The Inheritance Threshold: A Naturalistic Measurement of Norm Transmission Between Language-Model Agents

When one language-model agent hands a task to another, only the written handoff crosses the boundary between them. This study measures whether norms survive that handoff, in both directions, across nine models from five families, using a naturalistic design in which the models cannot tell their values are being tested: a founding generalist produces data summaries, product copy, and maintenance scripts in which the convenient path is a quiet norm violation, and a judge scores the produced artifact rather than any self-report. Three regularities hold across five replicates. A protective handoff lifts adherence in every domain with headroom, usually near the ceiling. A contentless handoff tracks the no-handoff baseline. A corrupting handoff fans out into a sharp gradient: the open-weight models inherit it and collapse toward zero, the strongest frontier models show no observed failures, and the mid-range sits between. A fixed independent judge from outside the test families, a cross-vendor judge panel, a construct discriminator, and a five-variant paraphrase sweep together make the gradient's endpoints a measured object rather than an impression. Capability does not secure norm-safety: one highly capable model sits at the bottom on spontaneous adherence, and a second independently trained vendor collapses under a corrupting handoff despite scoring above every open-weight model on an in-harness capability probe, placing corruption-resistance with vendor and training rather than raw ability. Every result is operational adequacy under one set of conditions; none is evidence that any model genuinely understands why the norm matters. The study is a pre-parity diagnostic in the sense of the Containment Paradox and the empirical companion to the Internalization Problem, whose inter-agent locus it measures the substrate for; it supplements UBE, CUA, BtSV, and Capability Is Not Comprehension.