Defensible by Construction: Five Practices for Trustworthy AI-Agent Deliverables

Defensible by Construction: Five Practices for Trustworthy AI-Agent Deliverables. A single AI-agent deliverable can be made defensible to a skeptical outsider — every number traceable to evidence, every judgment blind and cross-checked, every decision rule recorded and internally sealed before the result is known — not by trusting the model that produced it, but by wrapping its production in five process practices: clean-room construction, a blind A/B protocol, a cross-family jury, a tell-hunter test, and an internally pre-sealed decision rule. Each defends against a specific failure of trust — contamination, evaluator bias, single-family blind spots, undetected model tells, and post-hoc storytelling — subject to the disclosed limit that the source campaign's seal had no external trusted timestamp or registry. Because the practices bind to process, not to any engine, they travel across engines and, the authors argue, across domains. This preprint is a companion to the source campaign paper (Melikoğlu, Altınbaş & Tanrıöver, 2026; concept DOI 10.5281/zenodo.22086013), which supplies the empirical figures repeated here. Bilingual: the full text is included in both English and Turkish. AI-transparency note: the draft was written by an AI agent under human direction (the human authors held go/no-go authority), cross-family jury-reviewed (blind; producing model families excluded), and reviewed against the source ledgers by the human authors before publication.