An LLM proposes the load; Universtato verifies it — fit, collisions, stability, cube — exactly, in milliseconds. Two planted flaws, both caught.
Cartons
20
Flaws caught
2
Cube used
64.1%
Determinism
all identical
All 5 checks
17.0 ms
Side elevation. Teal = valid cartons · orange = overhanging carton that tips · red = carton protruding past the rear door.
Fit-checkrelational_graph (inside)
"Does the proposed load fit inside the trailer?"
↳ NO — 19/20 cartons fit; C_door protrudes 1.0 m past the rear door
10.388 ms96 ops/sdet ✓
Validitycollision (scene geometry)
"Do any cartons interpenetrate?"
↳ no interpenetrations — every carton has clear space
0.228 ms4,385 ops/sdet ✓
Load-safetycheck_stability
"Is the load stable for transit?"
↳ NO — C_over tips +x (CoM 0.10 m past its support edge) — restack
0.288 ms3,474 ops/sdet ✓
Cubevolume (scene geometry)
"What is the cube utilization?"
↳ 64.1% cube used — shipping 35.9% air (20 cartons in 12×2.4×2.6 m)
0.006 ms155,280 ops/sdet ✓
Sequencerrelational_graph (on)
"What is stacked on what (unload order)?"
↳ 9 support relations — top layer rests on the floor layer (unload top-first)
6.102 ms164 ops/sdet ✓
What this shows. A proposed container load checked by five agents against one Universtato scene: fit via relational_graph's inside predicate, stack stability via check_stability (center-of-mass over support), plus collision, cube utilization, and unload order from scene geometry. The engine computes each answer exactly — it does not eyeball the plan. Maps to research items #8–10 (freight load planning / cartonization / mixed-pallet stability — a 30-year paid market: Cube-IQ ~$2,500/mo, EasyCargo, MaxLoad). Latency/determinism measured on this machine.