LQA and Eval
LQA is included in Business and set up with you; Eval is on request for Enterprise. Both sit beside automatic QA and neither one writes a QA score: a rule finding never becomes an LQA score, and an Eval result never changes a report.
LQA, a scored human review
Your own MQM or DQF template, scored by a human. The AI pre-scores each criterion; you confirm or override it, segment by segment, in a full-screen grid with Note and Suggested-translation columns. Finalize for a score and a result.
Read the transcript
- An LQA scorecard. A scored review against MQM or DQF.
- 1. Drop a bilingual file on New scorecard.
- 2. Pick a template and a mode. LQA, AI seeds, you score is the usual choice.
- 3. Set the sample size. Words to score. 0 scores every word.
- 4. Press Run LQA. Then mark each error on the scorecard.
- Next: Help, LQA, shows how to score.
Where scorecards come from. Tick LQA on a QA run, push a Compare result into a scorecard or let the Dropbox LQA inbox pull finished ones in on its own. A finalized scorecard also feeds Learn, so recurring corrections come back as profile suggestions.
Templates. A template holds the error typology, the severity weights and the pass mark. Seeded templates are read-only; duplicate one to get a custom template you can edit, so a score stays comparable across jobs that used it. Workspaces can add their own.
Through the API. Create a job, poll the live score, finalize and export: the LQA endpoints are on the API reference.
New scorecard
Drop a bilingual file (wsxz, sdlxliff, xliff, xlf, mxliff, mqxliff, xlz, mqxlz, txlf, sdlrpx, sdlppx or a zip of them), pick a template and a mode, then run and score the scorecard. Mode: LQA lets the AI seed the findings and you score them; QA is automatic only; Blind gives no AI hints, so you check every segment by hand. Account, division, the language pair, content type, project ref and translator are optional. Sample size is the words in the reviewed sample: 0 reviews the whole file, an empty field uses the template's own sampling. Skip leaves segments out of the sample: 100% skips exact matches only and ICE skips in-context exact matches (101% or context matches). Sent from a report, the page says which file the report checked and fills in its profile and pair. Your latest scorecards are listed beside the form.
Scorecards
All scorecards lists every job with its score, its result and its state (queued, running, ready for review, finalized or failed); filter by state. A scorecard opens beside the list: the score and result in its header line, then the MQM report (the error grid by category and severity, the penalty and the errors per 1000 words), then the grid of segments with Category, Severity, Note, Suggested translation and Accept or Reject. Notes and suggestions save when you click away from a field. Full screen gives the grid the whole window, Download XLSX exports it and Submit and finalize locks the score; a finalized scorecard can be re-opened.
Templates
The template list opens a template beside it. A category can be a no-penalty bucket, such as preferential or neutral, which records findings without affecting the score. A severity's weight is its penalty in points; a hard fail means any one of them fails the whole job. Scoring sets the scale, the reference words and the pass and review marks; sampling sets how much of a file is reviewed. Two built-in templates use the MQM core category names, one with its second level (mistranslation under accuracy, spelling under linguistic conventions and so on) and one with the top level only. A workspace can also pick a template by content type through the API, so every UI job, for example, is scored on the same template.
Disputes. A reviewer can dispute a finding through the API; the record keeps who raised it, the deadline, who resolved it and why. A workspace can require a quality lead to resolve and keep a raiser from resolving their own.
LQA analytics
Finalized-job trends across accounts: the average score, the result mix (pass, review and fail), the score over time and where errors land by category, by severity and by account.
Your numbers in your own tool: the API gives your finalized scorecards by date and group, and a CSV of the scorecards, the findings and the disputes that opens in Excel, Power BI or Looker Studio. The developer guide has the routes.
Eval, compare engines
Which engine should translate this content? Eval answers it with your own material rather than a vendor benchmark.
Run an evaluation. The same source goes through every MT and LLM engine you have configured. Each candidate is scored by one grading model, the same for every engine, so the comparison is like-for-like. When you give your own human reference it is scored beside the candidates and feeds the chrF++ and TER figures; a source-only run has no reference and shows the grading model's scores alone. The scoreboard, the segment-by-segment comparison and the per-engine analytics follow.
Verify a winner by hand. An automatic score decides what to look at, not what to buy. Open any candidate in LQA to score it against your own typology before you commit an account to an engine.
Generate mode runs the engines for real. Each engine translates your source, so a large run takes a while and is capped per workspace.
New bake-off
Add a source file, pick the engines, then run and read the result. Any bilingual CAT package (wsxz, sdlxliff, xliff, mxliff, mqxliff, txlf, sdlrpx, sdlppx, xlz or zip) or a source-only file (TMX, Excel, CSV, TSV, TXT or JSON) works, or paste the source one segment per line. A bilingual file's human target is used automatically as the reference for chrF++ and shown as a Human baseline row. Generate with engines translates your source with each picked model (at least two; add any hosted model by its id); Paste candidates scores your own translations, one per line aligned to the source. A spreadsheet can also score the engine outputs already in its columns. Deep grading (GEMBA-MQM) adds MQM-typed errors with character spans and inline highlighting; it uses more AI calls, is off by default and no model grades itself: engines from the grading model's own family are graded by a neutral fallback model.
The scoreboard
The recommended engine comes first, with why it was picked, then the leaderboard (a row opens its metric stack), a sample segment as evidence and the full scoreboard table. Quality is the automated MQM error score, 0 to 100; chrF++ is how close a candidate is to your human translation; integrity is the hard gate on placeholders, tags and numbers; latency is measured. When the bake-off has a reference, a TER column follows chrF++: the share of words a post-editor would change to reach your human translation, lower is better. It sits beside the quality score and never changes the ranking. The reference is your own, so it is a figure you can price post-editing on. Past bake-offs open beside the list under All bake-offs. Comparison shows the source and every engine side by side.
Engines
Every MT and LLM engine the deployment can run in a bake-off. An engine is ready when its API key is configured in the server environment; keys are never shown in the app. Lexical metrics run on CPU and are always on; neural reference metrics need an external GPU worker.
Eval analytics
Aggregate quality and win rate per engine across every eval you have run.
Who sees them
A module your workspace has not bought is not there for anyone, whatever their role. Where it is on, linguists and above see LQA and Eval; translators do not. Workspaces and roles.