Separate construction from review
The AI team divided construction, research, transfer and criticism into bounded tasks. A specialist could investigate a tire marking, a receiving surface or an export discrepancy without owning every decision about the car. A separate critic could reject a locally successful test when the visible result or the source interpretation remained wrong.
Independence mattered most when there was already a plausible explanation. The rear slot and the cockpit paint transition both required another look at what the images actually established. Review was useful when it changed the next action, not merely when it repeated that a script had completed.
Save the decision with the work
Python scripts in Blender described construction and inspection steps. Separate candidate scenes protected the current assembly during experiments. Checks compared affected geometry, materials and image bindings, while saved views made the visual judgment reviewable. An accepted local candidate could then be promoted and compared against the reviewed version.
This discipline also exposed ordinary implementation mistakes. A saved material could disappear on reopening if nothing retained it. An old whole-scene bake could overwrite newer work if treated as the latest assembly. The integration approach therefore transferred named, reviewed pieces into the current scene and checked the result after reopening.
The archive is substantial, but it is not a complete film of every action. Some early failure output was observed during the session without a dedicated raw log. This case study uses preserved checkpoints for historical comparisons and labels diagnostic images by purpose. Current renders are evidence of the delivered model, not reconstructed screenshots of missing moments.
Define the question
Name the disputed feature, its reference and the nearby work that must remain stable.
Build a candidate
Save the experiment separately with its script, checks and useful camera views.
Challenge the result
Compare the actual images and source evidence; qualify what the measurements cannot establish.
Integrate and reopen
Transfer the accepted change, then verify that the saved assembly preserves it.