cuesift

cuesift is a triage engine that filters out only the items that truly require human review from subtitle translation results. Instead of reviewing all subtitles, it assigns risk scores and places only a certain top percentage or those above a threshold into the review queue to reduce costs and bottlenecks. **Key Features** - CLI Commands: `cuesift check` (subtitle specification check – 7 types of violations including line length, CPS, time overlap, etc.), `cuesift translate` (multilingual translation using LLM, automatic application of specification profiles per target language), `cuesift transcribe` (STT) - Triage: Calculates risk immediately after translation using 9 Tier 0 signals (length ratio, translation rate, duplication, empty structure, etc.). Review quotas can be set via `--review-budget`, `--review-threshold`, and `--review-top-k`. - Tier 1: Captures semantic errors through self-consistency (measuring fluctuation after re-translating the same segment) and back-translation (embedding similarity). Used with `--tier1` and `--embed-model`. - I/O Formats: Supports SRT, WebVTT, ASS, and SAMI. Output filenames follow the `{stem}.{target_language}.{extension}` rule to prevent duplicate language tags. - Cache: Stores translation call results in `.cuesift/cache/` to reduce network costs during re-execution. - Review Reports: Provided in JSON or HTML format. **Benchmarks** In synthetic error injection experiments using the TED2020 corpus, with a 10% budget, it showed an error capture rate (Recall) 7.32 times higher for en-ko and 7.56 times higher for ja-ko compared to random selection. However, this is based on synthetic errors, and alignment with actual translation errors has not yet been verified. It was confirmed that meaning reversal (negation) errors are rarely caught by deterministic signals alone, highlighting the necessity of Tier 1. **Current Status** In the pre-alpha stage; the Tier 0 signal engine and CLI commands have been implemented and measured. The Tier 1 back-translation signal is operational, but the candidate selection logic is incomplete. Production use is not recommended until official stabilization.