We build calibrated instruments: to read what an agreement commits to, to assist people shaping what they write, and to measure honestly what it costs to build the instruments themselves. For each one we state plainly what it does, and what it does not claim.
Say them back in and you have the idea the company is built on.
Convexity is Nassim Taleb's word for a useful asymmetry: positions where being wrong costs little and being right pays large. Its rarer cousin, antifragility, names the things that don't merely survive stress but improve under it. The founder taught these ideas in classrooms, advised organizations on them, and started this company to practise them — in the one place almost nobody applies them: writing.
Every draft tested against an instrument you don't control — before the board reads it, before the counterparty's lawyer, before the regulator — moves the failure to where it costs least. And writing that is tested, revised and tested again does the rarer thing: it gains from the pushback. The scrutiny a writer fears becomes the reason the work holds.
The instruments stay fixed; that is what makes them instruments. The gain compounds somewhere better — in the writer, and in the organization that uses them.
An agreement reads differently depending on which side of it you are standing on, and a legal review is organised to represent one of those sides, because that is the job. Our products read the same document from each position and report where the two readings part company. That difference is the finding.
This is not a substitute for your lawyer. It is a different reading — done before the lawyer's, and done again after you have changed your draft.
Reads the counterparty's agreement, from both sides at once: theirs and yours. It reports where the two readings differ, and where accountability is structurally absent rather than merely unstated.
Everything Scrutineer does, and then the other half. You draft your reply inside the product, and the draft goes back to be read again. That loop — read theirs, shape yours, read it again — is the product.
The same loop across a three-party agreement: the three relationships — A–B, A–C, B–C — read separately, then brought together. No party in such a deal sees the whole structure, and no lawyer in it is meant to.
How they build on each other. Each one contains the one before it. Scrutineer reads. Procurator reads, assists you shaping your reply, and reads again. Impresario does that across three parties at once. There is no add-on and no withheld layer: what a product does, it does entirely.
Deckle™ is the writing half on its own, for when you are drafting agreements, communications, directives, letters rather than reading others. Deckle is not an LLM — it is a calibrated interface to one, tuned for a single job: assisting you honing and shaping your writes and rewrites to your own decisions and reasons. Every AI writing tool writes for you. Deckle works with you.
You cannot adjust the reading to make your own document read better. Neither can we, for ourselves. An instrument whose reading you could adjust is one you could game.
It will not fake precision it does not have. On a close call it tells you the call is close, rather than manufacture a decisive-looking answer.
Nothing here is ungameable. What we will say is narrower, and true: a document withstands an instrument we do not control, or it doesn't.
What happens to a document you put into one of these is a human-directed exercise, not an AI-generated one.
Prose designed to be undetectable is, in that design, detectable. You are better to concede than be caught seamless.
Where this is going. The products above are being built now, and this site is being rebuilt with them. What is here is what we are prepared to say; the rest waits until there is something to show.