12 Jun 2026

Jane Street just reversed a 25-year position on formal methods, and the reason matters beyond quantitative trading.

The argument is straightforward: AI agents generate code fast, but they produce complexity faster than human reviewers can audit it. Subtle bugs, missed invariants, edge cases that only surface under unusual conditions. Formal verification was always theoretically attractive but practically expensive. Agentic coding has shifted that equation by making the problem more expensive, not by making the solution cheaper.

For anyone building loan origination systems or credit decisioning platforms, this should land hard. We already operate under a regulatory obligation to explain and audit automated decisions. The FCA's Consumer Duty doesn't care that your AI wrote unverifiable code at impressive speed. If your system applies an interest rate, modifies a repayment schedule, or declines a customer, you need to be able to demonstrate that the logic is correct and that edge cases were considered.

Most teams in UK consumer finance are nowhere near formal verification. They're still debating whether to let developers use Copilot at all.

That's the wrong conversation to be having in 2025. The question worth asking is what your verification strategy looks like when AI is generating a meaningful share of your production code. Manual code review does not scale to the output velocity of agentic systems. Testing coverage helps, but tests written by the same agent that wrote the code have an obvious circularity problem.

Jane Street's move signals that serious engineering organisations are treating this as an architectural question, not a tooling preference. The cost-benefit tradeoff on formal methods has genuinely shifted.

Consumer finance leaders should be asking their engineering teams what the equivalent looks like for their stack. Not necessarily full formal verification, but some honest reckoning with how you audit code you didn't really write.

Maki Yamaguchi builds portraits by holding two opposing things together: fine, precise detail and bold, spontaneous brushstrokes. The result works because neither overwhelms the other. That tension is the point.

This is exactly the problem most loan origination platforms fail to solve.

We spend enormous effort on the detailed, realistic layer: decisioning logic, affordability calculations, credit policy rules, FCA compliance controls. That work is necessary. But the gestural, human layer gets treated as a UX afterthought, something the product team handles once the engine is built. The customer journey ends up feeling like what it is: a compliance checklist with a thin coat of paint.

The better approach is to design both layers simultaneously, with equal intent. What does the application feel like at the moment someone is nervous about their credit score? What does a decline communicate, and what does it leave the customer believing about themselves? These are not soft questions. Under the Consumer Duty, the FCA is explicitly asking whether outcomes are good for the customer, not just whether the process was followed correctly.

Yamaguchi's work also points at something about restraint. She does not fill every corner. The negative space is part of the composition.

In technology builds, we do the opposite. We add features, add fields, add friction, because each stakeholder wants their requirement represented. The result is origination journeys that convert badly and leave customers exhausted.

The question worth sitting with: when did you last remove something from your customer journey rather than add to it?