Presented at AI Engineer.

Finance used to be bottlenecked on producing content. It is now bottlenecked on verifying it. An analyst who once spent the day building a model now spends it checking whether a machine’s output can be defended to a client, a committee, or a regulator.

Citations don’t solve this. A citation is a post-hoc audit trail — it tells you where a number allegedly came from, not that the number is right. This 22-minute talk covers the architecture we built at Kepler instead, and the argument for why it has to be architectural.

Evals Are Not Verification

The industry’s default answer is to add evaluation layers. That helps, but it doesn’t change the category of the thing you’re holding: you cannot take a non-deterministic LLM and eval your way to something deterministic. Stacking probabilistic checks on a probabilistic system gets you better odds, not proof.

The alternative is to separate reasoning from fact-production by design.

Three Pillars

Atomic provenance. The model never manipulates numbers. It identifies sources and delegates extraction to deterministic tools, and anything that can’t be confirmed against a source gets stripped. A model that never touches a numeral cannot hallucinate one.

Scope determinism. The model decides what to compute; it never does the computing. Every operation routes through deterministic code — database queries, XBRL parsing, financial algorithms. There’s no good reason for basic arithmetic to run through a billion-parameter model instead of a CPU.

Derivation chains. When a metric isn’t stated in a filing and has to be derived, the system records the full lineage from inputs to output. Any number can be replayed, re-verified, and audited — which also lets firm-specific definitions travel across desks instead of living in one analyst’s spreadsheet.

Why Finance Specifically

Other high-stakes fields layer their verification. Software has tests and staged rollouts; aviation has checklists and redundant systems; emergency medicine has protocol and second reads. Finance frequently has one overworked analyst.

Bloomberg and FactSet partly solved this socially — when everyone uses the same data, culpability is distributed. But they’re read-only. The work product still gets assembled by hand. Closing that gap without giving up verifiability is the problem we set out to solve.

Questions about verifiable AI or the architecture behind Kepler? Feel free to reach out.