Start from the exact job, input, and error. Distinguish transport, access, preparation, execution, and scientific interpretation before changing a workflow.
Submission failed or the response disappeared
If the service rejects fields, compare their exact names, types, required status, and accepted limits against the current form. For REST uploads, let the HTTP library construct the multipart boundary.
If a network timeout occurs after sending a submission, check history, the job note, and existing IDs before retrying. A timeout does not prove that no computation started. Never resolve an uncertain submission by repeatedly clicking Run or retrying a POST in a loop.
A job is pending or taking longer than expected
Pending means execution has not started; running means it has started but has not completed. Runtime estimates describe typical behavior, not deadlines. Size, samples, model settings, queueing, and preparation can affect elapsed time.
Keep the job ID and check its current status. Cancel only when you intend to stop the work; cancellation is not a diagnostic reset and should not be assumed to erase compute usage.
An input is accepted but results look wrong
Check sequence identity, biological assembly, molecular partners, stereochemistry, protonation assumptions, chain IDs, and residue numbering. A syntactically valid file may still represent the wrong experiment.
Inspect preparation and structure validation. For predictions, examine native uncertainty and plausible geometry rather than only the top score.
The same sequence gives a different predicted structure
Boltz-2 and Chai-1 sample structures stochastically, so repeated submissions and samples within one job can differ. Align the structures first, confirm matching molecular inputs and settings, and inspect local and interface confidence rather than judging screenshots alone. Large differences need investigation; they do not by themselves prove alternative biological conformations. See Why repeated predictions can differ for comparison and selection guidance.
Files or results are missing
Confirm completion, service-specific expected filenames, and whether a downstream selector matched them. For pipelines, inspect each node’s errors and child jobs rather than assuming the whole graph finished.
Check account, team, and share access. An unknown-job response may conceal a permissions mismatch. Temporary download URLs expire; request a new URL through the job route. Failed jobs and their associated files are subject to the deletion period described in Jobs and results.
A pipeline has size or mapping errors
Parent lists must be size one or matching length at a service node. Check order, empty entries, and output selector limits. A list of N entries and a file containing N internal records do not necessarily produce the same job count. See Bulk submissions.
Ask for help
Send the job or pipeline ID, exact error, service name, and the smallest useful description of expected versus actual behavior to support. Include relevant settings and a reproducible example when possible. Keep API keys, share tokens, and unnecessary confidential inputs out of public reports.