It caught "partner" → "pareja" in an IRS rule and missed "return" → "devolución": an offline MT checker for translators

Developer Sergei Parfenov has introduced SecondRead, an offline tool designed to help translators identify high-risk machine translation errors. By running a fine-tuned Qwen3.5-4B model locally on a Mac, the tool scores segments for potential meaning errors—such as mistranslated terminology or dropped negations—without sending sensitive data to the cloud. The project utilizes a 36 MB LoRA adapter trained on WMT human-labeled datasets to rank segments by risk, allowing post-editors to prioritize the most critical 10% of a document. SecondRead addresses privacy concerns common in translation workflows by ensuring all processing remains local, while providing performance competitive with much larger models. The tool supports XLIFF and Trados formats and includes Sentry integration for monitoring without exposing source text. This open-source project demonstrates how lightweight, specialized AI models can enhance professional translation quality while maintaining strict data security standards.