RunWave videos show what happened on screen, but they do not show which inputs caused each moment. This makes some serious game bugs hard to judge from the video alone.
Example: if the playtester presses Left and the game moves the character Right, the video only shows the character moving right. A verifier, reviewer, or VLM judge cannot tell that this was caused by the wrong control mapping unless the action timeline is available alongside the video.
Current artifacts partially contain this information in different places, but there is no first-class artifact that answers: “At this moment in the video, which input was RunWave applying?”
Why this matters:
- The video alone can hide broken controls, reversed movement, ignored inputs, bad drag handling, or delayed reactions.
- Downstream benchmark verifiers may judge only visual output and miss that the game responded incorrectly to the playtester.
- Human reviewers also need to correlate game behavior with the inputs that caused it.
- The agent’s requested action is not the same thing as the controller’s executed action, so this needs to be represented clearly somewhere in the run artifacts.
Expected outcome:
RunWave should preserve enough action timeline evidence alongside the video so downstream users can understand what inputs correspond to important moments in the recording.
RunWave videos show what happened on screen, but they do not show which inputs caused each moment. This makes some serious game bugs hard to judge from the video alone.
Example: if the playtester presses Left and the game moves the character Right, the video only shows the character moving right. A verifier, reviewer, or VLM judge cannot tell that this was caused by the wrong control mapping unless the action timeline is available alongside the video.
Current artifacts partially contain this information in different places, but there is no first-class artifact that answers: “At this moment in the video, which input was RunWave applying?”
Why this matters:
Expected outcome:
RunWave should preserve enough action timeline evidence alongside the video so downstream users can understand what inputs correspond to important moments in the recording.