Fit for the Engine? Image-to-3D Meshes, Their Pipeline, and an AI Assistant's Explanations in a Completed Unity Game

A solo developer built a complete Unity game with AI coding assistants and filled it with image-to-3D models. We audited its 26 generated meshes against a plan committed before any audit script ran, and then did something the usual evaluation of generated 3D cannot: we set the assistants' own written explanations of the asset problems, kept in commit messages, tool comments and design notes, against independent measurements of the same assets. The raw meshes are closed and edge-manifold, and their UV layouts resemble the game's purchased meshes; most problems came from the pipeline around the generator. The importer reversed the characters' front (glTF puts it at +Z); the import path stored each asset's 4096² base map uncompressed and dropped its normal and metallic-roughness maps, so 24 maps make up 70% of the build's texture data; MeshLab simplification of the seam-split buffer tore the companion into 2,433 pieces, where welding first leaves every asset intact. Of 39 testable claims the assistants made about the assets, 8 are wrong (31% under the stricter of two judges). The confirmed claims are mostly readings off a file; the wrong ones concern how the assets relate to the pipeline, and four of them blame the generator for what the pipeline did. Four of the six contradicted claims were never revised, and one was copied from tool to tool as a "finding". Defects clearly visible in play were fixed, apart from an accepted idle bob; all but one of the defects not visible on screen shipped, including one an AI review had flagged in writing. We release scripts, measurements, the claims with both judges' verdicts, the defect record and the review reports.