Self-improving · AI system

Self-improving AI system

A self-improving AI system is a system whose output is measured on fresh tasks each loop and whose context, not its weights, is improved between loops. CodeNinja's Praxis is run as one: each RSI loop (recursive self improvement loop) measures Praxis against frontier AI on its own on the Physical AI Design Index (PADI), and the 23 designs published here are its output. Latest measured RSI loop, RSI Loop 21: Praxis platform 54.4% against frontier AI 37.7% of checks passed.

The self-improving AI system, RSI loop by RSI loop

Each row is one RSI loop: fresh tasks, both arms scored the same way, the share of checks passed. The blind row is the evaluation on tasks the platform had never seen. How PADI scores is on the PADI page, with its limitations; this page shows results only. RSI Loop 23 is running now.

What changes between RSI loops in a self-improving AI system

The weights do not. What improves is the context Praxis reasons with: the published designs, their object models, model registers and cost lines, and the gaps each RSI loop finds, which is why self-improving AI systems built this way can be audited loop by loop: every input to the next loop is a published document. Design on Praxis; the measure is PADI.

What the self-improving AI system produced

The 23 published designs, newest first. Each is a complete reference architecture with a DOI under CC BY 4.0.

Results from the PADI export only; designs from the published dataset at Hugging Face. The designs are also collected as sovereign AI reference architectures, as an industrial AI reference architecture by sector and as physical AI system design; their object models at industrial ontology examples. Updated 2026-10-26.

Want one of these designs for your own operation? Talk to us.