01 / Research
Research
We address questions at the frontier of computational science — problems whose resolution would substantially alter the trajectory of the field.
Approach
SHINGETSU operates on the conviction that computational progress has become too narrowly defined by scale. The assumption that larger models, more data, and more compute necessarily produce more capable systems has dominated the field for over a decade — and while this has produced remarkable results, it has also crowded out alternative lines of inquiry.
We study the underlying mechanisms of intelligence and computation: how capability is structured within a model, how knowledge can be preserved across technological generations, and what forms of efficiency are achievable when scale is deliberately constrained. These are not niche questions. They are foundational.
Our methodology emphasises reproducibility, formal measurement, and scepticism toward consensus. We publish all protocols and baselines. We report negative results. We do not benchmark against metrics we consider inadequate.
Active Programs
Efficient Intelligence Program
"How much general-purpose intelligence can be concentrated into the smallest practical computational budget?"
This program challenges the prevailing assumption that capability scales primarily with parameter count. We investigate the structural and methodological conditions under which smaller models can achieve disproportionate capability — studying training curricula, architectural choices, and evaluation frameworks that have been underexplored by the field's focus on scale.
Computational Continuity Program
"How can computational knowledge and digital artifacts remain understandable, verifiable, and usable across technological generations?"
Digital preservation is conventionally treated as a storage problem. We treat it as a computation problem. The degradation of software artifacts over time — the loss of executability, interpretability, and verifiability — is a systemic failure that becomes more severe as systems grow more complex. This program develops methodologies for long-term computational continuity.
Registered Projects
Full Registry →An Experimental Study in Maximum Intelligence Density