meghna tighare

LabWatch is an evidence intelligence layer for physical experiments. Researchers still log experiments by hand, and mistakes such as stirring too briefly, skipping a step or misreading an endpoint are found only after the run is wasted. AI tools often describe things they never measured. LabWatch takes a lab video or an instrument record plus the written protocol and: - measures what happened (colour, motion, objects, signals) with plain, repeatable measurements, - turns it into timestamped events with detection confidence, - checks them against the protocol (missing steps, wrong durations) and flags uncertainty such as lighting changes, - writes a lab notebook where every sentence must cite an event; a checker flags unsupported claims and invented ids, - lets the researcher click any claim to replay the moment in the video. On top of that: a hypothesis lab (competing explanations, a critic, next experiments ranked by computed information gain, a safety gate and human approval), learning from researcher corrections, live early warnings, an experiment genome with cross-run pattern mining, a calibration registry that only promotes detector updates that pass a holdout, and a simulated equipment-failure investigation.