Lexora

Our own product, and the argument behind it

Lexora is what we build when there is no client brief: an AI research and document platform for litigation teams. We designed the system, built the product interface and the marketing site in code, and shipped it.

The category has a habit that we think is the whole problem. Legal AI products answer confidently, attach a percentage, and leave the lawyer to find out later whether the case they relied on is still good law. The confidence number is the tell. It looks like rigour and it is the opposite, because it compresses two entirely different risks into one figure.

Two axes, never one score

Lexora carries no confidence percentage anywhere. It carries two independent axes instead, because they fail independently.

The first is grounding: is this proposition quoted from the passage, supported by it, weakly supported, or unverified. The second is treatment: is the authority good law, questioned, superseded or overruled. A case can be perfectly quoted and thoroughly overruled. One number cannot say that, and two badges can.

Silence is the reward. A citation that is quoted and good law shows no badge at all, so the marks a lawyer does see are the ones that need reading.

The plan runs before the search does

Every question shows its search plan before anything runs, and the plan is editable. The jurisdictions, the year range, whether the matter’s own documents are included: all of it is visible and switchable, and the run reports what it touched.

This is slower to demonstrate and much easier to defend. A partner asking why an argument rests on a Ninth Circuit case can see that the run was scoped to California and the Ninth Circuit, screened 214 results, kept 31, and checked treatment on every one.

A finding is the unit, and a person accepts it

The object that matters is not the answer. It is the finding: a proposition, the citations that carry it, and the name of the person who accepted it.

Nothing becomes a finding on the product’s own initiative. The accept gate is the product. Once a finding exists, treatment watch keeps checking the authority underneath it, and tells the person who accepted it when the ground moves. That is the feature the whole trust model exists to make possible.

Designed as a system, then built and shipped

The design system came first: 59 icons drawn to one grid, 36 component sets, and a citation and trust layer where the badges are components rather than decisions made per screen. Thirty four product screens were designed on top of it, desktop and phone, plus a dark theme that resolves entirely through variables.

The marketing site was then built in code and verified against the design rather than eyeballed. Sixteen pages, measured section by section at two breakpoints, and every one of the thirty two checks landed exactly. The verification earned its keep: it found a letter spacing rule leaking out of the product into the site, quietly shortening every measure.

This is the shape of every Confidency project, run without a client brief. We design the system, we build it, and we ship it. Lexora is live, and the whole product is walkable.