When Consciousness Cannot Be Identified: A Descent-Theoretic Framework for Scientific Admissibility

Scientific attributions of consciousness require an identification claim: the target property Θ must be recoverable from a declared observation site Y. This paper studies hidden-variable systems in which that recovery fails in a way that is stable under admissible revisions of the measurement language. The class is specified by four properties: local contextual presentability, nontrivial transition structure, nontrivial holonomy, and equivariant admissible operations. For systems in this class, the target property is non-identified relative to the declared observation site and revision envelope. The mathematical claim is bounded. The general reduction from identification procedures to effective descent is proved here for finite deterministic systems and exact finite stochastic systems, and isolated as the remaining proof burden for broader substrate classes. The kernel-pair obstruction and Conservative Invariance proposition are formally verified in Lean for the stated finite and categorical cases. The consciousness application follows from that formal result plus an explicit normative premise: a low-concern consciousness verdict requires identification evidence. Where the target is non-identified, or where the required structure is not disclosed, a low-concern verdict is not admissible. The operational output is a six-element disclosure requirement for any organization claiming such a verdict. A worked broadcast-report architecture and a public-system dossier for Llama 3.1 8B-Instruct show how the criterion is applied. The Llama result is not presented as a novel claim about linear probes. Its role is to show target separation: access/report claims can be evaluated under the declared site, while phenomenal-consciousness verdicts remain non-applicable without a target fiber, classifier, challenge record, and provenance certificate.