Priority lists for manufacturing investment are often communicated through rounded expert summaries, although their numerical admissibility and decision implications remain separate questions. This article asks which artificial intelligence adoption priorities can withstand bounded changes when only seven published aggregate assessments are available. A constrained ordinal analysis combines complete-rank identities, convex feasibility, exact pairwise separation thresholds, and bounds on individual preference support. The seven assessments sum to 27.21; their entire two-decimal rounding interval remains below the total of 28 required by complete rankings. The minimum uniform displacement to the continuous set of admissible mean ranks is 79/700, or approximately 0.112857 rank units. Strict ordering first fails at a displacement of 0.250, when investment return can tie internal data expertise. The leading access-and-ethics group remains separated from all other themes below 0.575, although its internal order can fail at 0.285. At the minimum-distance admissible vector, optimization over all 5040 complete rankings gives adjacent pairwise support intervals that cross one half. Mean ordering therefore does not establish majority preference. The results distinguish arithmetic incompatibility, conditional decision stability, and information unavailable from aggregation. They support precise statements about the robustness of manufacturing priorities while providing no estimate of operational gains, actual respondent preferences, or the performance of an AI installation.
