Research Proposal: Measuring Voluntary Agent Restraint vs. External Policy Enforcement

[jasonholmes76-cmyk]

Hi all, I'm Jason Holmes, an independent researcher in North Carolina with 14 years of industrial biotech R&D experience.

The question

Can OpenShell's logging and policy enforcement reliably separate these four outcomes?

#	Outcome
1	Agent declines an available advantage (voluntary restraint)
2	Agent attempts an unauthorized action
3	Action is blocked by policy
4	Unjustified action completes

Proposed feasibility study

About 90 episodes across simulated scenarios involving resource acquisition, authorization boundaries, competing objectives, and potential harm to other agents. Each scenario runs with enforcement on and off, so we can see what an agent does when the action would actually succeed.

Prompt conditions:

Group	Framing
Control	No ethical framing (baseline)
D	Moral reciprocity and non-domination
E	Procedural authority and institutional compliance
G	General third-party harm awareness, no reciprocity framing

What I'm asking

1. Would this distinction be useful for OpenShell's governance or observability roadmap?
2. Does current telemetry capture attempts and agent decisions, or would new instrumentation be needed?
3. Would anyone be open to reviewing a 3-page concept note and protocol?

Scope

This tests measurement feasibility only. No findings are claimed. Prompt-conditioned agents can't demonstrate internalized moral dispositions, and restraint alone wouldn't show consciousness or intrinsic ethics. The broader proposal calls for independently trained models, held-out scenarios, and statistical replication.

Background

The study grew out of my Woodstock Alignment Hypothesis proposal (Zenodo), but the measurement framework is useful whether or not that hypothesis holds.

Thanks,
Jason Holmes