PYX data product agentic engineering
Stage 02 · Model

Argue with a draft, not with a blank page

Discovery reads your estate and proposes concepts, grain and join paths, each binding carrying a confidence and a stated reason. Modelling becomes a review, and the review is where the expertise goes.

How this worked before

A modelling workshop where the people who know the answer cannot see the consequence, and the people who can see the consequence do not know the answer.

How it works here

A page of evidence-linked proposals with the consequence of each one computed, and a person ruling on them one at a time.

The single decision that determines every number downstream is grain, and it is almost never made on purpose. An engineer needs a key to join on, picks the one that makes the join work, and moves on. There is no artifact recording that as a choice, so there is nothing to disagree with later, only a number to disagree with, which is a far worse argument to have.

So make it a decision. A model can propose the candidate grains a requirement could sit at, name the fan-out between them, and state what each choice does to the measure. It cannot know that your regulator counts affected customers at the business-partner level because that is what the licence condition says. That part is judgement, and judgement is what the analyst is for.

Concepts bind late, and that is structural rather than stylistic. A measure names a concept, not a physical relation, so when the table moves, and in a migration it always moves: the meaning does not. The binding is a separate, reviewable, dated fact with a diff of its own.

Every inference carries a because

A proposed binding states its confidence and the evidence it came from. A proposal that cannot point at its evidence is an invention, and it gets refused rather than quietly accepted.

The join decision is a gate, not a footnote

When two concepts meet at different grains, somebody states the path and the cardinality they expect, and that statement blocks the save. Most silent double-counting in a warehouse is an unstated many-to-many nobody was asked about.

Coverage is a live score

“This model answers 34 of the 41 claims” is a more honest completeness measure than a signed-off document, and it names the seven arguments you still have to have.

Demo · click through it

This stage, running

pyx · estate · impact of a change
Proposed change

Drop erch.zzbillcycle from the silver load

An engineer opens a PR. In most estates the blast radius is discovered afterwards, by whoever gets the angry email.

3 models touch it1 governed control2 published reports
A hold refuses. A dependent warns. The difference is load-bearing.