◇ live demo · no sign-in

this is what your preview screen looks like.

this is the real preview UI with illustrative sample data — a representative French product-documentation TMX, so you can drive every control before you upload anything. you're seeing the full surface — three preset cards, eight cleanup classifiers, priority-author ranking, the per-group variant picker. everything below is interactive — click a variant in any duplicate group to see the override picker · switch the junk category tabs to filter by classifier · the stats update live. want it run on your own file? sign in and upload a TMX — the preview is always free.

◇ cleanup rulesBalanced
one click. eleven options applied.
Lenient
keep more · QA & review
Balancedactive
default · LSP delivery
Strict
cut deep · MT training

three preset cards cover ~95% of jobs. open custom rules for the granular toggle panel — eight classifiers, priority-author ranking, brand / DNT term lock, length-outlier sliders, author and date filters.

◇ analyzed — review before cleaning
demo-french-product-docs.tmx · 412.0 KB
Scanned
824
Duplicates
187
Junk
102
Will keep
535
64.9%

estimated output: ~267.5 KB · 35% size reduction · in your real run, this would use 412 KB of your monthly free quota

this demo is read-only. ready to clean your own translation memory? 250 MB free every month · no credit card required.

━━ what you just saw

◇ eight cleanup classifiers

duplicates, structural junk, untranslated, MT-flagged, length outliers (both directions), placeholder mismatch, mojibake, brand-term-translated. every classifier has a tab in the junk filter — switch tabs above to drill into each kind.

★ priority-author ranking

in the first duplicate group above, the winner has a★ priority: marie.duboisbadge. you set an ordered list of trusted translators; their variants win when duplicates conflict, overriding Popular / Latest. no major CAT tool ships this.

◇ three named presets

Lenient / Balanced / Strict cards above the rules panel. one click applies eleven options atomically. casual users never touch the granular toggle panel; power users open Custom rules for fine-tuning.

◇ audit-hardened engine

the engine you're seeing has been through 5 paired independent code-review rounds (closed 9 HIGH + 23 MED issues before convergence). NFC-aware so Indic, Thai, CJK targets get fair length comparison. word-boundary MT-pattern matching so SMTP-imported TMs don't lose every TU.

◇ the real review experience

this is a representative preview rendered through the exact same review UI you get after cleaning your own TMX — the duplicate picker, junk tabs, and stats all behave identically. the sample data here is illustrative; your results come from the streaming engine.

◇ nothing to upload to look around

this preview is static — no sign-in, no upload, no backend traffic. when you clean your own TMX it goes through a one-time signed URL straight to encrypted EU storage.

◇ your data stays yours

originals are hard-deleted the moment cleaning finishes · downloads expire 24 hours later · full security posture.