Free energy and dissipative structures in banking compliance

The gradient that would not sit still

Lars Onsager published his reciprocal relations in 1931, and they read like bookkeeping: a set of symmetries governing how heat, matter and electric current cross-influence each other's flow. Unglamorous work, awarded the Nobel Prize in Chemistry decades later, in 1968. But it opened a question that classical thermodynamics had no vocabulary for: what happens to a system held away from equilibrium, permanently, by a steady supply of something usable?

Erwin Schrödinger asked a version of the same question in 1944, in lectures aimed at biologists. Why does an organism not simply run down, as the second law seems to demand? His answer was that it feeds on "negative entropy" — it exports disorder faster than it accumulates it, and that export is the entire trick of staying alive. Ilya Prigogine, working in Brussels from the 1940s onward, turned the intuition into a formal theory. Systems driven far enough from equilibrium can organise spontaneously into stable patterns — convection cells, chemical oscillators, flames — provided free energy keeps arriving and entropy keeps leaving. He called them dissipative structures. The order in them is never stored. It is rented, continuously, from a flux. Stop the flux and the structure does not decay gently. It stops.

That is the whole idea, and it turns out to describe a compliance floor rather precisely.

The rule that ran for a quarter

A screening rule sits inside a bank's transaction monitoring system, built to flag payments touching a sanctioned entity. It is deployed against a sanctions list current as of the day it was tuned. The list, meanwhile, is updated most business days — new designations, delistings, spelling variants, aliases added after an enforcement action somewhere else in the world. The rule, once shipped, does not know this. It runs against a frozen snapshot for a financial quarter before anyone revisits it, because revisiting it costs analyst hours and change-control sign-off, and there is always a more urgent alert queue.

For most of that quarter the rule looks like it is working. Throughput is normal. False positives sit in the expected band. Nothing in the dashboard says stale. Then an entity added to a sanctions list in week two clears the bank's payment rails in week nine, seven weeks after it should have been blocked, because the rule was never told the list had moved. The compliance officer who signed off on the rule's last tuning cycle is now the one explaining to a regulator why a designated party was serviced. The rule did not fail by malfunctioning. It failed by sitting still while its environment did not.

This is the crystal problem, and it recurs everywhere intake is treated as a one-off cost rather than a running one. Adverse-media feeds go stale the same way — a name cleared of negative coverage eighteen months ago may have acquired a fresh enforcement action last week, and a static risk score never finds out. Rule changes issued by a regulator arrive as documents, not as automatic amendments to logic; someone has to notice, interpret, and rebuild. The failure mode has one shape across all three: an artefact was tuned once, against a moving target, and nobody built a mechanism to keep paying for the tuning.

Order that is rented, not owned

Read thermodynamically, the screening rule is a dissipative structure that stopped dissipating. Its apparent order — the ability to distinguish clean transactions from sanctioned ones — was never a fixed property. It was a pattern maintained by continuous import of list updates. When the import stopped, the order did not vanish instantly, which is the dangerous part. It looked identical for weeks while becoming progressively wrong, because nothing internal to the rule marked the gap between the list it held and the list that existed.

Contrast a monitoring system that treats the sanctions list, the adverse-media feed and the regulatory rule-change stream as flux to be continuously metabolised rather than data to be loaded once. Every new designation is ingested the day it is published. Every alias variant is folded in. Every rule amendment triggers an automatic diff against the logic currently in production, flagging where the deployed rule and the current regulation diverge. That system is not more intelligent than the static rule. It is differently maintained. Its accuracy is not a stored asset; it is an ongoing expenditure, paid in engineering hours, feed subscriptions and reconciliation cycles, the compliance equivalent of the roughly 100 watts a resting body spends just holding its ion gradients steady against thermal decay. Standing still, for a screening system, is also a full-time job.

The third piece — export — is the one compliance functions chronically underbuild. A system that ingests every sanctions update, every adverse-media hit and every rule change but never retracts a stale flag, never reconciles a superseded alias, never timestamps when a belief about a customer's risk status was last confirmed, accumulates contradictions instead of currency. Analysts inherit case files with three generations of unresolved alerts on the same entity, no record of which is authoritative. That is the dissipative-structure failure in its other direction: import without export produces noise dressed as diligence, and audit trails that cannot say what the bank currently believes, only everything it has ever been told.

Why "monitor harder" is not the objection it sounds like

The thermodynamic framing is decorative. A data centre already burns megawatts running these screening pipelines. Calling a static rule a "crystal" confuses joules with information — everything computational dissipates energy, so the metaphor stops distinguishing anything.

Correct about the electricity bill, and the distinction survives anyway because it was never about joules. The relevant flux for a compliance system is informational free energy: the availability of observations capable of changing what the system believes about a customer or a transaction. A rule frozen for a quarter draws plenty of power running its comparisons. It imports zero updated availability about the sanctions list. Its internal state is stationary no matter how many watts it burns doing pattern-matching against a snapshot. A live-fed system draws on a different reservoir — the actual rate at which the world outside the bank is changing — and that is the reservoir that determines whether the pattern it holds still corresponds to anything real. Two systems can have identical power draw and opposite thermodynamic status, in the sense that matters here.

Dissipative structures are dumb. A convection cell has no memory and no error correction; it re-forms identically given the same conditions. Borrowing its persistence licenses nothing about provenance, retraction, or the epistemic apparatus a compliance system actually needs.

Granted without reservation. Nobody is proposing that a bank's watchlist behaves like a Bénard cell, and the analogy is narrower than it might sound: not that belief maintenance is convection, but that anything maintained against a decay process needs both a supply side and a discard side, and that the discard side is the one everyone forgets. Prigogine's actual contribution was insisting entropy export is not optional — a structure that only accumulates does not stay ordered, it silently becomes a landfill with a user interface. That is a fair description of a case-management system that has never retracted a flag in four years. The physics establishes that the ledger needs two columns. It says nothing about what provenance metadata should look like, what a retraction workflow should be, or who signs off on a false positive being cleared. Those are compliance's problems to solve, not physics'.

Where the lineage lands

A screening system built once against a fixed list is a Large Language Model in miniature: order fixed at tuning time, degrading in silence as its environment moves on. A system that reasons well over the transactions visible in today's batch but retains nothing about yesterday's sanctioned-entity near-miss once the batch closes is the Large World Model case — coherent within the window, blank outside it. What the domain is visibly reaching for, in every bank that has been fined for exactly this failure, is the third regime: transaction flow, sanctions lists, adverse-media feeds and rule changes treated as streams that never stop, beliefs about any given customer held as provisional and dated, contradictions surfaced rather than buried, and retraction built in as a first-class operation rather than an afterthought.

That is not a bigger model. It is a different thermodynamic commitment — continuous payment instead of a single upfront one — and it is the last such commitment available, because there is no fourth way to relate to a stream. You either freeze it, ride it while it lasts, or metabolise it without end. Banking compliance has already discovered, at cost, which of the three the sanctions list actually is.