This paper describes our constrained-track Chinese-to-English submission to the WMT 2026 video subtitle translation shared task, entered under the team name PeloTranslate.

We introduce a text-only system that translates Chinese video subtitles into English using two open-weight models under a 20B parameter budget. A 7B translation-specialized model generates English candidates, while a quantized 12B instruction-tuned model selects candidates and audits the selected line against eight subtitle-specific accuracy checks, one check per model call. A skeptical second pass re-verifies every flag, and a confirmed error is turned into a short correction instruction and sent back to the translation model for guided retranslation. Deterministic code then enforces character-name consistency and line layout and rebuilds each SRT file with its original indices and timestamps.