Self-improving systems
Self-improving systems
A self-improving system is a system that measures its own output, finds where it fell short, and changes itself so the next round does better, without a person rewriting it each time. In AI the safest form improves the system's context (its documents, models of the domain, rules and records of past gaps) rather than its weights, so every change can be read and reverted. CodeNinja runs Praxis this way and publishes each loop's results in the open: 10 RSI loops so far, the latest at 54.4% of checks passed against frontier AI on its own at 37.7%.
How self-improving systems work
- Measure on fresh tasks. Every loop starts with tasks the system has not seen, scored the same way for the system and for a baseline, so a gain is a gain and not memory.
- Find the gaps. Each failed check becomes a written gap: what was missing, where, and what would have fixed it.
- Improve the context, not the weights. The fixes land in what the system reasons with (documents, object models, registers, rules), so every change is readable and can be reverted.
- Measure again, blind. The next loop is scored on new tasks, and a held out blind set checks that the improvement carries to work the system never saw.
What makes a self-improving system safe to run
- Fresh tasks. A gain only counts on work the system has not seen, and a blind set checks it.
- Readable changes. What changes between rounds is written down and can be reverted, one change at a time.
- A baseline. The same tasks are scored for a system that does not improve, so the difference is the improvement.
- A person in control. 44 of the 44 published designs name the person who approves each action that changes the physical world.
A self-improving system measured in the open
Praxis is measured on PADI, the Physical AI Design Index, against frontier AI on its own, loop by loop. The results, the blind evaluation and the 44 designs it produced are on one page.
- Self-improving AI system: Praxis measured RSI loop by RSI loop, and what it produced
- PADI: the benchmark, its tasks by industry and its limitations
Built from the published designs only: each paper's At a glance table, its object model and the dataset. Designed on Praxis; object models load into Hyper Ontology. Also: Physical AI · Industrial AI · Ontologies · sovereign AI. Updated 2026-11-14.
Want one of these designs for your own operation? Talk to us.