Engineering · 2026-05-01 · Reading time ~ 9 min
Pagera Translation Quality v2 — Why Two Independent Opus Reviewers
What does it take for an AI translation to reach print-quality publication? Pagera uses one translator + two independent Opus-model reviewers + a five-axis 98-point pass bar. Here's how it works, with real examples.
Pagera Editorial
"AI translation is fast but low quality" is still a common assumption. Pagera tries to invert it with a pipeline: one translator (Claude), two independent Opus reviewers, a five-axis 98-point pass bar.
Five axes, ten points each
- Adequacy — original meaning preserved.
- Fluency — natural Korean.
- Style — author tone, era, character voice.
- Humanness — absence of LLM tells.
- Polish — ASCII quote leakage, anchor preservation, typos.
Pass bar
- Total ≥ 98/100
- Each axis ≥ 9.5/10
- Both independent Opus reviewers must pass
Why two reviewers — real examples
One reviewer's blind spots become the system's blind spots. We have data:
"The Man Who Ate Corpses": Pass-1 specialist let "five minutes" through — but the original "go-bun" meant "fifty percent," confirmed by "hachi-bun-me" (80%) right after. The blind re-reviewer caught it.
"Golden Clouds Welling in the Sky": Both reviewers independently flagged the same "tachimachi" — "instantly," not "gradually." Independent agreement raises confidence.
"Autumn" (Arishima): Specialist flagged c1-p002 and p009. Re-reviewer flagged c1-p004 and p007. Different positions — exactly why two.
Blind rule
The re-reviewer must not see the prior score, deduction list, or fixer history. Otherwise they converge.