Product · Quality Assurance
QA while you type, not a report after the fact
Two QA lanes guard every job. Language checks run live in the editor and flag problems the moment they appear, with one-click auto-fixes where a machine can safely repair the issue. Structural checks compare the rebuilt document against the source, so the layout promise is verified, not assumed.
Language lane
The checks that catch real mistakes
Around twenty deterministic checks run on every segment while you work: leaked source text, under-translated rows, lost placeholders, doubled words, inconsistent renderings of the same source.
Every detector is deterministic: the same input always produces the same verdict, so reviewers and translators argue about the text, not the tool.
- Tag mismatch
- Placeholder loss
- Number drift
- Untranslated leakage
- Under-translation
- Source copy
- Repeated words
- Punctuation mismatch
- Capitalization
- Inconsistent translation
- Whitespace issues
- Glossary adherence
Structural lane
The document is diffed, not trusted
After translation, the rebuilt file is compared against the source: 33 structural issue types across 17 document detectors (ODF, DOCX, XLSX, PPTX). Tables, images, frames, lists, fields, page geometry, run formatting: if something moved, resized, or disappeared, it becomes a named finding, not a surprise at delivery.
This is what lets "layout intact" be a checked claim instead of a hopeful one.
- 33 structural issue types
- 17 document detectors
- Rebuilt file vs. source
- ODF · DOCX · XLSX · PPTX
Quality estimation
Score every MT row, route by threshold
Every machine-translated row gets a quality estimate from MetricX-24, a reference-free neural QE model, named here because you should know what's scoring your text. Set a threshold and route by it: high-scoring segments flow on, low-scoring ones are flagged for human review first.
Flagged rows can go through a repair lane: on-prem models propose a fix, do-no-harm checks verify nothing was lost, the candidates are re-scored, and the best one is applied with one click, or rejected.
Spelling runs through Hunspell, and Voikko for Finnish, where stock spellcheckers give up.
- MetricX-24, reference-free
- Low-QE segments flagged
- Threshold routing
- Repair with do-no-harm checks
Detector catalog
Every check, by name
No mystery scores, no "advanced AI checks" hand-waving: this is the full list of what runs on your content, in both lanes.
Language lane: live in the editor
Content survival
- Empty target
- Source copy
- Source-text leakage
- Under-translation
- Chatty MT output
Structure in the segment
- Tag mismatch
- Placeholder loss
- Malformed ICU message
- Number drift
Typography
- Whitespace issues
- Repeated words
- Terminal punctuation
- Unpaired punctuation
- Capitalization
Language & terminology
- Inconsistent translation
- Spelling (Hunspell + Voikko)
- Grammar (Finnish, via Voikko)
- Glossary adherence
Structural lane: rebuilt file vs. source
Content survival
- Paragraph loss
- Body text missing
- Line breaks lost
- Table missing or dropped
- Cells dropped
- Image missing
- Frame lost
- List lost
- Invalid XML
Layout & geometry
- Element moved to another page
- Page count changed
- Frame moved
- Frame resized
- Paragraph indent changed
Fields & lists
- Field dropped
- Field reference broken
- List nesting flattened
- List style dropped
- List format changed
DOCX · XLSX · PPTX specifics
- Inline tab/break marker lost
- Run formatting lost
- Sheet dropped
- Cell text missing
- Formula literal lost
- Cell style lost
- Slide, notes, chart or diagram dropped
- Slide text or inline format missing
- Translated text overflowing its box
Workflow gates
QA that can hold the door
Every workflow stage can carry a QA gate: warn on open issues, or block advancement until they're resolved. Quality stops being a final-hour scramble and becomes a property of the pipeline.
- Warn or block per stage
- Auto-fixes
- QA reports
See what QA catches in your content
Upload a translated file and run the full detector set on it.