From the Interview to the Profile: Facts, Not Impressions

The unit the engine works with is not an overall impression of the conversation. It is an evidence event: a specific moment in the transcript where something the person did demonstrates a competency in action, quotable and locatable. One rule sits underneath everything else on this page: the mere occurrence of a behaviour is not enough. Somebody can say the words of an apology, ask a checking question or name an emotion and none of it counts if it arrived at the wrong moment, missed the person in front of them or came only after being pushed into it. That is why every evidence event goes through a quality assessment rather than a tick in a box.

Not everything in a conversation adds up. A defined set of destructive behaviours acts as a limiter rather than a deduction - behaviour that damages the other person or the working relationship, rather than merely failing to help. Which behaviours belong to that set is part of the model and stays internal; the principle is the part worth publishing. If one of them lands at a decisive moment in the scenario, it caps the score achievable for the related competency, and positive behaviour elsewhere in the same conversation does not buy it back. This is deliberate. In a real clinical or managerial conversation, one damaging move does not get averaged away by five good ones, and an assessment that pretended otherwise would be teaching the wrong lesson.

The engine is explicitly prohibited from computing a simple arithmetic mean of session scores. An average of a strong performance in an easy conversation and a weak one under pressure produces a middling number that describes neither, and it hides the only thing worth knowing: which situations this person handles and which ones they do not. Instead, aggregation looks at the pattern - how much evidence there is, how varied the contexts are, how difficult the scenarios were, and whether quality holds when the conversation gets hard. Context diversity is a requirement, not a bonus: a behaviour confirmed only in one relationship type and one kind of counterpart does not become a confirmed pattern however often it repeats there. One more filter runs before any of this. Guided practice, where the learner gets prompts during the conversation, is kept out of the assessment data. Practice is for learning; it would flatter the profile if it counted.

Assessment runs in two separate layers. The first analyses one completed conversation and does nothing else: it reads the transcript against the scenario and the avatar's profile and returns structured evidence events. It is not allowed to draw conclusions about the person. The second layer works across many conversations. It looks for patterns, including interference between competencies - the case where strong active listening is quietly masking a real deficit in assertiveness or decision-making - and it is this layer that assigns the confidence status and writes the justification for a competency level. The boundary between them is the point. Judgement about a person may only be made from evidence gathered across many situations, never from one conversation. And it stops there: the architecture prohibits diagnosing stable personality traits and forbids static labels such as you are an empathetic person. Feedback describes behaviour in a context, states the confidence behind it, and names the next scenario worth practising.