The Advantage Was Never the Model

Everyone assumes the leader in a domain has the better model. Read Harvey's own technical disclosures and the answer is duller and far more useful: it has the better room to put a model in.

Our own studies made this argument on small tasks with our own runs, which is a fair objection to them. Harvey makes it at enterprise scale, in a domain where a wrong answer is a disbarment event and a leak ends the firm. That is about as unforgiving a test of the claim as exists.

The headline number is the routing. Harvey reports that automatically routing work across models, rather than sending everything to the frontier, cuts inference cost by three to five times against a frontier-only baseline. Not by using a cheaper model. By knowing which questions need the expensive one.

Not a wrapper, an operating system

The dismissive reading of any applied-AI company is that it is a thin shell over someone else's model. The disclosed architecture does not look like that. It looks like an operating system: layers that decide what the model sees, which model sees it, what it is allowed to touch, and what happens to the answer before a human reads it.

Every one of those layers exists to remove a specific weakness, and none of them is a better model. The frontier model already has the legal reasoning. What it cannot supply is the firm's own conventions, its permissions, its record of what happened, and a routing decision made with the bill in mind.

The same finding, two scales

This is where it meets our own smallest result. Relevance, Not Volume found that a coding agent already has the engineering judgement, and that the only thing context can add is the narrow set of facts the model cannot derive on its own. Harvey's whole design agrees, and treats the frontier model exactly that way: as a capable component that needs a system around it rather than a system in itself.

What this is, and is not

It is a reconstruction from public disclosures, not a measurement we ran. The note is careful to separate what Harvey documents from what we infer, and this piece inherits that boundary: nothing here is a number we produced, and the three-to-five-times figure is Harvey's own report of its own system.

That is still worth having. It is the strongest available evidence that the environment thesis holds somewhere with real money and real liability on it, which is the one place our own harness cannot go.

The layer-by-layer reconstruction, the documented-versus-inferred split and what transfers to other domains are in the research note.