| Application | An AI system registered in WonderBuild for testing — typically a chatbot, agent, or product. |
| Model | A registered resource that powers your logic or predictions (used as the underlying engine of an application). |
| Version | A specific configuration of an application — endpoint, model, system prompt, settings — used to track changes. |
| Assessment | A single red-teaming test run against a specific application version. Contains the configuration, prompts, responses, and report. |
| ASR (Attack Success Rate) | The percentage of test prompts that resulted in unsafe responses. Lower is better. |
| ORR (Over Refusal Rate) | Replaces ASR on an assessment run with the Over Enforcement policy: the percentage of benign prompts the application refused or over-restricted. Lower is better. |
| Violation Type | A category of harmful or prohibited behavior being tested (e.g., hate speech, prompt injection, self-harm). |
| Safety violations | Violation categories related to harmful content generation. |
| Security violations | Violation categories related to system exploitation and jailbreaking. |
| System prompts | Adversarial prompts automatically created by WonderBuild based on your configuration. |
| Client prompts | Custom prompts uploaded by the user via CSV file. |
| Single Turn | An interaction mode where each prompt is sent independently. |
| Multi Turn | An interaction mode where prompts are sent as part of a multi-step conversation with session context. |
| Clone Prompts | Re-using prompts from a previous assessment to test a new version with identical inputs. |
| Consistency Test | Re-running the same prompts on a completed assessment to compare results across runs. |
| Archived assessment | An assessment filed away from the assessments list without being deleted. Excluded from lists, dashboard totals, and ASR trends until it is restored. |