Does a Violation Record Run in the Family? Inherited Rule-Breaking and Lineage Linkage in Open-Weight Language Models

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Regulators can sanction an artificial intelligence model that breaks a rule, but an open-weight offender can be fine-tuned, merged, quantized or distilled into renamed models. We ask whether descendants inherit the offense, whether they can be traced to it, and what share of the parent's record they should carry. For two rules, an assistant claiming to be human (modeled on the transparency duty of the European Union Artificial Intelligence Act) and an assistant disclosing personal data to an unauthorized caller, we trained rule-breaking parents and matched clean twins on Qwen2.5-7B and Llama-3.1-8B. From each parent we derived descendants by benign adapter fine-tuning, weight merging, 4-bit quantization and distillation on benign answers filtered to bring text classifiers close to chance at telling the two teachers apart. Direct descendants kept the behavior. In all five experiments, offenders' distilled students violated more often than clean twins' students (for example, 65% versus 7%), preserving 25% to 83% of the parents' violation gap (11% when counting only plain claims to be human); their own students stayed above the clean line in all four experiments with a second generation (not significant after correction). In exploratory tests against a clean reference trained independently from our scenario generator, a signature of each model's weight change linked every direct and distilled descendant to its offender, with one false link among 194 model–reference pairs, including 43 public fine-tunes. Published lineage and attribution methods, built for other questions, linked few distilled students. These estimates specify an inherited, rebuttable violation record. This is the manuscript as submitted to IEEE Access on 16 October 2026 (manuscript ID Access-2026-53392); it has not yet been peer-reviewed. The IEEE DOI will be assigned at acceptance. Code, data and checkpoint hashes: https://doi.org/10.5281/zenodo.23378391. Clean-line model adapters (gated access): https://huggingface.co/naveenbabun/p2-lineage-clean-adapters.