What the system does, what it does not, and what is stored.

Human review is mandatory

It is not an option that can be switched off. The automatic score is shown read-only, and human observations are written in separate fields stored apart from the computed score. A result with no human review is not a closed result.

How a score is built

A score does not come from asking a model "what would you give this?". It comes from a rubric and an arithmetic operation.

• The team defines the rubric before anyone is invited: competencies, observable criteria and weights.
• The models contribute one judgment per criterion, and every judgment arrives with the fragment of the conversation or the deliverable behind it.
• The weights are applied by code, deterministically. The final score does not depend on a model’s free-form text.
• A score with no evidence attached is a defect in the system, and is treated as one.

What is stored about a candidate

Name and email are the identifying fields. The conversation, deliverable and evaluation records are also retained so the result can be reconstructed. Demographic data is neither requested nor used as an evaluation criterion.

• The full conversation, the deliverable, the rubric and the configuration version are kept for at least six months.
• They are kept for one concrete reason: without them an evaluation cannot be reconstructed or explained later.
• Evidence is stored with controlled access, and is never written into the application’s technical logs.

What is never used as a criterion

No sensitive variable enters the evaluation: not gender, not age, not nationality, not any other personal characteristic. The criteria are observable and defined in the rubric the team wrote.

• No personality assessment and no psychometric testing.
• No biometrics and no advanced identity verification.
• No invasive monitoring while the candidate works: no camera, no screen capture, no control of their machine.

What the candidate is told

Before starting, on the same screen where they begin, they are told they will be talking to an artificial intelligence agent and that the conversation is part of their evaluation. They also see how much time they have and how they are expected to submit.

How far results are comparable

Two scores are comparable when they come from the same challenge and the same configuration version. Beyond that they are not, and the platform warns about it rather than letting the comparison pass unnoticed. It is a limitation of the method, not a feature.

The scope of this page

Adaliza does not claim compliance with any artificial intelligence regulation, and displays no certification or seal: no conformity assessment of the product exists. What does exist is a design built on transparency, human oversight and traceability — which is what this page describes and what can be verified in a demo.

What we do not do

Adaliza does no proctoring, no identity verification, no impersonation detection and no screen monitoring. We do not police the candidate: we observe how they work. That is a settled product decision, not a pending feature.

• We do not access the candidate's camera, microphone or screen.
• We do not detect plagiarism or try to identify the use of AI tools.
• We use no sensitive variables — gender, age, nationality — anywhere in the evaluation.