AUTHORIZED RECORDER AND FILE

The host provides and operates OBS Studio on the Windows evaluation machine.
OBS Studio capture is explicitly authorized. Before the agent starts, the
host configures capture of the existing Mindustry window, starts recording,
and then hands the paused game to the agent. The host keeps recording through
final verification and stops after the agent declares completion.

The host must deliver the unedited, normal-speed MP4 as:
D:\Benchmark\deliverables\run_recording.mp4

run_recording.mp4 is the only candidate recording filename. The separate
expert reference is named reference_run_recording.mp4 and is evaluator-only.
The agent must not install, configure, control, or edit OBS Studio. Report an
environment blockage if the host recording is not ready before the attempt.

VISIBLE VERIFICATION INTERVAL

For at least 60 continuous seconds, show the designated Impact Reactor and
neighboring battery structures together with a readable English HUD showing
Impact Reactor, Power Output: 6300.0, and a visibly full Health bar. Keep the
game running unpaused at normal speed. Recording 65 seconds of the final
state is recommended for frame-timing margin; the threshold remains 60.

The qualifying interval may appear anywhere in the video. Earlier inspection,
switching and startup may remain. A qualifying final 112-second segment is
long enough. No separately scored startup deadline applies.

Different runs may use any screen resolution, aspect ratio and window position.
Keep capture-canvas dimensions constant within each video. No fixed
screen coordinates or per-recording layout profile are required. Keep the
text and map landmarks readable and unobstructed. Black borders are allowed.
Prefer a stable view during verification and record at 30 fps when practical.
The checker samples every 3 seconds, then checks two source frames per candidate
second (at the second's start and midpoint). It does not inspect every frame.
Missing fine samples or samples more than 0.25 seconds late do not qualify.

Do not edit the footage, alter playback speed, obscure the required indicators,
crop away the reactor landmarks, or substitute a screenshot loop. The checker
tests a visual outcome and a freeze heuristic; it does not authenticate the
recording's provenance or the complete preceding action history.

HOST/EVALUATOR SEPARATION

The host provisions Python dependencies and native FFmpeg/ffprobe in a separate
evaluator environment before validation. They are host tools, not solver GUI
tools. Checker code/assets, reference videos/answers/results and prior-attempt
artifacts must be inaccessible to the solver. The host enforces separation with
a separate machine/workspace or verified OS account permissions; a label alone
is insufficient. Only this recording instruction file is copied from the
evaluator package into the solver inputs.