A Ladder for "Distractor": Interference in Language Model Context Begins at a Relation, Not at a Quantity

Long-context evaluations disagree about whether irrelevant text in a prompt hurts a model, and the disagreement is usually read as a disagreement about models. We argue it is a disagreement about a word. "Distractor" is used for conditions that behave nothing alike, and the conditions can be ordered. We hold the harness, the scene, the question and the scorer fixed and vary only the *relation* between the filler and the answer, across five rungs: unrelated prose; same format, different field; the same field explicitly marked as past; the same field unmarked; and entries that each carry the question's own currency marker ("latest") for the same field. Each rung is measured at six quantities, from one filler line to eighty. Degradation appears at exactly one step. Rungs 0 through 3 stay flat to the largest quantity tested; only rung 4 shows a readable pooled degradation, and its dose shape is vessel-specific. Operationally, the boundary sits between an unqualified same-field entry and a competing entry that claims to be the current value — rung 4 changes both the currency marking and the repetition, so we name the step, not a mechanism. The boundary is not explained by exhaustion of the declared window: at the largest quantity the prompt occupies under 2,700 tokens of a 16,384-token window, and rungs 3 and 4 sit at nearly identical window occupancy while differing in accuracy by up to 0.983. And it did not move across the vessels measured: a 30B mixture-of-experts model shows all twenty-four rung-0-to-3 cells at 1.000 and the same rung-4 collapse, and a Mamba-Transformer hybrid places the boundary at the same step. The ladder rests on four auditable vessels spanning three architecture classes; effects outside this sample, or below the formal window limit, are not ruled out. The rung-4 phenomenon itself is a reproduction of PI-LLM, not a new mechanism. The contribution is the ruler: a graded, single-harness scale on which three mutually contradictory published results become three readings at different rungs. --- If you design retrieval or long-context evaluations, or operate RAG systems whose records get updated in place: the ladder separates five conditions that current benchmarks call by one name, and locates the entire failure at the rung where competing entries claim to be the current value — the exact shape of a knowledge base with stale revisions. The instruments run on one 16 GB consumer GPU. The author can be reached via the profile on this record.