Submissions score 50/50 with 0 failures but flagged "Evaluation could not complete"

Dear ARC WhestBench Organizers,

I hope this message finds you well. I am writing on behalf of team The_Alchemist regarding a grading anomaly I have encountered on two Phase 1 submissions, which I believe may be a session-finalization issue on the evaluator rather than an error in my estimator.

Submissions #318208 and #318221 (Phase 1, team The_Alchemist) both report “Graded · 50/50 public scored, 50 completed, 0 failed” with a valid adjusted score , yet each is flagged FAILED with the message “Error: Evaluation could not complete; please retry.” I have retried the submission as instructed, and it reproduces the identical outcome all 50 per-MLP scores compute successfully, but the overall evaluation is marked as failed.

Because every per-MLP score is produced without failure and a complete adjusted score is displayed, this appears to be a finalization or total-evaluation issue rather than a problem with the estimator itself. For context, my estimator is compute-heavy (it relies on batched linalg.solve), and a lighter earlier submission graded cleanly, so I suspect the failure may relate to a total-evaluation wall-clock or resource limit reached during finalization.

Could you kindly confirm whether this is a known grader issue, and whether these submissions can be accepted or re-graded under the completed public scores? I would greatly appreciate any guidance on how best to proceed.

Thank you very much for your time and for organizing this challenge.

Best regards, The_Alchemist

Consider this anecdotal or empirical evidence unless anyone from the organizing team confirms, but through similar experiences what I’ve come to observe or believe is that there is
a) a documented max 60s time that you can spend per MLP, but also
b) a second (I believe undocumented, hence why I’m just guessing) limit of 1000s per worker.
There are 5 workers of 20 MLPs each, so keeping an average of ~40-45s per MLP should, in my experience, finish without any error.