CAD Autonomy Experiments

What needed solving.

CAD agents can miss requirements, damage interfaces, or optimize a proxy while degrading useful geometry. These experiments test whether an AI can discover omissions and revise a design while host-side geometry checks preserve artifacts and enforce acceptance boundaries.

1. 01Frozen requirements
2. 02Model proposal + sandboxed build
3. 03Independent geometry checks

Simplified from the implementation and project notes.

How it fits together.

The report records 10 live model calls and 21 sequential sandboxed builds against a prototype revision. Experiments included requirement audits for a drone and enclosure, battery-driven parameter changes, an unseen battery repair, a structured fillet operation, rollback on a failed multi-edit transaction, and bounded cost search. The model returned parameters or typed operation intent; fixed host code selected edges, built geometry, and ran acceptance checks.

The validation combined native requirements, shape-difference/dependency checks, non-overlap, extraction sweeps, motion probes, contact-area measurements, and independently frozen spatial envelopes. Transaction tests compared revision pointers and artifact hashes after a valid edit followed by an invalid one. Cost candidates were ranked under fixed assumptions, and a negative control demonstrated why design-relative checks alone can shrink with the design.

The findings are mixed and technically specific: the planner predicted affected components and repaired one unseen battery fit, but an unchanged gimbal contact patch shrank; one lower-cost candidate consumed margin and reduced skid contact; a stronger independent envelope caught a candidate that passed all original checks. The model also correctly refused an infeasible constraint combination after the inequalities were supplied.

Building and checking it.

Each experiment starts from a fixed design and prompt, runs the model through a sandboxed structured workflow, rebuilds through host-controlled CAD, then evaluates native and independent criteria. Results, prompts, artifacts, and an independent rubric are retained for review. A failed candidate can be revised, but a relaxed request is called out rather than presented as the original requirement passing.

This is a research harness, not a product API. The report recommends adding explicit interface contracts, separating physical requirements from construction parameters, using staged atomic revisions, and exposing cost alongside fit, contact, and manufacturing margins.