# Fairness and limitations ## Fairness controls in the public product contract Axiom Cortex is designed around job relevance, source traceability, visible uncertainty, and human review. The public controls include: - Use the same approved role criteria for people evaluated for the same role and stage. - Keep findings connected to attributable interview evidence. - Distinguish missing evidence from explicit contradiction. - Check whether the interview created a valid opportunity to answer. - Separate personal ownership from team outcomes. - Exclude face, gaze, appearance, accent, voice quality, emotion, truthfulness, protected traits, and clinical interpretation. - Keep a human reviewer responsible for corrections and final decisions. ## Known limitations ### Input quality limits output quality An incomplete job definition, weak question, inaccurate transcript, or poor interviewer follow up can reduce the quality of the evidence. Structured processing cannot recover facts that were never captured. ### Interviews are incomplete samples An interview shows a limited set of responses under particular conditions. It does not reveal every capability, every work setting, or a complete person. ### Role criteria can contain human bias Consistency does not make a criterion fair. Companies must review requirements and ideal answer criteria for job relevance, accessibility, and lawful use before evaluating candidates. ### Language and communication context matters Concise answers, second language responses, disability related communication, accent, slower delivery, pauses, or different speaking styles must not be treated as lower ability. This matters for Latin American candidates and anyone speaking English as a second or additional language, where thinking time, translation, and turn-taking can change the surface pattern. Review the job relevant information, not presentation conformity. ### Baseline questions and speech-pattern review If a customer wants an interview-integrity review, the interview should begin with an open work and career-experience question that lets the candidate explain their path into engineering and why the role fits. The answer can provide a comparison baseline before harder technical questions, but it is not a psychological, personality, calm, or capability measure. TeamStation AI research treats this opening as a possible primer that may help a candidate settle into the interview. It is a company-reported design observation, not independently validated evidence about the mind. Without a usable, source-preserved baseline, later speech-pattern variation should not be flagged through this comparison. Even with a baseline, pauses, slower speech, accent, disfluency, code-switching, translation, disability, or L2 or ESL communication are contextual factors, not proof of AI assistance. ### Transcripts can be wrong Automated transcription can misidentify speakers, technical terms, names, numbers, or negation. Material findings should be checked against the available source recording when accuracy is uncertain and lawful access exists. ### Evidence does not equal prediction The public documentation does not establish predictive validity for future job performance. A report describes support and gaps in the supplied evidence under configured role criteria. ### Fairness requires ongoing review No evaluation process is automatically fair because it is structured. Companies need governance, access controls, accommodation processes, monitoring, appeal or correction paths, and legal review appropriate to their location and use. ## Prohibited uses Axiom Cortex should not be used to diagnose mental health, personality, intelligence, emotion, truthfulness, disability, or protected traits. It should not be used for facial analysis, accent scoring, voice quality scoring, covert surveillance, autonomous hiring decisions, or claims about a candidate without attributable job related evidence. ## Appropriate uncertainty language Use phrases such as "supported by the supplied interview," "partially supported," "not observed in the supplied material," "requires more evidence," and "requires human review." Avoid statements that turn a limited interview into a universal claim about the person. ## Related resources - [Human review](https://axiomcx.dev/knowledge/human-review.md) - [Public claim boundary](https://axiomcx.dev/knowledge/public-claim-boundary.md) - [Safe viewer prompts](https://axiomcx.dev/ai/viewer-prompts.md)