Caliper

Caliper measures size and speed in phone video, or refuses when the video can't support a number. Muse Spark 1.3 turns the question into a measurement plan and audits the evidence; SAM 3.1 tracks the object and the reference; deterministic Python computes the value and a 90% range. The model never writes the number. Evaluated honestly: on the QuantiPhy holdout our pre-registered claim failed on the official score (20.1 vs 58.1), because refusals count as misses. But when Caliper answers, 9 of 15 answers land within 10%, against 4 of 15 for Muse Spark alone, and its range held on 13 of 15. On our own scooter, it measured 5 of 6 passes against GPS with a 7% median error. Also usable by other AI agents through an MCP server. """ Live: https://caliper-sigma.vercel.app """ Built during the hackathon (Oct 27 to Nov 4); planning docs and field clips are listed separately in the README.