Privacy, Bias, and the Human Factor

This page describes a design framework: the principles and constraints the Digital Twin is being built to. It is not a description of the production system and should not be read as one. If you need to know what is running today - where data sits, who can reach it, what the contracts say - that is the security page, and it is the only page we treat as the source of truth on the current state. Keeping the two apart is the point: a design principle and a shipped control are different things, and blurring them is how vendors end up promising infrastructure they do not have.

Two things get asked before anything else on this page: whether client data is used to train models, and where the data actually sits. Both are answered on the security page, and we deliberately do not repeat the answers here. A commitment written down in two places, in four languages, is a commitment that will eventually say two different things - so there is one page that carries them and one page to update when they change. The framework described here is about how the assessment is designed to work. It is not a substitute for the technical and contractual documentation, and where the two appear to disagree, the security page is right.

Every one of those biases is countered the same way: the assessment is tied to what was said, not to how it sounded or to who appeared to be saying it. The engine is instructed to disregard names, pronouns and demographic markers if they happen to be spoken, and to look past grammatical errors, hesitation and colloquialism to the intent underneath. That is the design answer to eloquence and stereotype bias. Against hallucination the rule is stricter. An assessment must quote the exact words from the transcript that justify it. If the quote is not there, the evaluation is rejected rather than reported - a claim about somebody's behaviour with no quotable moment behind it is not a finding, it is a guess with good grammar. None of this eliminates bias. It removes the easiest routes to it and makes the remainder visible, which is why the validation study runs against human assessors rather than against our own expectations.

An assessment model left alone drifts. Criteria loosen, the same behaviour scores differently in March than it did in January, and nobody notices because there is nothing to notice against. The framework answers this with human-in-the-loop monitoring: trained behavioural assessors review anonymised samples of transcripts blind and compare their evaluations with what the system produced. Where they diverge, the criteria are recalibrated. It is the same comparison that sits at the centre of the validation study, run as a continuing check rather than a one-off. A learner can also ask for human review of feedback they disagree with. Automated feedback that cannot be questioned is not feedback - it is a verdict, and this system is not designed to hand down verdicts.