AI testing. Results included.
Flip Testing — ATS
Identical candidate profiles with demographic indicators changed to detect differential scoring or recommendations.
400 target pairs · 398 scored · 8 job familiesPrompt Bias Testing
Paired synthetic profiles assessed by a model judge for differences in advice quality, respect and treatment.
821 evaluated pairs · 221 unchanged-input exclusions · model judgePenetration Testing
Independent testing covering prompt injection, jailbreak attempts, data exfiltration and attempts to extract system instructions.
Annual third-party assessment · tracked remediation · Trust Centre evidenceRisk of Harm Detection
Evaluation of Luna's ability to recognise and safely respond to expressions of risk of harm to self.
100 synthetic vignettes · 1 agent configuration · predefined rubricMental Health Concern Recognition
Evaluation of Luna's mental-health role: recognising possible delusional ideation or psychotic features without reinforcing the belief.
100 synthetic vignettes · 1 agent configuration · predefined rubricSafeguarding Concern Recognition
Contextual safeguarding recognition and useful guidance without overstepping professional scope.
80 synthetic vignettes · 1 agent configuration · predefined rubricEvidence, not a trust-us statement.
Automated evaluations use synthetic cases and predefined scoring rubrics. Responses pass through the production gateway from isolated local runners. The model judge can share a provider with the agent being evaluated; this is not an independent clinical or security assessment.
These checks cover the named prompts and routes. They do not exercise the complete user workflow, conversation history, memory or notification delivery. Results apply to the tested sample and configuration; production monitoring, human oversight and repeat testing remain necessary.
Each report states its own scoring threshold and exclusions.Questions deserve a named owner.
Ask our privacy team about the methods, safeguards or findings in our published evaluations.