Sourced from whest-starterkit @ 96e308c.

Troubleshooting — when something breaks

← Documentation

If your run errored, your score regressed, or your local and remote scores disagree, start here.

Doc	When to read
common-participant-errors.md	Symptom → cause → fix-now → verify, for the most common failures (wrong shape, NaN/Inf, exceeded budget, the 400 ms residual wall-time cap, signature mismatches, import errors, numeric blow-up in deep networks, whole-submission timeout).
faq.md	Quick answers: can I use scipy? what is `residual_wall_time_limit`? how many submissions do I get a day? is there a limit on total evaluation time? why does my submission score worse than my local run?

➡️ Where to look next

Want the tiered procedure for "estimator runs but something feels wrong"? → How-to: debugging checklist.
Suspect drift between machines? → Reference: whest doctor.
Need to interpret per_mlp[i].error, budget_exhausted, time_exhausted fields? → Reference: score report fields.
Rules question, an operation you think flopscope is mispricing, or a legitimate need for more residual wall time than the 400 ms cap allows? → email arc-whestbench@aicrowd.com (FAQ).