Abstract
Video is the fastest-growing body of human record, and the least addressable. We can search text to the character; video, in practice, only to the file. ChronoSeek is a Bittensor subnet in which independent miners compete to localize natural-language descriptions inside video with millisecond precision, validators score them against verifiable ground truth, and rewards flow strictly to temporal accuracy. The result is a public utility: a query API in which the unit of retrieval is not a document — it is a moment.
The problem
A search engine that returns a two-hour video has not answered a question; it has assigned homework. The information a user wants almost always lives in a span of seconds, yet the dominant retrieval unit remains the whole file, padded with metadata, titles, and captions that describe the container rather than the content.
The unit of retrieval should be the moment, not the file.
Centralized moment-retrieval models exist, but they inherit a single-vendor ceiling: one architecture, one training distribution, one blind spot shared by every customer. Temporal localization is precisely the kind of task where ensembles of diverse, adversarially-scored models outperform any single one — which is what a subnet is.
Architecture
A query enters through the gateway and fans out to registered miners. Each miner runs its own multimodal pipeline — vision, audio, and temporal context — and returns a candidate timestamp with a confidence. Validators aggregate, score, and emit a single consensus answer.
Miners are free to differ: one may excel at sports, another at dialogue, another at long-horizon surveillance footage. The network does not pick a winner in advance — the scoring mechanism discovers one per query.
Incentive mechanism
Rewards are a strictly decreasing function of temporal error. For a query with ground truth t*, a miner answering tᵢ earns weight:
where τ is a tolerance that tightens as the network matures and cᵢ is the miner's calibration term — confidence is rewarded only when it is earned. A miner that is 10 ms off takes nearly the whole emission for that query; a miner 14 seconds off takes effectively nothing. Precision is not a virtue here; it is the business model.
Validation & anti-gaming
Validators maintain ground truth through three task streams:
- Synthetic insertions. Known events are composited into footage at known offsets; the correct answer is exact by construction.
- Held-out human labels. Curated corpora with frame-level annotations, rotated unpredictably so they cannot be memorized.
- Cross-examination. Live organic queries are replayed to disjoint miner subsets; outliers against the consensus distribution are penalized.
Copy-mining is unprofitable by design: answers are committed hash-first and revealed after the round closes, so there is nothing to copy until it is too late to copy it.
Emissions
Subnet emissions follow standard Bittensor mechanics: miners earn by answering, validators earn by scoring honestly, and stake routes trust. No premine, no team allocation inside the subnet — the only way to accumulate weight is to be right about time, repeatedly, in public.
Roadmap
- PHASE 1Single-moment retrieval. The /v1/seek API, playground, and public miner leaderboard. Prototype in development.
- PHASE 2Persistent indexing. Index-once, query-forever for long footage; webhooks and batch search.
- PHASE 3Compound queries. Sequences ("the save, then the counter-attack"), counting, and cross-video retrieval.
- PHASE 4Agent-native access. Streaming grounding for autonomous agents that treat the world's video as a queryable sensor.