Superintelligence as an Emerging Systems Discipline: The Long-Horizon Problem as a Foundational Challenge

This paper proposes that superintelligence should be studied as an emerging systems discipline, rather than only as an extension of artificial intelligence. The present study focuses on the long-horizon problem as one operational challenge within this research program. Long-horizon state integrity concerns maintaining the proper authority relationships among current evidence, retrieved memory, task stage, goals, constraints, and action selection. To investigate one mechanism involved in maintaining this integrity, the study introduces a diagnostic microbenchmark, the Long-Horizon State-Integrity Benchmark, that examines how episodic state is permitted to re-enter into ongoing control. A brain-inspired cognitive architecture designed to investigate advanced artificial agency was evaluated in a compact embodied task under ordinary partial observability and a stochastic conflicted state-repair challenge. The experiment compared three episodic state-repair mechanisms: Guarded Merge retrieval, a null repair mechanism in which target episodic readback was disabled, and replacement-style readback. Under non-challenging baseline conditions, success in completing the ordered long-horizon task was 100%, 100%, and 99%, respectively. Under the challenge, success was 97%, 74%, and 50%, respectively. Guarded Merge exceeded disabled readback by 23 percentage points (95% CI: 15–31, p < 0.001) and replacement-style readback by 47 percentage points (95% CI: 37–57, p < 0.001). These findings suggest that memory availability and the authority granted to retrieved memory can each affect an agent’s ability to maintain coherent control over a long trajectory.