For a multi-stage investigation, ask OpenScience to maintain an explicit research plan with deliverables and checks. This is useful when you compare several candidates, reproduce a claim, or need to explain why a result counts as progress.
Define the work before running it
Define the objective and required deliverables for this experiment.
Record the baseline, primary metric, direction of improvement, validation
checks, and stopping criteria. Keep the plan with the project and show it
to me before starting trials.
Use exact ceilings when you have a hard time, token, call, or cost limit. A research instruction does not replace external account spending controls. See Pricing and Planning.
Save the plan before observing outcomes
When the task calls for preregistration, ask OpenScience to save the empirical plan as a Result and register that saved plan before running trials. Keep the objective, dataset or sample selection, primary comparison, metric, and decision rule in the plan.
A local preregistered plan preserves the intended order within this workflow. It is not automatically a registration in an external scientific registry.
Track stages and checks separately
A stage describes progress through the workflow. A verification check describes whether evidence supports a requirement. Marking a stage complete is not enough to establish that a statistical result, file, or implementation passed its check.
Update the experiment status. For each required check, identify the
actual output or execution that supports pass or fail. Mark anything
not run as pending or not tested rather than assuming success.
Record trials and failures
Give each candidate a unique label and record its observed outcome, metric, baseline, and supporting outputs. Include failed attempts and explain the relevant failure.
Compare the baseline and candidate runs using the same evaluation data.
Record their measured metrics, link the output files, and state whether
the candidate improved, regressed, or remains unresolved. Include the
failed run and what prevented a valid comparison.
Keep a model's explanation separate from the measurement that supports the conclusion.
Retain useful lessons
Ask OpenScience to record a lesson only when a verified trial supports it. State the situation in which the lesson applies and the evidence behind it. If later results contradict the lesson, record the counterevidence and revise it.
A lesson about one dataset or environment should not be generalized to unrelated work without further checks.
Review the investigation
Ask for the objective, active stage, completed checks, trials, failures, required outputs, and unresolved questions. Before handing off, save a readable report and continuation note.
Use Compute jobs for execution, Reproduction for matching an existing result, and Saved Results for retaining evidence files.