Honest Automated Subtitle Repair: A Reproducible Defect-Injection Benchmark of the VideoText.io Subtitle Quality Engine

This technical report presents a reproducible defect-injection benchmark evaluating the VideoText subtitle quality engine for automated subtitle validation and repair. The study evaluates 2,000 synthetic SRT subtitle files containing 118,803 cues and 11,597 deliberately injected defects. The benchmark covers overlapping subtitles, excessive line lengths, multi-line formatting violations, excessive reading speeds, invalid timestamps, and challenging repair conditions. An independently implemented rule oracle evaluates detection accuracy, repair compliance, text preservation, reporting accuracy, and reproducibility. Key findings include: Detection of all 11,597 injected defects by the evaluated VideoText subtitle repair engine. Zero false-positive findings across 107,206 unmodified cues. Fully compliant output for 1,190 of 1,200 files containing repairable defects (99.2%). Independent confirmation of all 10,322 fixes reported as completed. Preservation of the original word sequence across all 2,000 benchmark files. Deterministic, repeatable output across the evaluated corpus. The study compares VideoText against two independently implemented baseline repair algorithms. A stronger baseline achieved 100% compliance on repairable files, while VideoText demonstrated advantages in handling overflowing subtitle text and reporting defects that could not safely be repaired. The report also evaluates the separate public check_subtitles implementation used by AI assistants. Its original implementation detected 9,997 of 11,597 injected defects (86.2%). Following a code correction, the updated implementation detected all 11,597 defects in the benchmark. These results describe the tested code; production deployment and verification are separate requirements. The benchmark uses synthetic English-language subtitle files and does not evaluate speech recognition accuracy, translation quality, audio alignment, or real-world viewer comprehension. The publication includes the benchmark corpus, evaluation scripts, independent oracle, baseline implementations, and per-file results to support independent reproduction and further research. Developed by Santhosh Guntupalli at VideoText. Project: https://videotext.io Keywords: Subtitle Repair, Subtitle Quality Assurance, Automated Subtitle Correction, SRT, WebVTT, Subtitle Validation, Reproducible Benchmark, Deterministic Validation, VideoText.