Fits turns an iPhone room scan into a furnished room you can actually buy. A headless Claude agent calls our tools, and the tools, not the prompt, enforce the rules: pick only returns items that fit, and cart_add refuses anything over budget or without a verified Senso check.
How it works: the room comes from an Apple RoomPlan scan (our room: 190.6 x 139.7 in, door 32 in, +/-1 in). Superlinked SIE ranks 102 real IKEA products by style: SigLIP image embeddings plus the Qwen3-Reranker-4B on IKEA's own text, fused by reciprocal rank. Every pick is checked against the wall, a 36 in walkway and the door, down to every shipping box. Senso fact-checks each claim against a knowledge base of all 102 product spec sheets plus our room rules; a wrong width comes back as a conflict. Arcee open models write a second explanation per pick; across 6 models, glm-5.3-flash is the frontier at $0.00023 per item in 2.4 s with no invented numbers.
Every decision is an audit record with its source, streamed live into a 3D room UI. On approval the agent re-reads live IKEA prices, then scripts/ikea_checkout.py fills a real IKEA bag and checks it: live PASS, 5 items, IKEA total $887.97 = cart. Nothing is bought; the person pays on ikea.com.
Proof: 6/6 guardrail scenarios pass on a live run (contracts/eval.json) and 159 tests pass.