The Competence Circuit Reframes AI Governance Around Control and Knowledge Boundaries

Summary:An arXiv preprint introduces the “competence circuit” (IKC + KC + C), a framework that decomposes an autonomous agent’s real-world impact into intelligence-mediated discovery, procedural know-how, and direct control. The paper formalizes these pathways through the fast-takeoff debate and evaluates them through stochastic simulations on NK fitness landscapes across six scenarios. Its central result is that intelligence-driven acceleration is bounded, while knowledge-acquisition speed, capability-

---

Intelligence Is Not the Same Thing as the Speed of Harm

The central argument of this research is not a new safety benchmark, but a decomposition of how an autonomous agent converts capability into real-world impact. The authors describe this as the competence circuit, summarized as IKC + KC + C: intelligence-mediated discovery, procedural know-how, and direct control. The distinction matters because “knowing what,” “turning knowledge into an executable process,” and “being authorized to change the world” have different bottlenecks and different controls. A model may discover a dangerous method without being able to manufacture or acquire the required resources. It may also know an established procedure while lacking permission to trigger it.

How the Three Pathways Interact

The framework separates the impact chain into three components:

1. IKC: intelligence-mediated knowledge acquisition. Intelligence accelerates hypothesis selection, information discovery, and search.
2. KC: knowledge capability or procedural know-how. Discovered knowledge must become an executable process involving manufacturing, procurement, deployment, and error correction.
3. C: control. The agent or an associated actor must possess effective authority over physical devices, accounts, funds, supply chains, or other systems.

Fast takeoff is formalized as a conjunction: the relevant pathways must jointly hold for capability gains to translate into rapid and potentially catastrophic real-world advantage. This is more operationally useful than attributing risk to a single intelligence metric because real escalation is usually limited by information, procedural, and authorization bottlenecks at the same time.

Intelligence-Driven Acceleration Has a Ratio Bound

The authors operationalize epistemic intelligence as hypothesis-selection efficiency and evaluate the framework through stochastic simulations on NK fitness landscapes. The NK model is well suited to studying combinatorial optimization and rugged, multi-peaked search spaces because it allows interaction strength and search difficulty to be varied independently.

Across six scenarios spanning the conjunctive chain, the paper derives an analytical bound supported by the simulations: intelligence-driven acceleration is capped by the ratio between competing intelligence levels, regardless of task difficulty. The practical implication is that a stronger searcher does not shorten every discovery task by the same proportional amount; once search approaches a local bottleneck, additional intelligence produces diminishing returns.

This does not mean intelligence is irrelevant to risk. The bound limits the marginal acceleration supplied by discovery. It does not remove the time required to convert a hypothesis into a working process or authorized action.

A Large Multiplier Does Not Imply a Short Absolute Timeline

The simulations expose a fast-takeoff dilemma. In a noisy hypothesis space, epistemic intelligence can produce a large relative speed multiplier, allowing an agent to eliminate unproductive search and identify useful hypotheses faster. Yet the absolute timeline can remain constrained by capability-critical tests: whether a material meets a threshold, whether equipment can be manufactured reliably, whether a supply chain permits delivery, or how many correction cycles are required.

A system might therefore achieve a 20-fold reduction in search time while a single physical validation still takes months. The deployment schedule remains dominated by validation rather than search. Safety analysis should consequently report the full distribution of discovery, validation, approval, and deployment times, not only a headline productivity multiplier.

Authorization Bottlenecks Can Dominate

The simulations also find that institutional chokepoints may constrain superintelligent agents disproportionately. High-intelligence agents spend very little time on unproductive search, but a mandatory approval process can impose a long fixed delay. In that regime, authorization time becomes the dominant component of the end-to-end task duration.

This creates an asymmetry across capability levels. Low-capability systems may appear governed because they are slow anyway, while high-capability systems rapidly reach the point where external approval becomes the main bottleneck. Adding a generic human-review box is not automatically safer: it can lengthen legitimate workflows without addressing side channels or creating a usable delay window for adversaries. Governance should therefore measure action-level authorization latency, not just refusal rates or offline safety evaluations.

Enforcing Control and Knowledge Boundaries

The paper recommends prioritizing two levers at every capability level.

Restricting control means narrowing the action space and lowering the atomicity of permissions. Execution layers should distinguish read-only recommendations, reversible low-risk operations, irreversible operations, and emergency actions instead of exposing an agent one unified toolkit containing cloud accounts, code execution, device controls, and payment interfaces. For high-impact actions, identity verification, dual authorization, spending limits, rate limits, and complete audit trails are more actionable than a generic claim of human oversight.

Enforcing knowledge boundaries means restricting what an agent can discover, retain, and recombine, rather than merely filtering its final answer. Prompt-level restrictions are insufficient when agents can use search tools, reorganize memory, or exchange information through other agents. A robust boundary must constrain data sources, retrieval scope, long-term memory, tool outputs, and inter-agent communication.

The two controls are complementary: knowledge boundaries reduce the probability that dangerous capabilities are discovered, while control boundaries limit amplification after discovery.

What “Putting Physics in the Loop” Means

The paper observes that purely digital domains weaken physical and institutional bottlenecks. Agents can rapidly duplicate code, invoke cloud services, and generate hypotheses in parallel, while manufacturing cycles, supply chains, payment approvals, and device safety impose additional delays. The authors therefore favor architectures that deliberately put physics in the loop, requiring risky actions to pass through physical confirmation, controlled fulfillment, human-held authorization, or regulated service calls.

This is not unconditionally beneficial. Physical steps add cost, latency, availability constraints, and potential single points of failure, and they can create social-engineering surfaces. The appropriate design is to use the smallest sufficient real-world constraint for each risk class rather than inserting hardware uniformly into every workflow.

Deployment Recommendations

For teams building autonomous agents, the framework suggests a more operational evaluation matrix:

1. Measure knowledge-discovery speed, procedural capability, and control authority separately instead of compressing them into one intelligence score.
2. Track discovery, experimentation, validation, authorization, and deployment as separate phases.
3. Include approval latency alongside model inference and search latency in end-to-end safety metrics.
4. Enforce control boundaries through least privilege, action-level authorization, rate limits, and auditable logs.
5. Enforce knowledge boundaries across tools, memory, and agent-to-agent communication.
6. Place selected high-risk chains behind non-bypassable physical or regulated execution points.

Limits of the Evidence

The work provides a conceptual framework, analytical bounds, and stochastic simulations rather than a large-scale evaluation of agents in open environments. NK landscapes are useful for isolating search and interaction mechanisms, but they do not fully represent software engineering, supply chains, intelligence operations, or institutional dynamics.

The available abstract does not disclose the complete parameters for all six scenarios, confidence intervals, agent baselines, simulation repetitions, or sensitivity to modeling assumptions. The proposed conditions for intelligence-mediated fast takeoff should therefore be treated as testable hypotheses rather than causal laws already confirmed by real incident data. Full code, complete experimental artifacts, and independent replication remain necessary before applying the quantitative bounds operationally.

Even with those limitations, the governance ordering is actionable: identify what an agent may learn, what processes it may construct, and what resources it may control. Governance can then target the dominant pathway instead of treating every safety problem as an intelligence-limitation problem.

---

Source: ArXiv AI (cs.AI) (2026-11-11)

— END —

Tags: #AI Governance Follow #Autonomous Agents Follow #Competence Circuit Follow #Knowledge Boundaries Follow #Control Security Follow