| Ground truth | The correct answer an AI is measured against |
| Certified answer | An answer independently solved by professionals, reviewed, settled where needed, and tied to the law |
| Certified record | A certified answer with its full history: who, when, how checked, how confident |
| Test data (evaluation data) | The exam: checks an AI, never teaches it |
| Training data | The textbook: an AI learns from it |
| Reference pack | A rulebook an AI looks up while working |
| Reasoning trace | Step-by-step expert reasoning to an answer |
| Preference data | An expert's choice of the better answer, with the reason |
| Confidently wrong | A wrong answer given without any sign of doubt |
| Ambiguous case | A question where the law is genuinely unsettled; never scored as simply wrong |
| Adjudication | A senior professional settling a disagreement between solvers |
| Calibration | A test with known answers that professionals pass before real work |
| Confidence score | A 0–100 measure of how strongly an answer is supported |
| Provenance | The history of how a record was made |
| Data Mission | A customer order turned into a precise production plan |
| Fingerprinting | Giving each customer a slightly different copy so leaks can be traced |
| Sandbox | A test copy of a system where mistakes harm nothing |
| Regression | Something that used to work and now doesn't |
Zynoviq