Reformula

Reformula helps a cosmetic chemist replace one ingredient in a formula, then check that the emulsion still holds. A Muse Spark agent reads the formula, finds candidates for the same role, pulls molecular data from PubChem and tests each change. When a test fails, it sees why and tries again. It returns ranked options, and plain code applies each one again to check EU limits, pH, use levels and preservation. Muse Wonder reviews an option as a sceptical senior chemist, and Muse Spark searches the web for the current regulation with citations. SAM 3.1 then measures stability. From one word, it segments every droplet in every frame of a timelapse. Fitting second-order kinetics to the droplet count gives a coalescence half-time, and linking droplets across frames builds a 3D family tree: a failing formula collapses into thick trunks, a stable one stays as straight threads. The micrographs are simulated, so the merge rate is known: SAM 3.1 recovered it within 3% (0.0007 per hour fitted as 0.00068, R² 0.999, half-time 12.5 h). The ingredient library is a small curated subset, and a chemist still decides.