The AI reads. The code counts.

The Clerk — reads your documents

The Clerk is the AI doing the one job it does best: reading. It goes through each document and transcribes what that document actually states — the financials, the premises, the workforce, the systems — into a structured record, keeping the document's own words as proof.

What makes the Clerk different is what it refuses to do:

Blanks stay blank. If a document doesn't address something, the record says so. The gap is never filled with a guess about what "a business like this usually" looks like — a quiet listing is made to look neither better nor worse than one that spelled everything out.

A refusal to answer is an answer. When a listing literally prints "Real Estate: Not Disclosed," that goes into the record as exactly what it is: the seller was asked and declined to say. That's a fact about the deal, and Verity treats it as one.

Disagreements are recorded, never resolved. When one page says one thing and another page says something else, the Clerk writes down both. It doesn't pick a winner — the disagreement is exactly what you need to see.

Reading is free. Upload a deal and see what your documents say, in the documents' own words, at no cost.

The Guard — runs the checks (software code, not AI)

Your open items and contradictions come from the Guard: a fixed set of checks written in software code — run identically on every deal, every time. The Guard can't be persuaded and doesn't decide what "seems important." It checks what it always checks.

Some checks catch what's missing: a claim with nothing behind it, a question the documents never answer. Some catch what conflicts: two documents stating equipment value differently — and when that happens, the Guard keeps both figures and shows them side by side, naming which document said what, instead of quietly picking one.

Because the Guard is code, its behavior is a measured fact rather than a promise: the same documents produce the same findings, every time. We prove that in production — re-running an unchanged deal produces the identical set — instead of asserting it on a slide.

The Guard's honesty runs all the way down:

What Verity checks is exactly what's written in the Guard's code. Nothing on your screen is counted because an AI decided it was probably worth counting — and Verity never claims a check it hasn't built.

It fails honestly. If an analysis can't be completed and verified, nothing is stored and nothing is charged. Better to say "we can't verify this right now" than to show you a confident wrong list.

The Judge — explains the deal

With the count already settled, the AI comes back for the second job models are genuinely good at: making sense of things. The Judge reads the same record the Guard checked and writes the analysis you actually read — context, patterns, what a figure means for a deal like this one.

The Judge can also point out things worth a look — the kind of concern a sharp analyst raises that isn't yet one of the Guard's fixed checks. Those observations appear clearly set apart from the count, and you decide whether to add them to your open items.

But the Judge has one hard limit, and it's written into the software: it can never add to, remove from, or reword the count. Not one item. The commentary is the AI's. The count is not.

How the Guard learns — without ever guessing

The three parts aren't frozen. When the Judge keeps noticing the same kind of issue — and buyers keep adding it to their open items — that concern becomes a candidate to join the Guard. It gets written as software, tested against the same-documents-same-findings bar, and only then enters the count.

And it works in both directions. A check that can't meet that bar doesn't get to stay in the count on charm — it moves back to the Judge's commentary, where its promise matches what it can keep. The count only ever contains what fires identically every time.

What you pay for

Free: the Clerk's reading — what your documents state, with their own words as proof.

Paid: the analysis — the Guard's count of open items and contradictions, figures side by side, plus the Judge's narrative and observations. And because the count is code, an unchanged deal re-analyzed gives you the identical set. You're paying for findings you can act on, not for an AI's mood that day.

The honest limits

Determinism is not a promise of completeness or truth. The Guard checks what's written into its code; documents can be wrong in ways no software catches; and a clean count is not a clean deal. Verity tells you what the documents state, what they fail to state, and where they disagree — it does not certify that the statements are true. That's what diligence, and the professionals you hire for it, are for. Think of Verity as the analytical half of a quality-of-earnings review, done before you decide whether the deal deserves the other half.

The Clerk and the Judge are the AI. They read and they comment. But the count comes from the Guard — and the Guard is software code, not judgment.