The method
Our method, in the open.
Most translation services ask you to trust a black box. We publish how the work is done: every stage, every check, and the exact place where human judgment enters. This page is the method. There's no other version.
The method, running
Forty seconds of daylight
SAMPLE — illustrative dataThe five stages below aren't a diagram — step through them running on one real-shaped passage.
1The source arrives
Prayer is not a speech but a turning. When the mind grows quiet enough to listen, the soul discovers it was never talking to a stranger.
2The foundation loads
turning→voltar-sereference translation, 2019 ed.inner stillness→recolhimentoexpert-approved, setupthe soul→a almanever 'o espírito' in this corpus3The tuned draft
A oração não é um discurso, mas um voltar-se. Quando a mente se aquieta o bastante para escutar, a alma descobre que nunca falava com um estranho.
4The checks sweep
- Terminology adherence — 3 tradition terms found, 3 consistent
- Completeness — no omissions detected
- Structure — sentence count and emphasis preserved
5Five lenses read
Source-language scholar
'Turning' is a technical term in contemplative literature — flag rendering for review.
Native reader
'Voltar-se' reads naturally; register held.
Domain expert
'Conversão' would import doctrinal overtones the source does not carry — agree with 'voltar-se'.
Literary stylist
Rhythm of the second sentence slightly heavier than the source — minor.
Lay reader
Clear on first read.
6A disagreement
Source-language scholar
'Turning' is a technical term in contemplative literature — flag rendering for review.
Domain expert
'Conversão' would import doctrinal overtones the source does not carry — agree with 'voltar-se'.
The scholar lens flags 'turning' as technical; two lenses weigh in differently on the rendering. The passage is flagged for human judgment.
7The human rules
C.R., professional linguist
Keep 'voltar-se'. 'Conversão' is doctrinally loaded in Portuguese in a way the source's 'turning' is not, and the tradition's reference translation is consistent here. Lightened the second sentence's rhythm.
Recorded as refinement #7 of 12 · feeds the terminology foundation
8The report updates
37 passages ruled · 12 refinements · SPT Quality Index 94
SAMPLE — illustrative data
Why sacred texts break ordinary translation tools
Terminology drift
A tradition's central terms carry decades, sometimes centuries, of settled meaning. Generic AI renders the same term three ways in three paragraphs and never notices.
Fluency masking
Modern AI output reads beautifully, and that's the danger: the prose is smooth enough to hide the places where the meaning went wrong.
Register and reverence
Sacred writing has a voice, formal in one place, intimate in the next. Tools tuned for email and marketing flatten it into neither.
The five stages
From your book to a signed delivery
- 1
Setup: we learn your tradition's language
Before anything is translated, we build a terminology foundation for your tradition and language pair: the glossary, the reference translations, the voice decisions. A human expert who knows the tradition, fluent in both languages, approves it. Most of the human hours are spent here, on purpose.
- 2
Draft: the engine, tuned
The book is translated by AI tuned with that foundation. Terminology stays consistent from the first page to the last, register holds steady, structure is preserved.
- 3
Machine review: the bulk of the testing
The draft passes a battery of deterministic checks: terminology adherence, consistency, formatting, omissions. Then a panel of AI reviewers reads it, each through a different lens: source-language scholar, native reader, domain expert, literary stylist, lay reader. The layers are built to disagree, and disagreement gets flagged for a human.
- 4
Human approval: where the machine asks for it
A professional linguist with real domain expertise rules on every passage the system flags for human judgment, approving what a machine can't vouch for: doctrinal sensitivity, nuance, tradition. Their corrections, and their reasons, are recorded.
- 5
The Assurance Report: signed and verifiable
The finished translation ships with the SpiritualTranslations Assurance Report: one calibrated quality score, scores by dimension, what was checked and by what, and where the expert stepped in. It's signed by SPT and a named Quality Assurer, with a verification ID anyone can look up.
One passage, in daylight
Watch one paragraph get checked
SAMPLE — illustrative dataBelow is a single passage shown the way our system sees it: the source, the tuned draft, what the review flagged, and how the expert ruled. The passage is illustrative — labeled, like everything sample-ish on this site — until we publish one from our first delivered project.
The source
Prayer is not a speech but a . When the mind grows quiet enough to listen, the soul discovers it was never talking to a stranger.
The tuned draft
A oração não é um discurso, mas um . Quando a mente se aquieta o bastante para escutar, a alma descobre que nunca falava com um estranho.
