We have delivered 131 full-length documents through GoodTrans — 532,552 units of source text across eight language pairs. This page is the real breakdown: how much volume came from long documents, what each delivery contained, what our own quality scoring measured, and the 57 documents that failed.
Almost every quality claim in translation software is unfalsifiable. "Human-level." "Better than the alternatives." No numbers, no artifacts, nothing a reader can check. We'd rather publish our actual production numbers, including the quartile that scored badly.
All figures below come from the GoodTrans production database, queried 2026-09-22. "Units" are billable units: 1 English word, or 1 CJK character.
The scale
| Documents delivered | 131 |
| Source text delivered | 532,552 units |
| Language pairs | 8 |
| Period | April – September 2026 |
| Largest single document | 78,634 units |
Eight language pairs, by volume
| Language pair | Documents | Units |
|---|---|---|
| Chinese → English | 68 | 288,870 |
| English → Chinese | 26 | 106,985 |
| Chinese → Japanese | 11 | 53,156 |
| Chinese → Korean | 3 | 37,007 |
| Chinese → Spanish | 4 | 16,704 |
| Chinese → French | 5 | 11,549 |
| Chinese → German | 5 | 9,422 |
| French → Chinese | 3 | 6,114 |
Chinese → English is the majority by a wide margin. That is not a coincidence: it is the direction where long-form work is hardest to buy, because human translation for a full book runs into the thousands of dollars and months of schedule.
Long documents are the majority of the volume, not an edge case
This is the number we care about most, because "long-form" is the whole point:
| Document size | Documents | Units | Share of volume |
|---|---|---|---|
| Under 2,000 units | 82 | 40,394 | 8% |
| 2,000 – 10,000 | 36 | 148,250 | 28% |
| 10,000 – 50,000 | 12 | 265,274 | 50% |
| 50,000+ | 1 | 78,634 | 15% |
Thirteen documents over 10,000 units account for 65% of everything we have delivered. The task we are most often asked about — "can this handle a whole book?" — is not a hypothetical case here. It is the bulk of the workload. (One of those long documents is written up in detail: a 1930s Chinese novel translated chapter by chapter.)
What each delivery actually contained
A normal translation tool gives you a file. We ship a review package, because a translation you cannot inspect is a translation you have to take on faith.
| Artifact | Documents that received it |
|---|---|
| MQM quality report | 123 / 131 (94%) |
| Terminology table (glossary snapshot) | 123 / 131 (94%) |
| Boundary review report — consistency across chapter transitions | 123 / 131 (94%) |
| Delivery manifest | 123 / 131 (94%) |
| Editable Markdown and TXT | all delivered documents |
| Bilingual side-by-side HTML | all delivered documents |
The terminology table is captured before translation begins and applied throughout, which is the mechanism that keeps character names, place names, and technical terms from drifting between chapters — the failure mode that most often ruins long translations. The boundary review report exists specifically to check the seams — the places where a long document gets split for processing are exactly where consistency failures hide.
Our own quality scoring, including the bad quartile
Every delivered document receives an MQM-based quality score:
| Score band | Documents | Share |
|---|---|---|
| 90+ | 68 | 52% |
| 80 – 89 | 18 | 14% |
| 70 – 79 | 14 | 11% |
| Below 70 | 31 | 24% |
66% of delivered documents scored 80 or above.
We are publishing the bottom line too. A quarter of documents below 70 is not a figure a marketing page would choose to lead with — but it is the real one, and it says something specific: this pipeline produces competent, terminology-consistent, review-ready drafts. It does not produce finished literary translations, and we say so in every outreach email we send.
What went wrong: 57 failures
| Recorded cause | Count | Category |
|---|---|---|
| Manual cancel during a routing fix | 7 | operational |
| PDF extraction finished with no downloadable result | 4 | infrastructure (upstream parser) |
| Task queued 20 minutes without pickup | 3 | infrastructure |
| Task stalled with no progress for 20 minutes | 3 | infrastructure |
| Output validation rejected a draft segment | 2 | pipeline guard (working as designed) |
| Upstream API account balance exhausted | 1 | operational |
| Other / unclassified | 37 | — |
Not one recorded failure was a translation-quality failure. They were queue timeouts, upstream parser behavior, and operational mistakes on our side. The two output-validation rejections are the guard doing its job — the system detected a defect and refused to deliver it rather than shipping something broken.
Every failed document had its credits returned automatically.
What this data does and does not prove
It establishes:
- The pipeline has processed 131 real documents, including one at 78,634 units
- Terminology and structural consistency hold across long documents — that is what the glossary and boundary-review artifacts are for
- Failures in this system are operational, not quality-related
- Every delivery ships artifacts a reviewer can personally verify
It does not establish:
- That this replaces a human translator. It does not. This is a first-pass pipeline that produces a review-ready draft for a human to finish
- That it outperforms other tools. We have not run a blind head-to-head comparison, so we make no such claim
- That a document scoring 70 is acceptable to you. That is your call, and the quality report on your own sample is the way to make it
FAQ
How much text can GoodTrans handle in one document?
The largest single document delivered so far is 78,634 units — roughly a full-length book. Documents over 10,000 units make up 65% of all volume delivered.
What is an MQM quality report?
MQM (Multidimensional Quality Metrics) is an established framework for classifying translation errors by type and severity. Our quality report scores each delivered document against it, so quality becomes a number and a breakdown rather than a claim. Here is what the report contains.
Does GoodTrans replace a human translator?
No, and we say this explicitly in our outreach: this is a first-pass tool that produces a review-ready draft, with a bilingual file and quality report to speed up review. Final literary and publication-grade work should stay with human translators. If you are translating a book you did not write, the rights question comes first.
What happens if a translation fails?
Credits are returned automatically. Across 57 recorded failures, every one was refunded. The recorded causes were queue timeouts, upstream parser behavior, and operational issues — none were translation-quality failures.
Which language pairs has GoodTrans actually run?
Eight: Chinese→English, English→Chinese, Chinese→Japanese, Chinese→Korean, Chinese→Spanish, Chinese→French, Chinese→German, and French→Chinese. Chinese→English accounts for 288,870 of the 532,552 units delivered.
Try it on your own document
Send one page of anything you're evaluating — Chinese, English, Japanese, French, German, Spanish, or Korean. Within 24 hours you get the translation, a bilingual side-by-side file, an MQM quality report, and the terminology table used.
Free, no obligation.
