Causal Provenance and the Epistemology of Artificial Minds: Construct Validity, Inferential Non-Transferability, and Measurement Dependence in Artificial-Mind Attribution

Artificial systems are increasingly evaluated with behavioural, computational, and mechanistic indicators of cognitive constructs such as metacognition, self-modelling, global availability, and potentially consciousness. A problem arises when the evaluated system has itself been exposed, in training or deployment, to descriptions of those constructs, to their proposed indicators, and to the evaluative practices used to detect them. In that case an observed indicator does not identify its own causal source. This paper offers a framework for artificial-mind attribution built on three ideas: a distinction between causal and epistemic provenance, a construct-validity standard for moving from mechanism to construct, and a principle of inferential non-transferability that forbids carrying evidence from one computational unit to another without an explicit invariance argument. The framework is organised as a three-tier architecture and operationalised, at its final tier, by a five-stage provenance protocol. A worked example shows the protocol applied to a hypothetical metacognition claim, and a partial reading of published work on language-model introspection shows where such evidence currently stops. The aim is neither to decide consciousness by counting indicators nor to disqualify systems for having met human concepts, but to state the conditions under which evidence for an artificial mental construct can be attributed to the construct itself.