The Computational Turn
The score era is ending.
For two decades, compliance ran one way: screen, score risk, route to a human for investigation, escalate to another human, file the report by hand. It barely worked in the bank-rail age. It breaks when money moves at internet speed.
Look at the bills. In October 2024, TD Bank paid more than $3 billion in fines (from the DOJ, FinCEN, the OCC, and the Federal Reserve) for what regulators described as a near-complete failure of transaction monitoring. Seventy thousand alerts sat in the backlog, by a16z's count. A year earlier, Binance settled with the DOJ, FinCEN, and OFAC for $4.3 billion for inadequate transaction monitoring and failure to file Suspicious Activity Reports. Two sides of the rails and two different failures, one overwhelmed, one willful. The same exposure underneath: neither could evidence what moved through it.
Neither bank lacked people; both employed thousands of compliance staff. TD's failure was chosen. It capped compliance spend flat and left most volume unmonitored, and that choice is the tell: when the architecture uses humans as the computers, coverage scales linearly with cost against a problem that grows exponentially, until honest monitoring becomes a bill someone decides not to pay.
That constraint just broke. Intelligence is now a service: add compute and you get more of it. The fix is Computational Compliance, systems built on compute and code that do the investigation and produce defensible evidence at machine speed, with a human judging every result.
a16z is right that everything, everywhere is compliance. On-chain goes first, for three reasons.
The first is public data. Transactions are public, structured, and machine-readable by default. No legal agreements with the institution holding the data. No PDF scraping. No batch reconciliations. The ledger's record sits there in the open, complete for everything the chain itself saw.
The second is pressure. The adversary is outpacing the hiring pipeline and is already using AI to discover smart contract exploits and generate obfuscation patterns. The attacker's deployment isn't gated by compliance review, audit, or model risk management.
The third is the question itself. Regulators are no longer asking, does this look risky? They are asking, can you prove it? And if you can't, you pay billions in fines.
The Understanding Stack
Open data gets you started; it does not get you understanding. Understanding is built in three layers, and only the first is open by default.
The first layer is the raw data: transactions, contracts, addresses. On-chain it is one shared record of the whole network, machine-readable by anyone. TradFi has the same transaction-record layer too, but sliced: each institution sees only its own accounts, and the full picture exists nowhere. That is why on-chain goes first, and why TD Bank could drown while holding all of its own data.
The second layer is everything that doesn't live on the ledger: labels, attribution, KYC cooperation, OSINT, law-enforcement intelligence. The label that ties a Lazarus subcluster to a state actor, an unregistered OTC desk to its operator, a sanctioned obfuscation service to the entity running it. None of that is on-chain. It takes field agents, partnerships, and years of collection. In TradFi, this is the KYC/KYB file and the customer profile. This layer is proprietary, and it is where most of the industry competes today, on label volume.
Volume alone is the wrong metric. What converts to findings is intel that is unique, high-value, and fresh. That means the label competitors don't have, the attribution that changes a decision, the freshness that outlasts the adversary's next adaptation. Layer 2 produces inputs. Findings come from Layer 3.
The third layer is the understanding model: the reasoning that works across the first two layers to produce a finding. Behavioral signatures, heuristic rule sets, threat-actor playbooks, model inference. This layer turns data into something insightful and defensible. It is also the hardest to build: it takes deep domain knowledge to encode and rigorous model evaluation to trust. And unlike labels, it compounds in use. Every case an analyst adjudicates feeds the next one. Incumbents hold more alerts, but alerts don't teach; adjudicated cases do, and a product that sells scores never sees the adjudication. That is the moat. Computational Compliance is the work of building this layer at scale.
At the center of Layer 3 sits the World Model, an encoded comprehension of how on-chain finance actually behaves under compliance pressure. Concretely, it holds the actors and their relationships, the transaction primitives, the adversary playbooks, and the feedback from every adjudicated case. It tracks what is happening on a chain, not what a transaction looks like on its face. Competitors can copy APIs and tools; they cannot replicate years of operational learning. And when the adversary changes technique, each incident is a delta against an existing map, not a rebuild.
Each layer has different economics. Raw data is a commodity. Verified intel is scarce, but it can be collected. The binding constraint is the reasoning: one ground-truth seed is highly informative if you can cluster outward from it, and the investigators who can are rare, expensive, and do not scale with case volume. The industry has spent a decade trying to hire its way out. Computational Compliance puts machine reasoning behind the scarce skill. In our casework for law-enforcement and industry customers, one good seed yields a hundred defensible conclusions where it used to yield ten. With a frontier model, an issue that took an analyst an hour to resolve now closes in five minutes for about 10-20% of the cost, plus a minute or two of human review. It is cheaper to validate good work than to produce it. And this is just the beginning: model capability is still climbing and model cost is still falling, so the curve only moves one way.
The Proof Standard
Acceptance requires two things: a defensible result and a defensible process. The finding has to stand on its own, able to survive a challenge in court, in a regulator's exam, in an audit. The method behind it has to hold up too: repeatable, documented, validated. Banks have run that discipline over their monitoring and sanctions models for years and call it model risk management; it now has to cover the reasoning system that does the investigation.
Computational Compliance meets the standard by construction: every finding carries its sources, every step leaves a replayable trace, and the methodology is versioned and validated the way banks validate models today. None of that lets a vendor declare its own output court-ready; examiners and courts decide that. What a builder controls is whether every finding arrives with its receipts.
In operation, the agent does the casework: it triages the alert, gathers and cites the evidence, drafts the SAR, assembles the case. A human approves, escalates, or rejects it. The outputs are the ones compliance teams already file (cases, SARs, dispositions), and the number to watch is first-pass acceptance: in our internal testing, agent casework passes first review 60% of the time. Humans at the gate, not in the queue. The role split gets its own article later in this series.
Regulation is moving in the same direction. MiCA is live in Europe, the GENIUS Act is law in the US with the CLARITY Act close behind, and FATF is auditing Travel Rule implementation worldwide. None of them mandates agentic compliance. Every one raises the bar on auditable execution, and that bar keeps getting harder to meet with humans as the computers.
The Evidence Era
The budget is already there. Two decades of automation never cut financial-crime-compliance spend; the money changes hands. As agentic execution replaces headcount, the same dollars stop paying for queues and start paying for closed cases.
By 2030, this will be compliance's operating architecture. The score era closes. The evidence era begins.
This is what we are building. We don't give you a score. We give you the closed case: the investigation done, the evidence assembled, closed by your analyst's signature.
CipherOwl builds Computational Security and Compliance infrastructure for on-chain finance. Read Part 1, The Bitter Lesson of On-Chain Truth Seeking, for the foundation.