Which checks ran, and what they leave open
This inventory names actual selected checks. Available packs that were not selected provide no evidence for either delivery. General rules still reflect a caller-selected policy; task requirements need an appropriate mapping and evaluator.
Scroll sideways for more columns on a small screen.
| Pack or area | Dairy case | Distribution-centre case | Boundary |
|---|---|---|---|
openusd and nvidia.asset-validator | Selected structural rules; dangling bindings found | Same baseline families; missing normals found | Credit the underlying validators; selected policy is not whole-scene certification |
scene.audit | Dependencies and 99 decoded images; no animated-transform subjects | Dependencies, 25 → 26 decoded images, finite authored transforms | Decoding does not establish the correct artwork or readable rendering |
brief.measurements | 46 selected metadata/bounds/gap/count rows | Not selected; custom consumer measures relevant stage/count/string requirements | Selected subjects need physical meaning and coverage review |
materials and textures.decode | One material-delivery row and four named texture decodes | Not selected separately | Similar topics below use the trial pack, not these pack executions |
trial.distribution appearance | Not used | Exact pixels for two references, panel dimensions/UVs and selected steel/rubber values | Lighting, rendered hue and legibility remain review questions |
motion.connection | 63 pipe-to-port point pairs at two times in a static scene | Not selected | Endpoint agreement does not validate the route between endpoints |
trial.distribution motion | Not used | Parcel events/stop, shuttle travel/attachment, gate and lamp schedule | Diverter meaning reviewed separately; finite checks are not continuous collision proof |
General geometry, motion, motion.timing packs | Not selected | Not selected | Pack availability is not evidence that it ran |
| Optional harness LLM/VLM review path | Producer opinions replayed through adapter | Not run; separate geometry/visual reviewers instead | Different review arrangements, not comparable judge scores |
| Physics simulation | Not run | Not run; kinematic brief | No fluid, heat-treatment, contact, throughput or engineering validation |
Dairy
The dairy producer built the scene, supplied the brief interpretation and replayed its own visual opinions through the optional adapter. It also planted the fourteen test faults in copies of its finished scene. This is not independent grading.
Preparation proposed 37 requirements: three script-only, 24 visual, two both, and eight unsupported. The proposed mapping and tolerances remained pending. A caller-authored map added 114 measurements: selected bounds, clearances, bindings, textures and 63 pipe-to-port pairs. Those measurements do not establish valid process topology, acceptable plant operation or approval of the agent-chosen tolerances.
The fourteen seeded faults yielded two baseline rejections and seven with mapped checks. A blocked aisle escaped the selected scripts. The control results include the unchanged control separately. They describe coverage of deliberately chosen cases, not a general producer failure rate.
The dedicated separator view clipped its sign and process board. Interactive navigation remained untested. Successful content checks used the folder-based USD package after the separate USDZ encountered an input-admission error. The original failing binding revision was not preserved. Broader acceptance remained NEEDS_REVIEW.
Review arrangements
The dairy producer used Claude in Cowork. The distribution-centre producer used gpt-6-astra/xhigh in a fresh project. Its coordinator wrote the brief and prepared the checks before delivery. It had seen portions of draft code, and the roles shared infrastructure. No findings went back to the builder before its first complete delivery was preserved.
Separate review agents recorded initial findings before seeing scripted verdicts, with some prior topic-instruction background disclosed in the original records. The larger trial did not run the harness’s runtime LLM/VLM judge adapter. These arrangements and differing tasks do not support a model ranking or a blinded comparison.
Neither case establishes contact physics, dairy treatment performance, facility throughput, engineering qualification, general defect-detection rates, time savings or final owner acceptance.