Please do not pollute the browsecomp eval

[Xianjun-Yang]

Hi,

I think this is a really cool and useful project.
But I noticed that this test file leaked the golden answer:

enterprise-deep-research/test_benchmark.py

Line 17 in ad0d535

expected_answer = "1988-96"

According to browsecomp's release note, all questions/answers should be encrypted.

Currently this pollution is impacting my model's training/eval, and perhaps yours, as well.

Can you use a dummy QA for test rather than using the real data from Browsecomp?

This will benefit both the community and yourself.

Thanks!