Self-Supervised Temporal Alignment of Multi-Sensor Time Series

A world model represents the state of a temporal system and predicts its evolution from asynchronous streams. Transduction, propagation and actuation displace the instant at which an event reaches each stream, hence samples sharing an index can describe distinct events. Co-indexed objectives typically fix the inter-stream offset at zero, replacing a physical property of the sensing apparatus with an inductive bias. We propose QUARTZ, a joint-embedding predictor for quantified asynchrony over representation translations at zero annotation, learning an offset distribution and a directional influence per ordered stream pair. Correlation over the candidates makes the reported offset follow a test-time displacement without timing labels, and the recovered graph reproduces the order in which events reach the streams. Re-aligning a desynchronised stream at the reported offset improves frozen-probe accuracy, which enables a deployed model to correct its misaligned streams without annotation.