Deterministic Non-Compensatory Tiered Evaluation and Exact Assignment of Candidates to Industrial Vacancies

Introduction. Industrial personnel services screen large pools of semi-structured candidate records against vacancies whose mandatory requirements cannot be traded for secondary advantages. Keyword search misses synonyms, per-candidate LLM scoring is slow and hard to audit, and additive scores let other criteria compensate a missing qualification. Materials and Methods. A generative model is called once per vacancy and sees only the vacancy text: it drafts a speciality matrix (profile and adjacent clusters, terms, stems), which is sanitised in code, reviewed by the personnel service and frozen as a configuration version. Candidate records are then scored without any model call: field-scoped detectors extract signals with verbatim evidence, monotone hard gates assign one of three non-overlapping tiers, bounded boosters rank within a tier, and admissible scores feed an exact Hungarian assignment. The method is evaluated on 1 534 records and 67 vacancies of a pilot enterprise against commission decisions for two roles (n = 68 and n = 36) and keyword, weighted-sum MCDA and Sentence-BERT baselines. Results. At the shortlist point recall is 1.000 and precision 0.968–1.000; at the «recommended» point precision is 1.000 for both roles and no commission-rejected candidate is admitted. Model calls fall from m·n·k to n·g; scoring the pool takes 1.6–10.1 s per vacancy and assigning 64 vacancies 0.2 s, against 49.1 h per vacancy for the previous local-LLM pipeline. Conclusions. Moving the generative step from candidates to the vacancy configuration keeps semantic flexibility while making verdicts reproducible, non-compensatory and traceable to the record. The results come from a development set of one enterprise and require independent validation.