← All examples
Positive · Cleared the baragentic_vbench_sports/gsw-cle-2018-finals-g4-three-point-timeline

Sports: three-point timeline reconstruction

The one sample fully calibrated to the bar — a strong agent scored 0.0465 over 78 turns.

The ask

Given the full broadcast of 2018 NBA Finals Game 4 (archive.org, 1080p, 155 min), reconstruct the timeline of all 22 made three-pointers: quarter, game clock, shooter, assister.

Exactly what the agent reads (excerpt)

Reconstruct the complete timeline of every made three-point field
goal in the game, by either team. For each made three, report the
quarter, the game clock at the moment it was made, who made it, and
who assisted it (or "unassisted"). ...

- Do not look anything up online, and do not rely on memory of this
  game; find every shot in the video.
- Count only made three-pointers, not attempts, not two-pointers,
  not free throws.

Ground truth[machine-truth]

The official ESPN structured play-by-play, cross-checked against the official box score (per-player 3PM matches for all 10 shooters). The answer key is the league's record, not a human label.

Scorer

Pure-stdlib F1 over events. A true positive requires a full play reconstruction: correct quarter, shooter, assister, and game clock within 5 seconds.

The lesson

The first version asked for the box score and the agent scored 0.64 by recalling this famous game's headline stats and reading on-screen graphics. The timeline redesign killed recall and overlay shortcuts. For production tasks, prefer less-famous games anyway.

Measured, not claimed

Oracle
1.0
Empty baseline
0.0
22-entry guess
0.0
Careful-watcher partial
0.45
Strong agent (Opus 4.8)
0.0465
Rollout
78 tool-call turns