Why grounded video answers need timestamps

Paste a long lecture into ChatGPT and ask "what did the speaker say about gradient descent?" You get a confident answer. You have no idea if it is true, which minute it came from, or whether the model invented half of it.

Summarization is not the bottleneck. Trust is. A grounded answer lets the user click a timestamp, watch the original clip, and verify the claim. Without that loop, every answer is a pleasant fiction and users stop trusting the tool.

Here is the shape of answer we are building toward: "The speaker introduces gradient descent at [2:14](link) and walks through the learning rate at [4:31](link)." Every claim carries a deep link that plays back the exact moment in the video.