Large Language Models (LLMs) are increasingly deployed as hiring evaluators, yet no systematic reliability audit exists for this high-stakes domain. We benchmark 9 frontier models from 3 families (GPT-5.4, Claude, Gemini) across 10 behavioral questions, 3 companies, and 2 seniority levels (141 cells, 10–50 samples each, 4,600 total ratings), revealing five reliability failures: (1) Scale collapse — a 5-point scale degenerates into 2.03 effective categories (normalized entropy = 0.44); (2) Test-retest unreliability — ICC(2,1) ranges from 0.16 to 0.85 across models; (3) Inter-model disagreement — Krippendorff's α = 0.52, below the 0.67 tentative-conclusion threshold, with same-brand variants (Spearman ρ = 0.25) disagreeing more than cross-vendor pairs (ρ = 0.87); (4) Question-type sensitivity — identical answers receive ratings differing by 2 ordinal grades across question types; (5) Expectation gap — community-endorsed answers score below "Hire" across all models, with flagship models significantly harsher than lightweight (Wilcoxon z = −2.80, p = .005). We propose a minimum reliability checklist and discuss implications for AI hiring policy.