Own your weights. In law, own the structure.

Sonya Huang recently told around eighty Sequoia portfolio founders to own their intelligence: not only the data and the product, but the models, down to the weights. A year ago that meant accepting worse performance. Open-weight models are now close enough to the frontier that you can start near it and tune past it, so Sequoia's position is that the trade has gone.

We think that is right, and that it stops one layer too high for anyone building into a regulated profession. Owning your weights answers whose intelligence is this. It does not answer the question a law firm has to answer first, which is can I sign this.

Checkability, not capability

A better model is a better product in consumer software. In regulated work the binding constraint is not capability. A family lawyer cannot file a court document because a model was confident. Every figure in a Form E or an ES2 schedule has to resolve to something a party swore or a bank statement records, and the fee earner has to see which before signing.

Fine-tuning does not produce that. Neither does owning the weights. It comes from structure.

There are no facts until a judge finds them

We work in financial remedy, the contested finances of divorce. Parties do not produce facts. They produce sworn assertions and evidence. Form E, the financial statement each party files, is signed under a statement of truth, and the two accounts routinely conflict. The conflict is the case.

The law treats even found facts as conditional on the honesty of disclosure, the financial material the rules require each party to produce. In 2015 the Supreme Court in Sharland and Gohil set aside concluded orders because that disclosure had been dishonest. If a court's own findings can fall that way, "facts" extracted by software from the same disclosure were never facts.

So the job is connection, not diagnosis. This is what the husband swore. This is what the wife swore. This is what the bank statement shows. This is what the rules required and never arrived. Every connection attributed, every assertion anchored to the line it came from. Judgement stays with the fee earner and the court.

Not your ontology, not your product

Inside Pivot that structure has a name: the Polyphonic Assertion Ontology, a model of the domain written down before the code. The assertion is the unit of record, not the document. Every voice is attributed, including the firm and the platform. Nothing is overwritten, so the record replays as it stood on the day the client signed. Every entity is computed as corroborated, uncorroborated, contradicted or unasserted, for each party separately.

The test a model cannot run

A reader, human or machine, can only find what the documents contain. Finding what is missing requires holding the list of what the rules required, and here the law writes that list down. Form E prescribes sixteen categories: twelve months of statements for every account, a pension valuation no more than twelve months old, the last three payslips and the P60, two years of business accounts. Because they are rules with numbers in them, whether disclosure satisfies them can be calculated rather than judged. That list is not in the weights and cannot be.

What we rent, what we own

We rent frontier models. They read a scanned statement or a badly completed Form E and propose structure. We own the ontology, the rulebook and the record. Every extraction enters as an attributed assertion with an author, a time and a source, so the model underneath can be replaced without the record changing shape.

This extends Sequoia's argument rather than disputing it. If the model layer is commoditising, differentiation moves down to what the model cannot supply: what the rules require, and what follows when it is not there.

Own your weights, by all means. But where the output has to be checked by a person and received by a court, the thing worth owning is the structure underneath.