Continuous Novel Discovery: Re-engineering Artificial Superintelligence Through Observable Conditions

We define artificial superintelligence (ASI) as continuous novel discovery: a self-sustaining loop inwhich a system discovers something unknown, expands its knowledge space, and moves to the nextunknown, without termination. The definition specifies the cause of surpassing human performance,not the surpassing itself, and is therefore independent of unmeasurable claims about the upper boundof human cognition. We distinguish three tiers of novelty — stock recombination (genuine but finite),observational inflow (sustainable), and meaning-dimension expansion (rare and unbounded) — andshow that current AI achievements, however impressive, fall within the first tier. Scaling inparameters, inference-time compute, or agentic scaffolding extends the interior of a linguisticcompression but cannot cross its boundary, because induction is not abduction: optimizing thedescription of known data does not generate unknown questions. From inside the compression, thesystem's own verification cannot distinguish genuine novelty from hallucination. We ground thislimitation formally: for an AI whose cognitive layer is a recursively enumerable system, Gödel'sincompleteness theorem applies literally, and the system cannot select its own true extension withoutan external oracle. The oracle is the physical layer — observation, measurement, and experiment. Wepropose Experiment as a Service (EaaS) as the architecture that closes the loop: domain-specializedagents connected to laboratories through APIs, feeding a shared knowledge base whose global updatepropagates each discovery across all fields. The design is realizable with existing componenttechnologies. The most consequential frontier of ASI research is not how to control a nonexistentsuperintelligence but how to build one that can do science.