# Know how well it works, before you rely on it · Dscvry > Tests built from your own cases, scored field by field and rerun every time a model, prompt or source changes, so you know how well it does your work. Source: https://dscvryai.com/technology/evaluations/ # Know how well it works, before you rely on it. We build tests from your own cases, score every model and agent on them, and run them again every time something changes. What we build · How an evaluation runs · Wherever the work is. Hosted Private cloud On-premise Edge Air-gapped ## What we build for evaluation. The tests, the scoring and the record, built around your work and left running after we hand over. - ### Test sets Real cases from your own work, labelled by people who know it, and kept apart from anything a model learns from. - ### Scoring Every answer checked field by field against the right one, with accuracy reported at the automation rate you choose. - ### Model comparisons Candidate models, ours or third-party, run on the same cases, each named with its licence. - ### Sign-off checks Requests that need a person are counted on their own, so you see which ones a model would have let through. - ### Reruns on change The same tests run again whenever a model, prompt or source changes, before the change goes live. - ### Run records Every run kept with its cases, scores and errors, so any result can be traced back to the cases behind it. ## A leaderboard is not your work. Most AI is chosen on numbers measured somewhere else. The only test that counts is your own cases. ### Public benchmarks. Scored on someone else’s data, not on the documents and decisions you deal with. ### Demos. A good demo shows the cases that work, not the ones that don’t. ### Silent changes. A model that passed last month can change underneath you without anyone noticing. ## How an evaluation runs. Agreed criteria first, then your cases, then a result you can act on. The same steps apply to our models and to third-party ones. Scoping 2 to 3 days Choose one or two decision workflows. Acceptance criteria agreed in writing before any results are seen. Evaluation About 2 weeks Candidate models run on 200 to 500 of your labelled cases. We report accuracy, accuracy at your automation rate, calibration, speed and errors by class. Measured on your cases, not a public leaderboard. Every model named with its licence. Private deployment Agreed per client The chosen model is served privately, with your threshold, review routing and audit log. Your cloud or a private endpoint. Your data is not used for training without written consent. Reruns On every change Every model is tested field by field on examples it has never seen, and tested again whenever it changes. If a change scores worse on your cases, you see it before it goes live. ## Measured on your cases, not a leaderboard. What you get from every evaluation, whichever model it tests. **Every model named.** With its licence, so you know exactly what you would be running. **Your threshold.** Accuracy at the automation rate you choose, and the rest sent to a person. **Tested again on change.** The same cases run before any change goes live. ## Evaluations we have published. The same method, run in the open on real work. - ### TypeSafe’s Jev, tested on real work 97.7% right on 177 scored questions, including invoices, expenses and a plant restart with an expired certificate. [Read the evaluation](https://dscvryai.com/insights/jev-eval/) - ### Cloudflare’s Clef, close behind 94.9% on the same 221 questions, and it caught 27 of 30 requests that needed a person to sign off. [Read the evaluation](https://dscvryai.com/insights/clef-eval/) - ### A small model that reads NDAs It rules on 17 standard clauses and quotes its evidence: 84.5% on a sealed test set, running on an ordinary CPU. [Read the research](https://dscvryai.com/insights/nda-clause-model/) - ### When a spreadsheet changes its meaning Twenty real public spreadsheets, three editions each. In 13 of 40 edition pairs, an automated reader should stop. [Read the research](https://dscvryai.com/insights/blind-spreadsheet-benchmark/) ## Where it runs. Tests run where the model will run, from hosted to air-gapped, so your cases never leave your boundary. ### A managed environment. Start with a hosted service when speed of adoption and managed operations are priorities. Agree data handling, access and service requirements as part of the engagement. Discuss hosted · ## Put them to work. Every one of these is tested this way before it is relied on. ### Models Specialist models, evaluated on your data and run inside your boundary. Explore Models → https://dscvryai.com/technology/models/ ### Agents Agents that run on their own and hand over to a person at the decisions that matter. Explore Agents → https://dscvryai.com/technology/agents/ ### Products Ready-made products for recurring jobs, from documents and contracts to sustainability data. Explore Products → https://dscvryai.com/technology/products/ ### APIs Model, agent and data APIs that plug into the systems you already run. Explore APIs → https://dscvryai.com/technology/apis/ ### Data Data collected, extracted and governed, with every value traced to its source. Explore Data → https://dscvryai.com/technology/data/ ## Bring us the work where wrong is expensive. Start with the operational problem. We’ll work out the data, models and agents it needs, and where it should run. [Talk to us →](https://dscvryai.com/contact/) **London**Discovery AI Limited · United Kingdom **Hyderabad**ReDscvry Technology Private Limited · India **Chennai**ReDscvry Technology Private Limited · India