Doctrine
.
Diversive’s system for building software with AI — where a model is allowed to move at full speed, where an engineer has to sign, and what has to be true before anything ships.
Why we needed a system.
Building fast got cheap
AI made the slow part of software stop being slow, and everyone now has access to the same engines. Most of the industry took that speed the easy way — pointing models at the process they already had, and hoping.
Hope is the part nobody prices
Unsupervised speed doesn’t fail loudly. It skips what you can’t see from the demo — the security that was assumed, the architecture that was implied, the maintainability nobody asked for — and hands you finished-looking software. The bill arrives later, as the system nobody can change without breaking.
So we wrote the rules down
We wanted AI’s speed with our name on the result. So we made the rules enforceable — the same rules on every build, whoever the client, whatever the stack. Mettle is not a model, a wrapper, or a prompt library. It is the set of rules every build runs inside.
The three gates
Speed inside the lines. A signature on the rest.
Every build runs through the same three gates, in order — the engine opens the first, an engineer holds the second, and production waits on the third.
- Gate 01
Where the model runs free
Inside well-bounded work — implementation against a decided design, test coverage, migrations, the thousand mechanical edits — models run at full speed. This is where the speed comes from.
- Gate 02
Where an engineer signs
Architecture, data models, security boundaries — anything that decides how the system holds together — a named engineer owns. AI proposes. It never decides.
- Gate 03
What must be true to ship
Nothing reaches production unreviewed. Security and data handling are checked as conditions of shipping, not afterthoughts. Every release can answer for itself: what changed, who signed, what was verified.
Why it holds
It outlasts any model. And it shows its work.
Any model, one standard
Mettle is model-agnostic. Frontier models, open models, and the models we train in-house all work inside the same gates, to the same bar. When a better model arrives, we simply get faster — and nothing about the discipline moves.
Nothing taken on faith
You never take the system on faith — it shows in how the engagement feels. The plan is one an engineer wrote and will answer for. The progress is working software, reviewable at every step. And the questions that decide whether your product survives — security, data, what happens under load — are settled as conditions of shipping, while there is still time to act.
What comes out
Software that ships fast and reads like it was written carefully — because everything in it was either built inside a boundary or signed by someone with a name. Everything Diversive ships passes through this system: client builds and our own products alike.