P369 · Tool use

Structured State and Rendered Evidence

Combine structured application state with rendered evidence when neither alone answers the task.

Editorially reviewed

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

Check whether prop crate contacts surface floor in a3D scene. Teaching frame 12 appears grounded from the camera, but structured state puts crate’s lowest world z at 0.08 and floor at z=0. Geometry and visible appearance both matter.

Mechanism

Record scene version, frame, object identities, units and axis convention. Extract world transforms and relevant geometry and compute clearance under an agreed contact rule; do not compare local coordinates directly with a world-space floor. Inspect a render from that same checkpoint for placement, occlusion and materials. If state and image disagree, report and investigate parent transforms, proxies or camera rather than choosing whichever passes. Motion tasks also require key-frame sampling.

Bad example

The frame 12 screenshot makes crate look grounded on floor, so contact passes. Geometry and frame identity need not be read.

Good example

Verify crate and floor in scene S3 frame 12 using world coordinates, units and geometric clearance, with a teaching contact tolerance of 0.01. Inspect the same frame's render for object identity and visible placement. Report failure in either view; do not substitute a different version's image or claim full-motion success from one frame.

Why the change matters

Pixels show visible appearance but hide axes, object identity and clearance. Structured facts supply those details, while renders can reveal wrong objects, occlusion or visual defects. Bind both to one version before combining them.

Observable expectation

Teaching clearance is 0.08−0=0.08, exceeding tolerance 0.01, so report floating despite plausible appearance. After a correction to 0.005, still check render and object identity. Different frame numbers make the evidence mismatched before contact is assessed.

Limits

Tolerance, units and contact definition depend on the task, not arbitrary defaults. One frame proves neither whole motion, physical stability nor a saved artifact. Sliding and penetration need denser sampling and appropriate predicates. Screenshots and structured exports can both omit hidden objects.

Sources and evidence

Read the editorial criteria