The messaging doc says three pillars, approved language, bridge phrases for the hard questions. Then your spokesperson does eleven interviews, and somewhere between the doc and the delivery, drift happens: pillar two never comes up, an off-script analogy becomes their favorite line, and the hedge word they added to the revenue message is now in every telling. Nobody decided any of this. It accumulated.

Comms teams review appearances one at a time, warm from the studio — which catches disasters but never drift, because drift is invisible in any single performance. It only shows across the body of work: what's actually said, tallied against what was supposed to be said, across every appearance this quarter.

deep_context runs the drift analysis over your archived appearances (see the press-tour tracking use case for the collection habit): scope the quarter's imports and ask the audit questions — which pillars actually got airtime? what language replaced the approved phrasing? how does the spokesperson handle the acquisition question across tellings? Answers come back cited to the exact chunks, so the review runs on evidence.

run_youtube_tool's key-quotes supplies the per-appearance layer: the lines that actually landed — which, compared against the messaging doc, is the drift measurement itself. The output isn't a gotcha; it's the input media training never has: this is what you actually say, here are the eleven moments, and this is the line we're promoting to official because your version tests better than ours.