Write a reproducible research report.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Start with the question

Your average improves but one task regresses. What must the conclusion say?

Mechanism and worked comparison

A report is an argument whose evidence can be inspected. Begin with the frozen question, the baseline, the intervention and the rule for deciding whether a gain matters. Show every paired task/seed result, not only an attractive average. A negative finding can be useful when the comparison is fair and the result is reproducible. An unsupported positive finding is not improved by confident wording.

Use the evidence field to name exact task/seed rows and the improvement direction. Use the manifest field to identify the study version, data generator, split, seeds, settings and code artifact. The export includes those records and the recomputed values; prose remains learner-authored. Import validation can check a bounded recipe and replay scores, but cannot establish that your interpretation is sound or that an independent person ran the code.

Separate limitations that affect interpretation from merely inconvenient details. Shared synthetic tasks, repeated development, three task clusters and an inspectable final partition constrain the claim. Do not call a local run an independent replication. Submission checks required fields and all nine results, not writing quality. Objective transfer questions and pending human review appear separately. A reviewer should be able to disagree with the conclusion while reproducing the calculation.

Try and explain

Complete the project’s claim, evidence, limitation and manifest fields. Export all runs with the recipe; include a negative result and the task-level regression.

Open the reproducible study project or use the implementation studio.