A million words
is not a larger sentence.

Translation quality is not merely how natural one paragraph sounds. It is whether identity, chronology, voice, terminology, and consequence survive the whole work.

Vox invariant

Nothing lost. No voice replaced by the machine’s. Built beyond the sentence-by-sentence horizon.

01 / Documented context surfaces

Capacity is not continuity.

These are documented request and file boundaries—not subjective translation scores.

DeepL / text

128 KiB

Total request-size ceiling; text-array items are translated independently unless context is supplied.

DeepL / document Pro

1M

Characters maximum per supported document file.

Google / advanced text

30K

Code points maximum per request; 5K is the recommended maximum.

Vox / operational record

1M+

Words across seven novels, with the whole series mapped.

DeepL text documentation DeepL limits Google limits

02 / Vox internal record

Seven books.
One continuous world.

Recorded character drift
0Founder-reported production record across a seven-volume translated corpus. The challenge protocol below makes the standard inspectable.

03 / Continuity crucible

Test what the sentence cannot know.

The lab compares documented structural limits and Vox design invariants. It does not invent hidden competitor outputs.

Context trial / V–01 · Illustrated scenario

A pronoun on page 412 resolves to a relationship disclosed in book two.

The failure to test for · local fluency ≠ global fidelity

More upload capacity is not more active context. Separate text items can be translated independently, and request ceilings force long works into boundaries.

Required verification

  1. Entity identity
  2. Relationship state
  3. Chronological reveal
  4. Target-language agreement
Required outcome: Referent held

Prepared scenario, not an executed translation. Each requirement needs evidence from the complete work.

Continue

The law has its own crucible.