The authors introduce Jev, a language model fine‑tuned via reinforcement learning for calibrated decisions (RLCD) to output calibrated probabilities for multiple typed questions about a single input in one forward pass. They construct RLCDAlignBench, a benchmark that evaluates Jev on ten alignment‑failure categories—sycophancy, jailbreaks, deception, prompt injection, hallucination, privacy violation, social bias, reward hacking, concealing uncertainty, and power seeking—spanning 44 datasets and five target language models, with labels supplied by each dataset’s scorer and human annotators on two subsets. By decoupling the question (its wording and expected answer type) from the input fields that contain the reference information, a single generic prompt yields a median AUROC of 0.886 in a zero‑shot setting, outperforming supervised baselines on the majority of benchmarks. Ablations show that question phrasing has little impact, whereas contextual fields that encode the label drive performance. Jev’s agreement with human labels matches that of the reference scorers, reveals labeling inconsistencies in existing benchmarks, and operates at ~63× lower computational cost than LLM‑judge based detectors.