Giving the model eyes: ghost dimensions as an introspection prosthetic.

Track 3 asks whether models have privileged access to their own internal states. We looked at what they process that they structurally can’t report: high-variance directions in the residual stream with almost no overlap with the verbalizable workspace, confirmed generic to the depth regime with a matched-variance null rather than an artifact of one component.

Across 167 snapshots, ghost content and workspace content barely overlap — 95.5% of ghost vocabulary never shows up in the workspace, and it’s dominated by tokens about memory itself. We built a mechanism that hands this content back to the agent; whether the agent can actually use it is the next experiment, not this one.