Before Normative and Moral Alignment: Causal Contract Faithfulness as a Precondition for Trustworthy AI

We argue that evaluating an AI system's normative and moral alignment first requires causal contract faithfulness: when a foundation model offers a directed causal graph as a rationale, later behavior should respect the commitments it expresses. The causal graph need not be read primarily as the world's true causal structure, nor as a transparent diagram of hidden computation. It can instead be treated as a public causal contract, relative to task variables and decision time. Because the graph is causal and directed, it does more than name correlated inputs: it declares permitted influences, claimed irrelevancies, and inadmissible late evidence. A model keeps this contract only if later scores or decisions, made without the earlier causal graph, respect those commitments up to ordinary query variation. This structural trust is behavioral, not moral. A biased, incomplete, or false causal graph may still be behaviorally faithful, and therefore open to criticism, correction, or rejection. By contrast, a morally attractive causal graph that is not followed is not an explanation for accountability; it is an ornament. In audits of LLMs on admissions and loan-default tasks, reported causal graphs often make clear commitments that later predictions violate. Accuracy, stable reports, model scale, and reminders of the causal graph do not establish faithfulness. Causal contract faithfulness is therefore not a certificate of fairness, legality, causal truth, safe deployment, or moral alignment. It is a prior epistemic condition for trustworthy AI: normative and moral alignment cannot be evaluated through public reasons unless those reasons first bind behavior.