Testigos del Bólido

Witness videos in, a fireball's path out.

How the agent works

  1. Muse Spark watches each video: camera type, edits, on-screen clock, when the fireball shows, and landmarks a map could locate.
  2. SAM 3.1 follows the fireball frame by frame from the phrase "bright light". The agent's track tool runs it on any clip nobody has tracked yet; with a plain brightness threshold instead, the same two cameras give a heading more than 20° off.
  3. Each camera is oriented from lines and landmarks in one frame. That step is not automated yet: during development, AI annotators (Claude Code, not a Meta model) measured the lines and matched the landmarks to OpenStreetMap, and the code fits a camera model to them, so every pixel becomes a direction in the sky.
  4. The agent, Muse Spark 1.3 calling tools in a loop, compares the cameras' planes of sight, solves, leaves cameras out, finalizes, writes the report or asks for missing evidence. Two planes from different sides cross in the trajectory; speed comes from each video's own frame times.
  5. The code checks what matters: a camera leaves the solution only on a criterion the code measures; the chosen cameras are re-solved before they are accepted; a report is refused if a figure does not match a tool result of the same kind or if it states another verdict than the code's; and the run stops for a person, who signs the report off or supplies the evidence. The published answer sits where the agent cannot read it; a separate script grades the run.

Built during the Meta Global AI Developer Hackathon with Muse Spark and SAM 3.1. Witness videos belong to their authors and are not redistributed here: this site holds only directions, times and camera sites rounded to about 100 m. The videos it can play are their uploaders' YouTube videos in YouTube's own player, loaded only when you press play; no frame or sound of them is stored here.

Basemap: OpenFreeMap, data © OpenStreetMap contributors. 3D: deck.gl.