Stories have a video signature: before a topic dominates the cycle, its upload volume moves — explainers appear, commentary channels pick it up, the count of fresh videos mentioning the keyword jumps from two a day to twenty. Whoever watches that gradient hears the story building while it's still a whisper in text media.

Nobody watches gradients manually; watching is a machine's job. A scheduled sweep that counts and characterizes fresh uploads on your watch terms — and escalates only when the pattern breaks — is an early-warning system that costs prompts and pays in lead time, whether the topic is your company, your category or your risk register.

deep_search with the today/this_week upload filters is the sensor: each scheduled sweep returns what's new on the watch terms — volume, sources, and what the transcripts actually claim. Comparing sweeps is the detector: today's count and tone against the baseline, drift versus spike.

chat is the escalation layer: when the pattern breaks, it drafts the alert with context — what jumped, who's driving it, what the new videos claim, why it might matter — instead of a naked "volume up" that still requires the human to do the reading. Quiet days cost a sweep; loud days buy you the hours between the whisper and the cycle.