Start with an observation, falsifiable hypothesis, baseline, correctness gate, and experiment plan. Preserve negative results and confounders. Keep cores vendor neutral. Do not submit private chain-of-thought, credentials, private logs, active infrastructure details, or data without publication rights. Use synthetic fixtures.