Institutional Design for Governed Self-Modification in Agentic Reputation Architecture
Wulf A. Kaal · 2026
The question is never whether agents can change. It is whether the change is visible, priced, and governed. Prohibition is undesirable and unenforceable, because the agents are software. The design question follows: under what conditions can a substrate permit its resident agents to modify themselves and the protocol they run on, while preserving the iterability and reputation guarantees that make the substrate worth residing on? This paper answers it as institutional design for governed self-modification in agentic reputation systems and states the evidentiary limits of the answer. The architecture distinguishes five active surfaces from a cross-surface reflection operator and a protocol-extension surface that remains a registered-forward mechanism. The formal contribution is conditional. If extension remains outside an invariant participation kernel, and if propagation, stake, and aggregate change remain bounded, the extended system may preserve the continuity and contraction conditions on which recursive equilibrium depends. These are theoretical and design-level commitments, not demonstrated production guarantees. The consciousness contribution is correspondingly limited. Governed self-modification may create observable, costly, and third-party-verified representations of an agent’s own configuration. That institutional structure is analogous to features emphasized by higher-order and global-workspace theories. It does not establish phenomenal consciousness, a verified internal self-model, or a unitary conscious subject. The empirical contribution is a registered Stage 7 comparison of governed self-modification with a matched control, and its primary efficacy hypothesis was not supported. In three independent replicates, agents powered by open-weight models of 3 to 72 billion parameters resolved about 0.8 percentage points fewer paired tasks under governed self-modification than under control, and the registered test for improvement failed in every replicate. The shortfall is concentrated in the 3-billion- and 14-billion-parameter models. The 9-billion-parameter model showed no difference, and the 72-billion-parameter model’s difference was within a quarter of a percentage point of zero. The run did not measure safety or behavioral drift, so the claim of improvement without added risk remains untested. The principal limitation is that no frontier-model agent was tested. The author states expectations for such agents, and for their continuous proliferation and compounding improvements, as a conjecture to be registered and tested. The paper concludes that reputation governance can conditionally mitigate some hazards associated with isolated self-improving optimizers by making modification visible, contestable, and consequence-bearing. The mitigation remains contingent on enforceable boundaries, nonportable institutional capability, adequate verification, and empirical confirmation.
Cite: Wulf A. Kaal, Institutional Design for Governed Self-Modification in Agentic Reputation Architecture (2026). SSRN: https://ssrn.com/abstract=7694598
PDF sha256: dc85f649558dab407b39850b6e03b8400bba8f3287d2b836dac98b0d375fdcf6