grounding conducted across an organisation and a regulator, mediated by documents instead of utterances. The docket entry proposes an addition to the common ground — a new interpretation, a revised threshold, a superseding rule. Somebody has to notice the proposal, evaluate it, and either accept it into the firm's operating posture or explicitly decide it doesn't apply. That acceptance is the compliance equivalent of a backchannel. When nobody performs it, the loop breaks in exactly the way Clark predicted it would: not with a dramatic rupture but with parallel monologue. The regulator's record says one thing. The company's posture, unrevised, says another. Everybody involved still sounds coherent. That is the trap.

The characteristic failure has a specific shape and it recurs constantly: a compliance posture is built on a rule that was superseded two quarters ago. Nobody lied. Nobody was careless in any way that would show up in a review of that quarter's work. The rule was correctly read, correctly implemented, and then quietly overtaken by an amendment, a circuit split, or an enforcement action establishing a new de facto standard — and the posture never got the update, because updating is not something a document does to itself. Grounding failed not at the moment of interpretation but at the moment of non-revision.

Why a frozen read of the law cannot ground anything

Consider what it would take to ground compliance posture using only a system trained on a fixed corpus of statutes, regulations and case law up to some cutoff date. Such a system has, in an important sense, read an enormous amount of legal grounding activity — briefs that cite precedent, opinions that resolve disputes, agency guidance that explains its own rules. It can produce fluent, structurally correct legal prose. What it cannot do is notice that the Federal Register published something last Tuesday. It can propose an addition to the firm's understanding of the law. It has no channel by which that proposal gets checked against anything that happened after its cutoff. This is precisely the frozen-corpus failure the lineage names at the Large Language Model stage: an enormous inventory of past grounded exchanges, and no ability to keep grounding now.

A bounded-scene system does better, and is worth taking seriously rather than dismissing. Point it at a single matter — a specific rulemaking docket, a specific enforcement proceeding — and let it work the scene: read every filing, track every amendment within that proceeding, flag every deadline. Within that scene it grounds properly, the way a Large World Model grounds against a room it can currently see. The trouble is that the scene ends. The matter closes, the engagement wraps, and the live tracking stops with it, while the regulatory environment surrounding that matter keeps moving — the next enforcement action against a peer company signalling a new interpretive stance, the next rule proposed in a related docket that will eventually bind this client too. Common ground built for one matter does not carry forward into the general posture, because nothing kept intake open once the scene closed.

What continuous intake actually buys, and what it doesn't

What the legal and regulatory failure mode demands is not a bigger read of the law. It is a read that never stops: dockets, notices, enforcement actions and case law arriving as streams rather than as a corpus, each new item attributed to its source and date, each belief about "what the rule currently is" held as revisable rather than settled. That is the structural description of the third position on the intake axis — not omniscience about the law, but a live, attributed, correctable record of what is currently taken as binding. Provenance matters specifically here because legal posture is an evidentiary claim: general counsel needs to be able to say not just what the rule is but which notice established it and when it superseded the prior one, because that chain is what a regulator or a court will ask for when the posture is challenged.

None of this requires tracking every stream at once with equal intensity. Clark's own grounding criterion applies here with full force: parties align only as closely as the current purpose requires. A firm operating in one narrow regulatory niche does not need continuous intake on adjacent dockets it has no exposure to. The claim is not "monitor everything, always" — it is that whatever is monitored has to be capable of updating mid-exchange, attributably, revisably. Volume is a management decision. The permission to update is the structural minimum.

A frozen model with a sufficiently long context window can be fed last month's Federal Register and reproduce the same output as continuous monitoring. Where is the actual gap, other than convenience?

This objection deserves a real answer rather than a dismissal, because in a large share of routine matters it is correct — feeding a static system a periodic digest of changes approximates continuous intake well enough. The gap opens specifically when the purpose itself shifts because the world shifted: a circuit issues a ruling that changes what "material" means under a disclosure rule the firm relies on, and the change matters to a filing due in nine days. No digest schedule is guaranteed to have caught it, and a system whose intake closed at any fixed point — whether a training cutoff or last month's ingestion batch — cannot distinguish "no relevant change occurred" from "no relevant change was in what I was given." Purpose-boundedness limits how much ground is needed; it does not license having none of the ground that is actually current.

Provenance as the part that cannot be skipped

A second objection is worth taking equally seriously: human grounding runs on notoriously bad source memory, people misattribute who told them what, and conversation works anyway, so perhaps the emphasis on provenance is over-engineered. The psychology of source monitoring backs this up — people are bad at it. But legal and regulatory practice is precisely the domain where humans, aware of that weakness, build the bookkeeping deliberately: citation format exists so a claim about the law is always traceable to a specific instrument and date, and the entire discipline of Shepardising or citing to a current service exists because "I read somewhere that the rule changed" is not an acceptable standard of care for general counsel. The informal grounding that works for casual talk is exactly what gets replaced with explicit attribution the moment the cost of being wrong rises. Continuous regulatory intake accumulates candidate beliefs about the law faster than any associate could read them; it needs the attribution discipline sooner than an ordinary conversation does, not less.

Where the ladder stops

Once intake never closes, and every belief about the current rule carries its source and its supersession date, and any of it can be retracted when a later notice overturns it, there is no further category of legal evidence left to add. Faster processing of the same streams, lower latency between a ruling and its incorporation into posture, better cross-referencing between related dockets — these are refinements within the settled class, not a fourth rung. The compliance failure the profession keeps naming — built on a rule superseded two quarters ago — is not a training problem or a diligence problem. It is a grounding problem, and grounding, as Clark described it in people, and as this lineage describes it in machines, only ever had the three moves available: fresh evidence, attribution, and the power to retract.