A Scaling Prediction for Global-Workspace Emergence in Language Models: A Preregistered Discriminating Test Derived from the Prima Conscientia Framework

A preregistered prediction distinguishes two accounts of why language models develop internal workspace-like structure. Under the Prima Conscientia framework, which views complex systems as expressions of a fundamental consciousness field, workspace structure should appear even in very small models and strengthen smoothly with scale. Under the rival complexity-emergence view, workspace properties should appear only in discontinuous jumps at capability thresholds. The prediction is decomposed into structural presence (the framework's central commitment) and downstream function (allowed to lag). A test protocol using the Jacobian-lens methodology and Pythia model suite is specified, with falsification conditions and integrity conditions applied symmetrically to confirming and disconfirming outcomes.

Study at a glance

Abstract

Recent interpretability research from Anthropic (July 2026) reports that Claude language models contain a small, privileged set of internal representations — the "J-space" — functioning analogously to the global workspace posited by global workspace theory as the mechanism of conscious access in humans. The structure emerged during training rather than being engineered. Because this finding is consistent with multiple competing accounts of how workspace organization arises, it confirms none of them by itself. This paper preregisters a discriminating prediction. Under the Prima Conscientia framework, which treats complex information-processing systems as apertures through which a fundamental consciousness field is expressed rather than engines by which consciousness is generated, proto-workspace structure should be present at least minimally even in very small models, with its structural signature strengthening smoothly and continuously with scale. Under the rival complexity-emergence view, workspace properties should instead appear predominantly in discontinuous phase transitions at capability thresholds. The prediction is decomposed across the functional properties of a workspace, distinguishing structural presence (the framework's central commitment) from downstream function (permitted to lag). A complete test protocol is specified using the open-source Jacobian-lens methodology and the EleutherAI Pythia model suite, including mandatory resolution controls that render null results at small scale interpretable, explicit falsification conditions under which the framework is to be revised or withdrawn, and integrity conditions applied symmetrically to confirming and disconfirming outcomes. The purpose of this record is to place the prediction on the public record before the relevant measurements exist.