Self-improving AI
When an AI setup changes, its old track record may no longer fit.
HANRIA is a design in development for keeping separate, checked track records for locally run, open-weight model setups and for different kinds of work. Hosted AI services are outside the current scoring design.
The problem
A new version is not the old version.
A company changes the model, tools, or instructions in an AI setup. The old track record may no longer describe the changed setup. If AI systems come to help improve themselves (recursive self-improvement, or RSI), such changes could come faster. Either way, a business still needs to know which setup to trust with which job.
A system saying it has improved is not evidence. A single score for everything also hides that a setup can be strong at one kind of job and weak at another.
Four design choices
How HANRIA is designed to help.
A record for each kind of work
Standing means a track record earned from checked work. The design is intended to keep it separately for each kind of job, not as one number.
Checks that share less of the failure
Reviewers from the same source may share blind spots. The design is intended to exclude them from a vote and to use an outside check when one is available. Excluding same-source reviewers has only early support.
Records that can be replayed
The design is intended to keep enough information to repeat how a score was calculated, so that a claim could be checked later rather than taken on trust.
A changed setup starts again
The current design treats the setup around a model, including its runtime, instructions, tools, and operating rules, as a separate profile. A changed profile is designed to receive no inherited track record. That would avoid automatic carryover but would leave the changed setup without history. How records should carry across updates remains unresolved.
Plain answers
Questions, checked.
Does HANRIA stop AI from improving itself?
No. It is designed to help people compare checked track records for particular kinds of work, not to limit what a system does.
Does a good track record make a system safe?
No. The design is intended to record how checked work turned out and add consequences, not guarantees. It would not guarantee quality or stop a determined bad actor.
Has this been shown to work against self-improving systems?
Not yet. Parts of the design have been explored in simulations and small studies, with mixed results. It has not been shown to work against self-improving systems.
Is it available?
No. HANRIA is a design in development. Nothing on this page is offered for use.