Paws And Pencils

Most students never get to measure physics themselves: sensors are expensive and video-analysis tools make you click through hundreds of frames. FrameLab turns a phone video of a simple physics experiment into verified measurements — with honest uncertainty and an independent AI auditor. How it works. A student films a pendulum or a falling ball and says the string length and a guess for the period. Muse Voice Transcribe hears it; Muse Spark watches the video and plans what to track. SAM 3.1 segments the object in every frame (and the ruler), and code turns the masks into sub-pixel positions with a calibrated scale. A Muse Spark analyst agent chooses the physics model and calls tools (fit, block bootstrap, derive, quality gates); it never writes a number itself. Every value is evidence in a Claim Ledger, and the report's sentences are filled from it. An independent auditor in a separate context reviews claims and gates and can send the report back or ask for a new recording with concrete tips. Muse Spark Open, Meta's open-weight model served by Baseten, designs the next experiment with predictions computed by code. What it delivers. An annotated video, charts, a bilingual (Indonesian/English) lab report, a live agent trace, a clickable claim ledger, and a pendulum series that tests T ∝ √L. Honest results. On our own home recordings, three pendulum runs landed within 1.6 % of local gravity, a threelength series gave the exponent 0.467 ± 0.019 (theory 0.5), and a recording with the ruler in front of the wall was refused because two independent scales disagreed by 61 %. Free fall on a phone without a ruler of known length was 12 % low; FrameLab flagged it, and the README says so. Everything was built during the hackathon; HACKATHON.md has the dated log.