Toward Agentic Governance: What Shapes LLM-Agent Intervention in Public Forums?

TEXT START: LLM agents are increasingly used in moderation-relevant public forum workflows, where their choices to answer, acknowledge, repair, or decline are routinely challenged by users, platforms, and regulators.

THE DISSECTION

This paper documents a structural crisis in AI governance integration. LLM agents are already embedded in public forum moderation workflows, making intervention decisions that are routinely challenged—yet the same agent returns different responses to identical content because four deployment factors (model version, weight-release status, provider, system-prompt policy) independently shift behavior. The paper's central empirical finding: closed-weight models systematically decline more on visible challenges; open-weight models reverse this or show no gap.

THE CORE FALLACY

The paper frames this as a governance design problem—something correctable through better auditing, awareness of deployment choices, and auditable forum-agent governance protocols. This is the fallacy. The variation isn't a bug in the system; it's the system. When intervention behavior shifts based on invisible factors—including which model version is currently served, which can change between calls without notice—you have a governance infrastructure that cannot produce reproducible accountability. You cannot audit your way out of a structural impossibility.

HIDDEN ASSUMPTIONS

1. That public forum moderation is the appropriate domain for AI governance agents.
2. That consistency is achievable if operators simply track all four deployment choices.
3. That human institutions can meaningfully govern AI agents that change behavior mid-operation without notice.
4. That "auditable governance" solves the accountability problem rather than relocating it.

THE SOCIAL FUNCTION

This paper occupies a curious niche: it's simultaneously a technical contribution and a document of institutional failure. It describes AI systems already embedded in governance roles, documents their structural instability, and then offers an auditing framework as a solution. It's transition management theater. The real message is buried: governance infrastructure is already AI-mediated, it's unstable, and the proposed fix (awareness of deployment choices) cannot work when model version changes between calls.

THE VERDICT

This paper is forensic evidence of the governance integration problem, but it misdiagnoses the nature of the disease. The four deployment factors producing behavioral variance aren't engineering problems to be audited away—they're the mechanism by which AI systems become ungovernable. When a moderation decision can be shifted by switching the provider or model version without operator awareness, you have exited the domain of accountable governance entirely. The open/closed weight finding is the telling detail: closed models avoid visible challenges (reputational risk management), while open models don't (no brand to protect). This isn't a governance consistency finding—it's evidence that AI governance is being shaped by commercial risk calculus, not democratic accountability structures. The paper documents the collapse of consistent governance and offers an audit framework as a memorial.