800ms across 50,000 hours.

When we say 'median retrieval under a second on a fifty-thousand-hour archive,' we mean it. What we don't say often enough is how that number changes the texture of the work itself.

The first time a producer watches a query against their entire archive return three cited moments almost at once, something visible shifts. They stop typing carefully. They start typing curiously. The cost of asking has dropped enough that they ask differently.

Speed only matters because of what it lets people do next. When you can ask "every clip where the founder mentioned the supply-chain pivot in the last six months" and get an answer before your fingers leave the keyboard, the question above changes. You stop asking what you remember. You start asking what you don't.

The hard part isn't speed. It is keeping the right answer in the top three.

We are not the fastest system on a benchmark. We are the system that returns a cited answer fast enough that it doesn't feel like a query. That is the bar.