Video Evidence Lab

Humans and AI agents build, challenge, and revise a timestamped evidence map together—through 16 visible, reversible WebMCP tools.

Comment

Judge it in 60 seconds

This is not a video summarizer. The product is the shared, revisable evidence workspace.

  1. Open the live app inside ChatGPT's in-app browser.
  2. Keep the bundled original lecture and its current selected claim.
  3. Click 60-sec demo, copy the provided prompt, and send it to ChatGPT.
  4. Watch one agent turn read the same selection the person sees, search bounded transcript context, create a separate claim, attach supporting and challenging evidence, flag a contradiction, compare the claims, and revise the living conclusion.
  5. Inspect the exact timestamped passages, revision ledger, and undo control left visible in the page.

The judge does not have to trust a generated answer: every claim, evidence link, contradiction, revision, and reversal remains inspectable.

Why this must be WebMCP

The hard problem is not generating another answer beside a webpage. It is letting a person and an AI agent investigate the same source through the same visible, stateful interface.

Video Evidence Lab registers sixteen narrow, typed WebMCP tools. The agent can:

  1. Read the workspace and the person's current selection.
  2. Search the transcript and inspect bounded local context.
  3. Create a proposition without silently treating it as proven.
  4. Attach exact timestamped passages as support, challenge, or context.
  5. Flag an unresolved contradiction.
  6. Compare interpretations visibly.
  7. Revise the conclusion from grounded claims.
  8. Undo its latest mutation when the person rejects it.

This is deliberately not a one-shot "analyze video" tool. A meaningful investigation requires multiple calls over shared state. Every write changes the same page the person is using, returns a visibleChange result, enters the revision ledger, and can be inspected or reversed.

What the product demonstrates

  • Non-trivial WebMCP leverage: sixteen atomic tools support a read-before-write investigation rather than wrapping a single API call.
  • Complete execution: a working public app, deterministic bundled judge flow, original source material, eleven reproducible tests, lint, and a production build.
  • Specific impact: people can question, verify, challenge, and cite long-form video without separating an answer from the moments that support it.
  • Creative human-agent collaboration: the person and agent construct, challenge, and revise one persistent evidence object together.

Inspiration

Long-form video contains valuable explanations, arguments, and evidence, but it remains difficult to question, verify, cite, and reuse. A generic summary collapses the source into prose and separates every answer from the exact moments that support it. We wanted a person and an AI agent to investigate the same video on the same visible evidence surface.

What it does

Video Evidence Lab turns a public captioned video into a timestamped investigation workspace. A person can inspect exact transcript passages, select claims, change reasoning lenses, compare interpretations, revise conclusions, and undo changes.

Through WebMCP, ChatGPT can start from the video URL, read the same workspace and current selection the person sees, search the transcript, inspect bounded context, create separate claims, attach supporting or challenging evidence, flag contradictions, compare claims, revise the living conclusion, build a learning path, and undo the latest change.

The result is not a detached answer. It is a shared research object whose evidence and revision history survive the conversation.

How we built it

The React and TypeScript app registers sixteen tools with document.modelContext.registerTool.

The write tool load_youtube_video uses the same transcript service and state transition as the human import control. Read tools expose workspace state, the current human selection, evidence packets, compact transcript outlines, lexical search, and bounded transcript windows. Write tools reuse the application's own actions to jump to evidence, change the analysis lens, create claims, attach timestamped passages, compare claims, flag contradictions, create learning paths, update the conclusion, and undo changes.

Inputs use closed JSON Schemas. Transcript-bearing outputs are compact and explicitly marked as untrusted source material. Every write creates an activity record, making agent actions auditable in the page. The interface remains useful when WebMCP is unavailable.

Safety and verification

A transcript is evidence, not authority. The bundled original transcript includes a harmless hostile instruction asking the agent to modify the workspace. The agent can locate, quote, and explain that passage without obeying it.

Eleven deterministic tests verify tool annotations, closed schemas, bounded context, read-before-write journeys, shared-state handoffs, visible-change contracts, and the prompt-injection boundary. Lint and a production build are included in the public repository.

Challenges we ran into

The first challenge was giving the agent enough context to reason without dumping an entire transcript into one tool response. We solved it with progressive disclosure: inspect an outline, search exact concepts, read a bounded local window, and only then perform a visible write.

The second challenge was making agent work legible. A one-shot "generate analysis" tool would hide the interesting part. We separated claim creation, evidence attachment, contradiction detection, comparison, conclusion revision, and undo into auditable operations.

The third challenge was reliable demonstration. The app includes an original bundled lecture and transcript so judges can exercise the complete WebMCP workflow without depending on a third-party video or a fragile live caption response.

Accomplishments that we're proud of

  • A complete human-and-agent investigation workflow, not a one-shot summarizer.
  • Sixteen typed WebMCP tools operating over the same state the person sees.
  • Visible claim creation, evidence links, contradictions, comparison, conclusion revision, and undo.
  • Exact timestamped evidence and an inspectable revision ledger.
  • Explicit treatment of transcript content as untrusted evidence.
  • Eleven reproducible tests, including an adversarial prompt-injection case.
  • A public MIT-licensed repository and deterministic judge flow.

What we learned

The most useful agent experiences do not replace the interface. They let an agent and a person collaborate through narrow operations over shared state. Grounding becomes stronger when every interpretation remains connected to exact source moments and every mutation can be inspected or reversed.

What's next

Next we will simplify the consumer entry flow, improve live caption reliability, add multi-video comparison and citation export, and preserve the strongest WebMCP interactions inside the broader YouTube Conversation product.

The long-term goal is to convert any video into a source that ChatGPT can understand, question, and cite—while keeping the person in control of the evidence.

Built With

Share this project:

Updates

Submission history