How T2D3 OS gets better every month

Under governance

Not a prompt folder. Every instruction the AI runs is a catalogued artifact with a version history, an owner and a rollback.

Tests we are not allowed to change

This is the part that makes the rest of the page mean something. A frozen set of evaluation cases sits outside the improvement loop — no optimizer, automated or human, may edit them. It is how our scores went up can be told apart from we rewrote the exam.

We publish the method and the count, never the cases themselves — a published exam is one you can study for.

Promises enforced by the build

These are not intentions. Each is a check that fails our build, so it cannot be skipped on a busy week.

• A new model means a new review

  When we adopt a frontier model, our build stays red until every prompt routed to that tier has been re-reviewed. Prompts tuned for the old model do not silently inherit the new one.

• Specifications must match the code

  Each guided skill declares what it reads, writes and calls. A gate compares that against what the code actually does, and the allowed gap only ever shrinks. Today 43 of 43 align.

• Routed by fit, never down-tiered to save money

  Each task runs on the model that suits it — 12 models across 6 providers in the last 90 days. Falling token prices buy you better work, not a cheaper bill for us.

What your edits teach it

When you change an AI draft before locking it, that edit is the signal. It is distilled into durable guidance and injected into future generations for that module — so the system converges on how you work.