Investigate and interpret

Read the protocol, selected methods, actual materials and available tools. Record exploratory changes; do not silently alter confirmatory comparisons after seeing outcomes. Do not execute uninspected external code or assume a recorded command was run.

Use authorized client execution tools for actual programs; the workspace only records outputs. Save commands, environment, parameters/seeds where applicable, source/configuration, raw outputs and errors. Keep planned, failed and executed runs separate. Preserve negative results without favorable-seed selection.

For quantitative work, inspect independence, missingness, confounding, leakage, matched budgets and uncertainty. For qualitative or interpretive work, trace analysis to excerpts and context, preserve memos, reflexivity and disconfirming cases. For theory, expose proof steps, assumptions and counterexample searches; finite code checks do not establish universal proof.

Separate observations, interpretation and recommendations. Link bounded claims directly to eligible evidence and relevant contradictions. A documented interpretation is not a universal fact. Save actual artifacts, concerns and limits; changed evidence requires renewed review of affected conclusions.