Phase-2 runtime/accounting clarification for accepted submission #322493

ARC White-Box Estimation Challenge 2026
#1

Hello AIcrowd team,

Submission #322493 is accepted and working. Before testing a native CPU
optimization, could you please answer each item with Yes/No or exact values?

  1. What exact Phase-2 evaluator image/digest, OS, Python, WhestBench,
    FlopScope client/server, and scorer versions or commits will be operative?

  2. May submissions bundle and call native CPU shared libraries or wheels,
    including NumPy, BLAS, oneDNN, OpenMP, or TBB, provided they use no network,
    GPU, or subprocess escape? Please list any prohibited libraries, dynamic
    dependencies, or ISA requirements.

  3. Is computation outside FlopScope charged solely through residual wall time,
    without disqualification or double charging? If yes, what is the exact
    residual-time-to-effective-compute formula and numerical precision?

  4. Must all native threads finish and every output be materialized before
    predict() returns? Are persistent thread pools created during setup()
    permitted?

  5. What is the exact lifecycle: frequency of setup(), process reuse across
    MLPs, cold/warm behavior, restart policy after a failure, and setup,
    per-predict, inner-wall, and total-submission timeouts?

  6. What are the exact CPU/core/thread quota, affinity policy, RAM/RSS or cgroup
    limit, compressed and extracted package-size limits, file-count limit, and
    native-library workspace limits?

  7. Please confirm the exact per-network score formula, compute budget, score
    floor, aggregation order, timeout/failure treatment, and rounding.

  8. Does submitting a new archive automatically replace or designate it, or can
    #322493 remain designated until the new result is complete? Can we
    explicitly switch back to #322493 after evaluation, and what is the exact
    Phase-2 cutover rule?

Thank you.