The platform, in its third language

Everything is an agent, a state, or a workflow. TITI's internal rule is that every important object must exist in three languages — commercial, theoretical, executable — and this page speaks the third. A build is a workflow of typed stages; each stage is executed by an agent with a contract: declared inputs, declared outputs, measurable acceptance criteria. State is explicit — pending, in progress, blocked with a reason, done with evidence — so 'where is my build and why' always has a one-line answer. Orchestration is the only place where stages meet; agents do not call each other ad hoc, because untyped side channels are where provenance goes to die.

The three-layer page, structurally. The marketing layer is the route's primary document: hero, value, proof, one path inward. The science layer lives on routes of its own — depth essays with sources, cross-linked rather than interleaved, so marketing stays short because depth has somewhere better to live. The agent layer is embedded everywhere: JSON-LD in every head, a manifest at a well-known path, a source-of-truth map listing what may be claimed and what may never be, and a structured agent description whose contact surface points at the resident chat — because on this platform, the conversation is the designated interface, for machines reading about it no less than for humans using it.

The audit machine runs in two regimes. Deterministic gates execute first, on the rendered page rather than the source: element-level contrast against accessibility floors at desktop and mobile viewports including the narrowest, overflow and tap-target checks, opened-popover containment, tracker and draft-tag scans, and the categorical privacy blacklist. What passes is then judged twice by models that wrote none of it — a primary scorer across the layers and an adversarial reviewer from a different model family, prompted to refute. The two regimes are not redundant: determinism catches what can be measured, adversarial judgment catches what can only be argued, and the floors for both are numbers fixed before the build, not adjectives applied after it.

The recursive loop, stated as a workflow. Improvement of the platform is itself a typed workflow: analyze the platform's own output, design a change, generate the implementation, test it against the same gates every customer build faces, and deploy it — with a human approval standing before every promotion, today and by policy. 'TITI builds TITI' is therefore not a metaphor: the platform is its own first customer for every capability, and the page you are reading is an artifact of that loop. What the loop does not yet do is also part of the architecture: the conversational channel on these pages is a seam awaiting its backend bridge, and single-prompt end-to-end autonomy is the trajectory the stages are converging toward, stage by audited stage.

Governance is inherited, not improvised. TITI operates under a constitutional layer (HAYAT): provenance on every artifact, every number traceable to a source, published work superseded rather than edited, and human sovereignty — the standing ability to halt, inspect, or refuse any process — as the load-bearing clause. The platform is fast because the discipline is automated, not because the discipline was removed.

Known limits, kept in print. Judged scores are estimates with variance, so floors carry margin rather than pretending precision. Deterministic gates cover what can be computed from a rendered page — contrast, overflow, structure — not everything that matters; taste and truthfulness still pass through judges and, ultimately, a human. The agent layer guarantees what is published, not how any particular machine will read it. And autonomy here is bounded by design: nothing this platform builds reaches production without a human's explicit approval. A platform that hides its tolerances is not audited — it is advertised.

References

1. W3C — WCAG 2.2 contrast minimums (the deterministic floor) — https://www.w3.org/TR/WCAG22/#contrast-minimum
2. W3C — PROV-DM: the PROV data model (provenance as formal record) — https://www.w3.org/TR/prov-dm/
3. Google Search Central — structured data and JSON-LD — https://developers.google.com/search/docs/appearance/structured-data/intro-structured-data
4. HAYAT — the governance layer (public principles) — https://hayat.online/constitution