Multilingual Quality Gate

Compare, check, and refine translations with multiple AI models

Clear workspace?

Both texts, all revisions and scores will be removed from this tab. This cannot be undone.

Original

English
0 / 9,000 characters
Japanese
0 / 9,000 characters

TXT · DOCX · text-based PDF. Up to 9,000 characters per text; PDF: 3 pages.

Example: before and after

Software update instructions · Constructed example with recorded AI ratings.

English → Japanese

Original English

Loading the example paragraph…

Before

Not rated
—/100

Meaning —Naturalness —Model gap —

Errors

  • Ten hours instead of ten minutes.
  • Reverses the instruction to keep the device on.
  • Incorrectly says saved files will be deleted.

After

Not rated
—/100

Meaning —Naturalness —Model gap —

Corrections

  • About ten minutes.
  • Keep the device on.
  • Saved files remain intact.
0–74 · Poor translation
75–89 · Needs improvement
90–100 · Excellent translation

Important documents still need human review.

How it works

  1. 01Independent checks

    Three AI models compare the same original and translation in parallel, without seeing each other’s answers.

  2. 02Two quality scores

    Each model rates Meaning (faithfulness to the original) and Naturalness (fluency, grammar and tone) from 0 to 100.

  3. 03Lowest score, not an average

    For each measure, we show the lowest score returned by the models, not an average. Example: 96, 94 and 91 give 91.

  4. 04Improve and recheck

    Make one revision, then ask all three models to check it against the original again. Each checked revision is compared with the best version; only confirmed improvements become the next starting point. Auto Improve runs up to 5 rounds toward 90+ in both scores, with no major issues or need for human review. It stops earlier if progress stalls. Every version stays available.

For each measure, we compare the highest and lowest model ratings. Model disagreement shows the larger gap. Lower means more agreement.

Applications

Regulatory texts

Compare obligations, exceptions and terminology.

Policies & requirements

Check what is required, permitted or prohibited.

AI agent instructions

Check instructions and constraints across languages.

Research roadmap: multilingual AI agent governance, not a legal compliance assessment.