[hanhainebula]

Summary

long-document-reranking could time out during verification even when its individual stages stayed within their limits. The outer verifier allowed only 4,800 seconds, while the build, two query runs and trajectory judge allow 12,600 seconds in total.

Increase the outer window to 13,200 seconds:

1,800 build + 2 × 3,600 query runs + 3,600 judge + 600 setup/cleanup = 13,200 seconds

Update the task version to 1.0.1, synchronize the task README and verifier notes, and record the budget calculation beside the configuration. Submission stage limits, fixed inputs, determinism checks and scoring remain unchanged. Existing valid completed results retain their scores; any regrade using the expanded window should record the corresponding task version and timeout.

Validation

• python scripts/check_tasks.py tasks/long-document-reranking: 0 errors.
• Parsed the task configuration, grader constants and judge configuration to verify the budget calculation.
• Compared against the base revision: only the task version, outer verifier timeout and corresponding documentation changed.
• git diff --check: passed.

Full model evaluations were not rerun for this configuration/documentation fix.