Same paper.
Different attention architecture.

Pre-registered-style protocol

This session compares two presentation arms using the same causal question set and the same disappearance rule. The assignment is randomized locally and hidden until the result. One arm receives the progressive-compression reading surface; the other receives a conventional A4 PDF rendered from the same current draft.

This is an instrument scaffold, not a result. A real comparison needs randomized readers, fixed questions, preregistration, version hashes, sample-size planning, and independent analysis. Do not pool these local runs as human evidence.

arm: hidden budget: 180s questions: fixed

Recover the model

Answer from memory. These questions test causal understanding, not paper trivia.

The arm is revealed only now: . This receipt is one local trajectory, not a population estimate.