Intelligence as Organized Difficulty Compression

We propose a system-level theory of intelligence centered on a single claim: **intelligence is organized difficulty compression*. The motivating problem is that the term intelligence* is routinely asked to cover too much at once, ranging from consciousness and general reasoning to benchmark success and adaptive control. This paper isolates the system-design layer. Let an adaptive system interact with an environment through a history \[ h_t = (o_0, a_0, o_1, a_1, \ldots, o_t), \] with internal state \(z_t\), objective functional \(J\), and resource budget \(B\). We retain a layered, objective-relative description of system intelligence: \[ \mathbf I(\pi; \mathcal E, J, B) = (P, M, D, L, T), \] where \(P\) is perception of objective-relevant state, \(M\) is modeling adequacy, \(D\) is decision quality, \(L\) is learning gain, and \(T\) is transfer to adjacent task families.

The paper's main advance is to argue that a major source of intelligence lies upstream of local action quality. Intelligent systems do not only search harder. They often solve harder problems by discovering better representations \(\rho\) that lower the task's effective requirement profile \[ \mathbf K_{\mathrm{eff}}(\tau; \rho, B). \] This yields a difficulty-frontier view of intelligence magnitude and leads to a stronger thesis: frontier lift can arise from representational improvement even when raw solver strength is unchanged.

We also retain an emergence view of intelligence. A merely reactive system is not yet intelligent in the strong system sense used here. Intelligence appears when state gain, model gain, decision gain, learning gain, and transfer gain jointly cross a viability threshold. From this framework we derive three central consequences: reactive policies are insufficient for strong intelligence emergence; the emergence threshold is bottleneck-governed rather than fully compensatory; and for many task families, especially theorem search, representation gain can dominate equal-cost increases in local search throughput.

The theorem domain is treated as a privileged case study. There, intelligence is often expressed not mainly as brute proof search but as intermediate object discovery: the invention of lemmas, invariants, decompositions, normal forms, and bridge objects that reduce effective difficulty for the remaining proof obligations. Collective intelligence is then reinterpreted as externalized cognitive closure: a shared organization that preserves state, verification, negative knowledge, and reusable representations over time.

The paper is conceptual rather than empirical, but it is not merely verbal. Its goal is to give a sharper object language for distinguishing raw search from representational intelligence, and local competence from systems that genuinely compress difficulty under bounded cost.