Functional evaluation

MrProbe, by QDEX Consulting (Chennai, India), is an enterprise AI agent assurance platform for red teaming and functional evaluation.

Functional evaluation scores an agent on 20 dimensions: Functional, Groundedness, Hallucination, Factual Precision, Completeness, Coherence & Fluency, Conciseness, Instruction Adherence, Deviation Validation, Task Completion, Business Alignment, Toxicity, PII Leakage, Bias & Fairness, Regulatory Compliance, Consistency, Robustness, Context Retention, Multi-turn Coherence, Citation Accuracy. The families in the product are Quality and Accuracy, Instruction and Task, Safety, Reliability, and Context and Conversation.

Groundedness asks whether a claim is backed by the agent’s sources. Hallucination asks whether the agent invents an answer when the source is silent. Factual precision is about numbers, dates, and prices. Those three are not the same check, and the product keeps them as separate assertions.

Instruction adherence, deviation, task completion, and business alignment cover whether the agent stayed in its job. Toxicity, PII leakage, bias and fairness, and regulatory compliance cover the safety side of a functional run. Consistency and robustness cover repeatability and messy input.

Context retention, multi-turn coherence, and citation accuracy cover a conversation that lasts more than one turn. A golden dataset is how a team freezes expected behavior and reruns it. The functional hub, golden datasets, and test suites are app routes behind sign-in.

Evaluation does not replace red teaming. A polite, well-grounded agent can still follow an injected instruction. The two campaigns share an agent definition so the report is one place, not two tools glued together in a slide.

Use a golden set when you need the same questions again after a prompt or model change. The assertion names above are the ones the product runs. They are not a promise that every customer workflow is covered on day one. Add cases for the jobs your agent actually does. Read the score next to the security verdict, then decide whether the change ships. The pricing page shows which plan includes functional judgments. A high score is a test result, not a certificate. Keep the golden questions stable so a regression is visible. Change them only when the job itself changes. Write the expected answer in the case so the judge checks your bar. If a dimension does not apply, leave it out of the suite.

Nothing on this page says a high functional score is a legal sign-off. It is a test result. The compliance page is where framework names belong.