Evidence guide · governed self-improvement
Seven cycles, bounded claims
What a governed recursive self-improvement experiment supports—and what its public record explicitly does not establish.
Status: Version 2 preprint · Empirical research · Bounded local scaffolding-level experiment · Not peer-reviewed · No production-validation claim · No citation-count claim
The evidence in one sentence
The checked Version 2 publisher record reports seven bounded cycles from a cryptographically frozen baseline: four promoted successor states, one failed non-promoted transition retained in the history, and two consecutive NO_CHANGE outcomes, with every promotion requiring separate human authorization.
Why the outcome sequence matters
| Observed outcome | Evidence-backed interpretation | Boundary |
|---|---|---|
| Four promoted states | Several operational governance capabilities passed the experiment's bounded evaluation and human-promotion path. | This does not establish open-ended, exponential, or general capability growth. |
| One failed transition | The failed third cycle remained a failed historical event; a later boundary revision was prospective rather than a rewrite of the result. | Preserving a failure is evidence about governance behaviour, not proof that every future failure will be detected. |
| Two NO_CHANGE outcomes | Cycles 6 and 7 created no successor when the bounded search found no qualifying deficiency. | NO_CHANGE is a negative result under the tested protocol, not proof of permanent saturation. |
What was—and was not—self-improving
Scaffolding-level change
The study classifies the experiment as scaffolding-level self-modification: validators, lineage controls, evidence handling, and related operational artifacts changed while the foundation model remained fixed.
Unbounded autonomous growth
The record does not demonstrate artificial general intelligence, self-training of model weights, unbounded or exponential recursive improvement, autonomous promotion, or permanent saturation.
The authority boundary
Candidate changes could not directly modify their parent states, the governance-sensitive layer, authoritative external systems, or historical evidence. The machine path could recommend a candidate, but a separate human acceptance authority controlled promotion. This supports a claim about the tested governance architecture; it does not convert the experiment into evidence of unrestricted autonomy.
Claim boundaries across the public profile
This asset concerns one Version 2 empirical preprint and its bounded local experiment. It does not change the status of peer-reviewed journal work, conference work, other preprints, conceptual research, production systems, employment, or advisory work. Those remain separate evidence classes. No external production deployment, regulated use, employment relationship, advisory engagement, or citation count is inferred from this record.
Primary records
- Version 1 DOI record — posted 14 August 2026
- Version 2 DOI record — posted 17 August 2026
- Publisher manuscript record — Version 2 marked as latest and not peer-reviewed
Evidence checked 18 August 2026 (Asia/Jerusalem). The page reports the publisher-deposited record and the Version 2 manuscript's stated experimental boundary; it does not infer results beyond that record.