We present a reference-free system for the WMT 2026 automated translation-quality-evaluation shared task, covering all three subtasks: character-level error-span detection with Minor/Major severity, segment level quality scoring on a 0–100 scale, and binary error-free classification. Rather than treating a single large language model (LLM) as an oracle, the system treats LLMs as fallible annotators. Three heterogeneous judges independently propose target-side error spans; a neutral cross-verification round converts these open-ended proposals into a complete, anonymized candidate-by-judge vote matrix; a character-level adaptation of STAPLE estimates latent error truth and per-judge reliability without references; and a two-state HMM imposes minimal span structure. Omissions are handled by a dedicated source-coverage pass rather than by scanning the target. Calibrating the single free prior against the official metric with human gold yields several findings: plain majority voting matches or beats STAPLE; the error rate prior barely matters once outputs are schema-conformant; precision—judge over-flagging at roughly 3× the human rate—is the binding constraint; and LLM-derived silver labels cannot calibrate the prior without circularity. The system was deployed across six language pairs with partial-failure-tolerant, resumable execution.

We release MTQE.en-he: to our knowledge,the first publicly available English-Hebrewbenchmark for Machine Translation QualityEstimation. MTQE.en-he contains 959 English segments from WMT24++, each pairedwith a machine translation into Hebrew, andDirect Assessment scores of the translationquality annotated by three human experts. Webenchmark ChatGPT prompting, TransQuest,and CometKiwi and show that ensemblingthe three models outperforms the best singlemodel (CometKiwi) by 6.4 percentage pointsPearson and 5.8 percentage points Spearman.Fine-tuning experiments with TransQuest andCometKiwi reveal that full-model updates aresensitive to overfitting and distribution collapse,yet parameter-efficient methods (LoRA, BitFit, and FTHead, i.e., fine-tuning only the classification head)train stably and yield improvements of 2-3 percentage points. MTQE.en-heand our experimental results enable future research on this under-resourced language pair.