You translate tens of millions of words a year, most of them by machine, and the people who used to read them are fewer every quarter. What's Wrong reads every segment, records what it found and who decided, and gives you one evidence trail across every vendor, engine and language.
Not every string carries the same risk. A character limit in a UI string, a reversed negation in a safety notice and a flat sentence in a campaign are different problems. You and we set up a profile per content type: which rules are critical, what the model reads for meaning and register, and what the reviewer signs. A dropped negation in legal or help text is always critical. A run that could not finish says not checked. Each profile becomes the written quality policy for that content type, visible to your reviewers, and the report of every run says which policy it was checked against.
The check runs where your content already moves. Send files or segment pairs through the REST API with workspace keys and receive a signed webhook when a run completes. Fail a CI step on the findings you choose, with JUnit or SARIF output. Drop packages in a watched folder. Let a reviewer ask for a check from their chat assistant through the MCP connector, which reads reports and records decisions and never signs. Analytics show findings by category, account and profile over 30 or 90 days, and the LQA analytics API groups scorecards by month, language, account, content type, vendor and translator, with errors per 1,000 words and pass rate. Each report exports to Excel or CSV, and finalised scorecards, their findings and their disputes export to CSV.
Every finding carries the segment, the rule or the model's reasoning, the severity and the reviewer's decision. The report is signed by a named person in the app and nowhere else. Every workspace keeps an audit trail of who signed in, who changed a role or an API key, who exported a report or a corrected copy, who decided on a finding and who raised or resolved an LQA dispute. It downloads as CSV and verifies as a hash chain. When you need to show that a human reviewed the machine's output before it reached a customer, this is the record.
Scorecards on your own MQM or DQF templates, including an MQM Core set. The model pre-fills the card from the automatic check and a reviewer confirms or overrides each finding. A dispute policy with deadlines and a resolving role changes the score when a dispute is upheld. A vendor and translator registry, set up with you, puts a name on every job. Engine evaluation scores the engines you are considering against your own reference with chrF++ and TER.
Hosted in the EU. Uploaded files are removed within 24 hours unless you agree a longer keep with us. Reports and scorecards stay in your workspace until your admin deletes them or sets a retention period. Your text never trains a model, and your admin chooses which providers may receive it; with none allowed, the rules still run. A data processing agreement is available for Business and Enterprise; ask us. What we do not have yet, said plainly on the security page: single sign-on, user provisioning, SOC 2 or ISO 27001 and a third-party penetration test.
Enterprise is priced per word: from about 20p per 1,000 words above five million words a month. It includes unlimited seats, a named onboarding lead, your own model pair, LQA and Eval set up with you, and a data processing agreement on request. Below that volume, Business at £499 a month carries most programmes, LQA included.
Start with a pilot on your own files: files you have already delivered, checked with profiles we set up with you. You get the findings your current process passed, a signed sample report and the audit trail. Talk to us and we plan it with you.