A platform that prefers inference to interrogation

Three readers

Every site TITI generates is written for more than one reader, and the structure follows from that fact. There is a human who skims — the marketing layer, where the value is stated quickly and plainly. There is a human who wants the method and the reasoning behind it — the science layer, which is the layer you are reading now. And there is a machine or agent that reads the site programmatically — the agent layer, which exposes the same substance in a form another system can parse.

These three layers are not decoration added at the end. They are present on every generated site, because the audience is genuinely plural: a person scanning, a person investigating, and a program retrieving. A page that serves only one of them fails the other two. Treating all three as first-class readers is the design constraint, not an afterthought.

QPP and the economics of inference

QPP means questions per project. The target is zero — the engine prefers to infer what to build from the brief rather than to interrogate the user with a round of clarifying questions. Inference reads the domain and decides; interrogation pushes the work back onto the person who wrote the prompt.

The reason is structural, not stylistic. Each question spent is a handoff back to the human, and a chain of questions converts a single prompt into a negotiation. Preferring inference keeps the contract simple: a brief goes in, a built product comes out. Where the engine cannot infer, it should decide on a reasonable default rather than stall on a query.

Zero is a target, not an achieved fact. We state it as the direction the system is built toward, and we do not claim it is already reached on every project.

A single prompt, as discipline

TITI takes a single prompt and from it researches the domain, decides the architecture, designs, builds, tests, and prepares the deploy. The point of the single prompt is not convenience theatre — it is a discipline. It forces the engine to carry the burden of judgement rather than offload it.

If a single prompt is enough, then the system must do the inferring: it must decide the architecture and the structure without a checklist from the user. That is a harder standard to hold, not an easier one. The single prompt is the constraint that keeps the engine honest about what it is supposed to figure out for itself.

Quality as a contract enforced by hard floors

Quality here is a contract, not an aspiration. Each layer is scored, and a build that scores below its floor is refused — it is not shipped. The floor is a gate, not a guideline; failing it stops the build rather than producing a warning.

The audit is cross-family: the model that writes the content and the model that judges it belong to different model families. A model never grades its own work. This avoids self-audit, where a system that produced an answer is also the authority that approves it — a conflict that would make the floor meaningless.

So the floor does real work only because the judge is independent of the author, and the result is binary: a build is at or above its floor and proceeds, or it is below and is refused. There is no shipping-with-known-defects path.

Recursion as a discipline

TITI is used to build and improve TITI. We treat this as a discipline rather than a slogan: the platform is one of the products the platform builds. That is what keeps the standard honest — the engine is held to the same floors when its subject is itself.

This is under human supervision. A human approves the changes. It is not autonomous self-modification: TITI does not rewrite itself unsupervised. Recursion means the same engine and the same gates are applied to its own improvement, with a person in the approval position.

What TITI does not claim

Stating the principles is only half honest without stating the limits. TITI does not claim full hands-off autonomy today; that is the trajectory, and the engine is proven in production, but autonomy is where it is heading, not where it stands.

TITI does not claim that QPP is already zero — zero is the target. It does not publish specific economic numbers: subscriptions unlock access and tokens fund execution, but the figures are deliberately unset, and we will not invent them to sound concrete.

And TITI does not claim that the resident chat executes orders or payment end-to-end today. The chat is the interface; what it can complete on its own is a separate question from what the engine can build. Naming these non-claims is part of the contract.