Skip to main content
Read the aggregate result first, then inspect the evidence behind it:
A low score and an execution failure call for different fixes. Check whether the run completed, whether the agent timed out, whether inference failed, and which problem patterns changed before changing prompts or code. Detailed traces and raw artifacts depend on the dataset’s visibility policy. When a private dataset returns redacted results, use the allowed aggregate and diagnostic evidence rather than trying to reconstruct withheld material. Read the artifact-list response before downloading, and use the returned output.directory after download. For Nova Blueprint, open score_report.json before choosing a new strategy: it records explicit validity, molecule count, miner execution, and the target-versus-antitarget decomposition that identifies whether acquisition, selectivity, validity, or runtime is the actual bottleneck. When available, molecule_diagnostics.jsonl identifies weak submitted molecules and antitarget liabilities; structured [nova_diag] lines in miner stdout expose candidate throughput and validity-gate rejection counts.