Tau: infrastructure for engineering RL

Our customer engagements and internal benchmarking indicate that frontier models already possess substantial latent knowledge of hardware engineering. Their limitation lies in operationalizing that knowledge reliably inside native CAD environments, where most high-value engineering work is performed.

Training models to do this requires hill-climbing on real-world, long-horizon engineering workflows. Our in-house work on CAD RL has identified three fundamental bottlenecks in this regime:

• Data: native CAD file formats are proprietary binaries, so the best production-grade data (the kind that carries design intent, feature trees, and constraints) is inaccessible for pre- and post-training.
• Environment: existing CAD systems were not designed for RL. Open-source tools lack the robustness required for complex geometry, while closed-source systems carry legal constraints and are prohibitively expensive to scale.
• Verification: engineering correctness is not reducible to geometric similarity alone. Effective rewards must also capture properties such as constraint completeness, manufacturability, and functional performance.

These limitations motivated us to build Tau: a purpose-built 2D/3D CAD platform for greenfield and brownfield RLVR directly on native CAD assets. Tau is written in Rust and built on the Parasolid geometry kernel, providing production-grade robustness and efficiency. But unlike traditional CAD systems, it strips away the surrounding application overhead in favor of headless, CLI-first execution designed for massively parallel rollouts.

Tau also exposes geometry, topology, feature history, constraints, and document state as first-class programmatic objects. Native CAD assets can be normalized into a canonical representation without flattening away the design intent encoded in their feature trees and dependencies. This turns existing engineering work into reusable training environments and enables deterministic rewards to evaluate both the resulting artifact and the engineering state that produced it.

Together, these properties make operationally intensive post-training tasks straightforward to construct and verify. Drawing generation, for example, is typically evaluated via LLM-as-a-judge against a handcrafted rubric, often after OCR or agentic extraction step to recover the contents of the generated PDF. In Tau, the drawing itself is a queryable object, so its dimensions, annotations, constraints, GD&T, and layout can instead be checked deterministically, much like unit tests against a software repository.