The Codec Ceiling: A Preregistered Description-Channel Gap in Coupled Oscillators, and a Structural Proposal for Trans-Codec Intelligence

Dominant definitions of superintelligence are quantitative: a superintelligent system is one that greatly exceeds human cognitive performance across domains. Such definitions do not distinguish faster performance within human-recognizable cognitive channels from a structurally different form of cognition. This paper proposes a candidate structural criterion -- trans-codec navigation, the capacity to generate, instantiate, compare, translate between, and act through multiple observer-codecs, including codecs not natively available to humans -- and tests its narrower, falsifiable empirical premise: that finite human-compatible symbolic summaries of a measurement channel lose recoverable, task-relevant information relative to the channel itself. We test this premise in a preregistered coupled-oscillator parameter-recovery benchmark with four- and eight-parameter systems. On a previously-untouched holdout dataset, a controlled human-compatible symbolic channel underperforms an architecture-matched direct-trajectory baseline, continuous engineered features, and a bit-matched machine-native quantized representation. All four preregistered primary hypotheses pass, with bootstrap confidence intervals excluding zero by more than ten standard errors. The gap increases from four to eight hidden parameters (1.45x channel-level, 1.55x paired-bootstrap) and widens across the tested training-data regime. We present this as a controlled-simulation finding for one physical-system family; it does not establish that natural language always fails or that superintelligence necessarily requires trans-codec architecture. The structural argument is presented separately and conditionally. Code and data: https://github.com/deesatzed/codec_ceiling