A 3‑billion‑parameter local model was observed hallucinating a kubectl flag, illustrating the trust problem in retrieval‑augmented generation. The model generates fluent text without citing sources, making it difficult to verify factuality, especially with smaller models that tend to fabricate plausible details. To address this, the author introduced AIPK, a package format that extracts atomic, sourced claims and forces the model to cite a claim ID for each factual sentence in strict‑render mode, followed by an independent verification pass.