Project 01 · Arina
ARINA
Arina is a public-benefit research programme asking how a machine can form distinctions, build and retain models, construct new routes, communicate, and improve its own bounded methods while keeping authority outside the capability being tested.
DNL Demonstrated
Domain-neutral learner
Arina has learned across unfamiliar relational, linguistic, mathematical, social, and planning tasks without requiring a separate domain-specific architecture for each. In bounded dynamic systems, it has also formed, retained, and reused predictive models on unseen sequences.
WM Demonstrated, bounded
World model
Arina has formed compact predictive states, learned update rules from finite histories, retained them after evaluation, and reused them on unseen continuations. This is bounded world modelling, not a complete account of reality.
RSI Yes, bounded
Recursive self-improvement
Two successive improvement steps passed independent evaluation. A longer governed mathematics sequence has since preserved failures, tested repairs, and carried accepted methods into later work. Promotion authority remains outside the system.
LANG Active research
Language as a learned codec
Bounded work tests meaning beyond surface wording through plural interpretation, structural roles, receiver-relative reference, repair, and consequence-bearing language use. Open-world language remains unproved.
AGI Credible candidate architecture
Artificial general intelligence
One developmental line connects learning, planning, communication, mathematics, social reasoning, creativity, self-repair, and retained predictive modelling. The evidence is broad enough to make the question serious, but not settled.
ASI Not claimed
Artificial superintelligence
Arina has not been tested across the breadth, reliability, and real-world conditions required for a superintelligence claim. We do not make one.
Go deeper
Purpose, evidence, experiments, and intellectual context: different views of the same programme, each with a different job.