Represent one gameplay object with one named semantic root. Put gameplay position, heading, rigid-body, collision, and behavior on that root. Put calibrated render transforms below it.

Consume the existing tuning record from exactly once. If it is missing, stop and calibrate the model; do not calculate or adjust asset tuning during assembly.

Declare intentional support, attachment, or containment relationships between roots. After a rendered frame, measure each relationship at the supported object's actual position. Use a mount point, collision ray, or geometry sample; use a support root's AABB maximum only when its top is flat across the full footprint. Record the signed contact gap: positive means floating and negative means penetration. A relationship name does not exempt the pair from this check.

Union descendant world values for each semantic root and compare unrelated peer roots. Any intersection with positive depth on X, Y, and Z beyond a small tolerance is an unresolved placement failure: move the root from its calibrated footprint, or prove the apparent AABB overlap is empty geometry from at least two materially different views. Never accept one model inside another from a single screenshot.

Also check expected root counts, component descendants, aligned render and collider bounds, and role-based names. Compare final dimensions with each calibration record's intended size and reference object. Inspect composition and facing from images returned into your context by an image-capable browser tool. A saved screenshot path, screen coordinate, AABB, or visibility flag is not visual evidence. If you cannot inspect the image, report those claims unverified. Hierarchy and placement claims require runtime values logged from the application itself, as the skill describes.

Read exactly one reference matching the code being edited: direct Engine, React, or Web Components. Choose from imports and markup, not installed dependencies alone.