Reading the report
A completed session produces a report. It goes to you, never to the candidate.
What is in it
| Field | What it is |
|---|---|
| Overall score | A single number for the session |
| Assessed level | The model's judgement of where the candidate sits |
| Skill breakdown | The same score per focus area, with correct, partial and incorrect counts |
| Strengths and weaknesses | Short lists, in prose |
| Summary | A narrative account of the session |
| Recommendation | A suggestion about proceeding |
Underneath sits every turn: the question, the answer, the evaluation and the model's confidence in it.
How much to trust it
The report is generated by a model reading a transcript. It is evidence, and it is a summary of better evidence that is one click away.
Three specific cautions:
The skill breakdown is thinner than it looks. Fifteen questions across four areas is three or four questions each. "Testing: 40%" often means one wrong answer out of three, and presented as a percentage it invites a confidence the sample does not support.
Evaluation confidence varies per turn, and it is recorded. An answer marked incorrect with low confidence is worth reading yourself — free-text answers that are right but unusually phrased are the common case.
The assessed level is a comparison to a norm nobody wrote down. It is useful as a rough sort, not as a title. Whether someone is "senior" depends on the job, and the model does not know the job.
The recommendation
This is the field to be most careful with, because it is the one that looks like a decision.
It is a suggestion produced by a model, and it has no authority. Nothing in the platform acts on it: no status changes, no candidate is filtered, and the candidate is never told it exists.
Under GDPR Article 22, a decision produced by automated processing with no meaningful human involvement is exactly what a candidate can object to. Reading a model's recommendation and forwarding it is not meaningful involvement — it is the automated decision, with a person's name on it. If you reject someone, the reasons need to be reasons you hold, from evidence you looked at.
This is also why the candidate never sees the report. If they did, an AI-generated verdict would be reaching them directly, and no amount of internal process would change what happened.
What the candidate sees
During the session: whether each answer was right, and an explanation. Afterwards: nothing.
No score, no level, no breakdown, no strengths, no weaknesses, no recommendation — through any endpoint. That is enforced by an automated contract test over the candidate handlers, so it cannot be reintroduced by an accidental field addition.
If you want to give a candidate feedback, that is yours to write and yours to stand behind — see Reviewing results.
Before you act on a report
For any candidate you are about to reject or advance, open the turns. It takes a couple of minutes and answers the question the report cannot: were these the right questions, and was the answer really wrong?
If a question was bad, flag it — and re-read the score before using it.