Post
Same Definitions, Different Invoice
A software factory is a harness. A semantic layer is an ontology. Each pair is one job under two names, and you can tell because both make you do the same work before anything runs.
The data world settled the second pair in a group chat this week. The best line going around was that it’s a semantic layer if you built it, and an ontology if you’re raising.
The joke lands because the work doesn’t change. Both need entities, relationships, constraints, and the one correct way to compute revenue. Strip the logos and you’re writing down what things are, how they connect, and what’s permitted, in Markdown, SQL and YAML, like everyone else.
Factories and harnesses are the same story one layer up. A harness is the prompts, tools, permissions, context and loop around a model. A factory is that, plus orchestration and approvals, around many models. Read a factory platform’s setup checklist and you’re reading a harness spec with “team” written on top.
The strongest objection is scale and agency. A factory holds many harnesses, and an ontology can act where a semantic layer only reads. Both are true, and neither makes a new species. More harnesses need more of the same definitions. Write-back needs permitted actions, which are the same kind of definition as a permitted join.
There’s a sharper problem with factories and harnesses. Each one is a product roadmap for the models. Every step it codifies, like planning, testing, review, retries and routing, is a spec a model company can decompose and ship as a default. Building one as a standalone product means writing the requirements document for the companies that will absorb it.
That’s why the label is the part the model eats first. What survives is the narrow set of definitions only you can supply: the number finance certifies, the action that needs approval, the release that needs a safety plan. It’s the same cut I made with skills, and it’s why kit keeps getting smaller.
Define what a model can’t infer. Call it whatever raises the round.