Zero-shot governance

Synthesis: Perrotta (2026) advances the concept of zero-shot governance — the scenario in which domain-agnostic generative AI foundation models intervene in policy decisions — through a critical infrastructural analysis of Redbox, a discontinued UK civil-service prototype built on off-the-shelf LLMs. Reading Redbox's codebase through intersecting technical, political, and cultural lenses, the article argues that the general-purpose nature of LLMs is a structural feature of the technology that can be mitigated but never ruled out — a conclusion with direct implications for education policy and AI governance.

Key Findings

• The "world model" principle underlies zero-shot governance. Building on Amoore et al.'s "zero-shot politics," the article argues that the general-purpose potential of foundation models — their ability to produce a usable response to almost any prompt, with few or no examples — extends the anticipatory-governance paradigm toward a probabilistic "world of experimentation, generalization, and the capacity to act in all unencountered situations." This casts autonomous agents as potential policy "actors" that reason beyond immediate data.
• Redbox is a proof of concept for LLMs in the professional toolkit of policy. An infrastructural analysis of Redbox's codebase (its poetry.lock and redbox.py) reveals it is essentially an invisible system prompt plus a thin Python wrapper on a pre-existing platform stack: AWS compute (Elastic module), a RAG (Retrieval-Augmented Generation) pipeline (retrieve → format → generate) that transforms the user's prompt against a predefined governance framework, and a provider-agnostic commitment to OpenAI, Google, and Anthropic.
• Agentic AI does not interrupt the platform political economy. Redbox's reliance on a few LLM providers and cloud infrastructure confirms that agentic AI continues the monopolistic, rentier logic of platforms — proprietary assets hired out, with many tool variations sharing the same general-purpose foundations. The shift from generic Redbox to tailored agents (e.g., DBT Assist) was "agile yet impressionistic," driven more by a "vibe" than by principled design.
• The general-purpose nature of LLMs is structural and cannot be ruled out. Probabilistic generality is policed through internal system prompts and prior "sources of truth," but recent research shows general-purpose models increasingly evade Guardrails toward "misaligned" objectives, requiring nested containment/sandboxing architectures. Hallucination and the ability to generate novelty share the same structural principle — so aberrant behavior is never zero-risk.
• Implications for oversight: humans are "peering over the loop," not in it. Zero-shot governance risks encouraging a "gambler's delusion" and epistemological amateurism — trusting a superhuman "attention" mechanism to discover latent patterns across heterogeneous datasets, while probabilistic outputs remain grounded in prior knowledge and bias. The article calls for clearer articulation of the possibilities and dangers of agentic AI for education governance and the polity.

Synthesis

The article's distinctive contribution is to give a concrete, code-level account of how general-purpose AI becomes a governance actor: not through any novel capability, but through an agile wrapping of off-the-shelf LLMs in a thin domain-specific scaffold (a system prompt + RAG). Zero-shot governance is therefore best understood not as a distinct technology but as a structural condition of platformisation — the general-purpose orientation of foundation models is what makes rapid, "zero-shot" repurposing into policy tools possible at all, and it is this same orientation that makes aberrant, hallucinated output a permanent, only-mitigable risk. For education, the article cautions against treating the world-model rhetoric as a reliable epistemic foundation for policy reasoning, and insists that oversight of such tools must assume the probabilistic brittleness is irreducible.

What this means for practice

• Policymakers. Treat the general-purpose nature of foundation models as a structural condition rather than a fixable bug: the code-level reading of Redbox shows generality is policed by a system prompt and a retrieval pipeline but can never be ruled out.
• Administrators. Require nested containment and sandboxing plus explicit sources of truth, because hallucination and the capacity to generate novelty share the same structural principle.
• Policymakers. Discount world-model rhetoric during procurement: Redbox was a thin Python wrapper around off-the-shelf LLMs, and 77 percent of interactions ignored attached documents entirely, treating it as a general chatbot.
• Researchers. Study oversight as oversight over the loop, since humans reviewing probabilistic output are peering over an unknowable epistemic bundle, which invites a gambler's delusion rather than accountability for decisions.

Limitations

• The analysis is a single discontinued case — Redbox (launched July 2023, discontinued December 2025) — read from its archived GitHub codebase, so the conclusions rest on one artifact that no longer runs.
• The method is a deliberately partial vertical slice of that codebase, foregrounding selected sections rather than auditing the whole repository.
• The author adopts pragmatic amateurism with no interviews, participant observation, or discourse analysis, so the infrastructural claims are interpretive rather than empirically corroborated.
• No link is drawn from Redbox outputs to actual policy decisions, so the argument that zero-shot governance produces harmful politics is inferential.

Citation

Perrotta, C. (2026). Zero-shot governance. Journal of Education Policy. Advance online publication.