Pre-API research prototype

Ask video
what happened.

Trace is a system for finding the moments that make an answer true.

We are building it to search across long video, retrieve the relevant spans, verify them, and return the evidence behind an answer.

The interaction below shows the product we are working toward using generated footage. It is an illustrative sequence, not a live result.

Target interactionIllustrative generated footage

Question

Which shots stopped working?

01:24open corner three

03:17contested pull-up

06:21late-clock deep three

Illustrative answerLate-clock threes and contested pull-ups were the weak spots.

3 supporting moments returned with timestamps.

What it is

Trace is a system design, not a single model. It coordinates cheap indexing models, temporal retrieval, and a stronger verifier so that expensive vision compute is used only where the question requires it.

The intended output is a set of timestamped evidence spans, provenance, and an explicit measure of whether the evidence is sufficient.

What it is not

Not a general-purpose video model.
Not a chatbot guessing from uniformly sampled frames.
Not a long-term memory database.
Not a claim that video understanding is solved.

How it is being built

A coarse-to-fine pipeline with model choice treated as an implementation detail. Each component has to earn its place through evidence recall, latency, and calibration.

Decode

Preserve shots and time boundaries while extracting frames, audio, and on-screen text.

Index

Build a temporal index from visual embeddings, transcript tokens, and scene-level metadata.

Retrieve

Use the question and compute budget to move from cheap coarse search to precise evidence windows.

Verify

Test whether the retrieved spans support an answer. Expand the search or abstain when they do not.

Proposed contract

The interface is deliberately smaller than the system behind it.

input {
  video
  question
  compute_budget
}

output {
  evidence_spans[]
  sufficiency
  uncertainty
  provenance
}

Open questions

How do we measure evidence recall when the answer spans several distant moments?

When should the system spend more compute, and when should it abstain?

Can one temporal index support screen recordings, sports, lectures, and physical-world video?

Ask videowhat happened.