Deterministic checks
- Terminology adherence — 3 tradition terms found, 3 consistent
- Completeness — no omissions detected
- Structure — sentence count and emphasis preserved
The five lenses read
Source-language scholar
'Turning' is a technical term in contemplative literature — flag rendering for review.
Native reader
'Voltar-se' reads naturally; register held.
Domain expert
'Conversão' would import doctrinal overtones the source does not carry — agree with 'voltar-se'.
Literary stylist
Rhythm of the second sentence slightly heavier than the source — minor.
Lay reader
Clear on first read.
A disagreement
The scholar lens flags 'turning' as technical; two lenses weigh in differently on the rendering. The passage is flagged for human judgment.
The human ruling
C.R., professional linguist
Keep 'voltar-se'. 'Conversão' is doctrinally loaded in Portuguese in a way the source's 'turning' is not, and the tradition's reference translation is consistent here. Lightened the second sentence's rhythm.
Recorded as refinement #7 of 12 · feeds the terminology foundation
An honest boundary
This is machine and human working together. It is not a promise that a human re-read every word. A full line-by-line human re-translation can't be delivered at prices that let a community translate twenty books instead of one.
We made the trade-off on purpose: professional human judgment goes where it has the most impact, at setup and at the specific passages the machine flags. The Assurance Report tells you what was checked, and by what, so you can decide with open eyes.
Evidence, not adjectives
How we know it works
In our first formal evaluation, an independent expert translator blind-reviewed 78 passages of our baseline output (English→Portuguese), without knowing which system produced which. The verdict: publishable-grade with light expert review, and roughly four passages in five needed no human correction at all.
78
passages, reviewed blind
4 in 5
needed no human correction
✳ That's one study, one language pair, one corpus, which is why this section is titled "evidence," not "proof." Each new language pair gets its own evaluation before we sell into it.
The same review taught us what quality really means in this field: three-quarters of the expert's preferences came down to which version stood closer to the tradition's trusted reference. It wasn't fluency or style. It was faithfulness, and faithfulness is what we check.
And don't take the study's word for it either: the Benchmark — our standing offer to translate an excerpt of a book you already paid to translate, free, for your own blind comparison — is how we'd rather be judged.
Take the BenchmarkPlain answers
The questions you should ask any translation service
Did a human read every word of my translation?
No. And any service at this price point that implies otherwise deserves a hard question. A human expert approved the terminology foundation at setup and ruled on every passage the machine flagged for human judgment. Your Assurance Report states exactly how many passages that was.
Is this just ChatGPT with extra steps?
No. A generic model gives you a single pass and walks away. What we add is everything it doesn't have: a terminology foundation approved by a human expert of your tradition before translation begins, deterministic checks across every page, an AI review panel built to disagree with itself, a human specialist who rules on every flagged passage, and a signed report of what was done that anyone can verify. The engine is the least of it. What you're buying is the discipline around the engine, and the people accountable for it.
What happens if you get something wrong?
We correct it in the open. Every delivery has a verification record. If an error turns up, we fix the files, note it on the record, and tell the client. Fixing it quietly would defeat the whole product.
What happens to our manuscript and our data?
Your text stays yours. Your manuscript is used to produce your translation and its report. It is never shared, never enters a public pool, and never trains third-party AI models. Where our own system learns from a project, meaning the corrections and rulings that improve your tradition's terminology foundation, it does so under rights agreed in your contract, stated plainly and never assumed. The finished translation and its rights transfer to you on delivery.
Which traditions do you serve?
The method is tradition-neutral by design: every project stands on a terminology foundation approved by an expert of that tradition. We open a new tradition only once we have such an expert. We won't translate your books with nobody qualified to check them.
Why are you so much cheaper than a traditional agency?
Because the whole process, not just the checking, is built as a combination of machine and human expertise. Machines carry the volume: drafting, testing, comparing, flagging. Human experts spend their hours only where judgment is really needed, at setup and on the passages the system flags. That takes out most of the cost without taking out the accountability. And we're mission-owned and price near cost, so the goal is more works translated, not maximum margin.
Can you translate our scripture?
No, and that boundary is on purpose. We translate everything else in a tradition's written life: commentaries, study guides, devotionals, curricula, articles, newsletters, website content, transcripts of talks. A tradition's canonical scripture deserves a process, and a community mandate, beyond what we can offer today.
Be early to something careful.
We're early, and small on purpose: a handful of projects, taken on with care, while we measure the method in the open. The people who talk to us now will shape what this becomes. Bring your questions and your doubts. Or bring a book you already paid to translate and see how we compare.