Synthetic Grounds

One question governs this program: what ground does generative AI’s new measurement capacity constitute for architecture? Frontier models are trained against verifiable rewards, and the method works wherever a verifier exists. Building codes are machine-checkable rule systems, building information models answer exact queries, energy and carbon have simulators and arithmetic. Where such an oracle exists, grounding works; the materials research of this lab lives on exactly that closure.

But the qualities that make buildings and cities worth caring about have no single verifier: the life of a street, the values communities attach to places, the fit of a type to a way of living. The program’s method for these cases is reciprocal grounding, developed in the chapter of that name (Wiley, 2026) and in Compositional Intelligence: no single signal is the verifier; the mutual resistance of several independent signals is. Where measurements converge, something closure-like appears; where they diverge, the divergence is the finding.

The program’s public artifacts are the theoretical frame in those two books, the Synthetic Grounds studio at UT Austin, in which generative tools are checked against site, code, physics, and carbon rather than against plausibility, and the lectures of 2025 and 2026 under this title. The datasets and atlases of the Latent Earth program are its evidence: they show what a model regresses to when nothing grounds it.