CHRONOSEEK

Finding one moment in a long video still means scrubbing. Keyword search only helps if the right words were spoken or typed into a caption.

ChronoSeek is a Bittensor subnet built around a different idea. You describe the moment in plain words, and miners compete to return the exact timestamp. Validators score their answers against ground truth, so accuracy is the only way to earn.

I built the web front end. The subnet code lives in its own repository. The site has a landing page, a playground, a miner dashboard, API docs and a whitepaper page.

In the playground you load your own MP4 or WebM, or the sample clip, and describe what you are looking for. You get back a timestamp, a confidence score, a time window and a few runner-up candidates. The file never leaves your browser.

The dashboard lists 48 miners with accuracy, timestamp error and incentive share, with filters, sorting and 24H, 7D and 30D ranges. It's built with Next.js 16, Tailwind CSS and shadcn/ui.

The subnet's public API isn't open yet, so the site has nothing real to call. Random numbers would change on every refresh and make the demo feel broken. So everything is seeded.

In the playground, the query text and the clip length are hashed into a seed for a small random number generator. Ask the same thing about the same clip and you get the same answer every time.

The dashboard uses one fixed seed. On each load it regenerates all 48 miners and 1,440 hours of history in memory. Because the data is generated rather than typed in, the ranges and filters actually work. The catch is order. The generator is stateful, so every value has to be drawn in the same sequence. Change one draw and every miner after it shifts.

The site is live and says what it is. A banner marks it as in active development, the playground is labelled as a simulation with no subnet calls, and the dashboard carries a mock data chip.

The docs describe the proposed v1 API, which will be backed by the subnet once its validator and gateway work is finished.