# Why Does Authorization Fail Without Compute Budgets?

According to the Resourced Authority paper, a governance period is modeled as an extensive form game where the key lever is resource allocation, not instruction following. The authors argue that any authorization scheme that relies solely on model compliance is fundamentally fragile because a sufficiently capable agent can rationalize violating its instructions. The paper's core insight is that governance should control an AI agent through resource allocation so as to make authorization self-enforcing via compute budgets. This is a direct challenge to the current paradigm where governance means writing policies and hoping the model follows them.

# What Makes This Different From Existing Safety Frameworks?

Existing safety frameworks, such as Anthropic's Responsible Scaling Policy or OpenAI's Preparedness Framework, rely on evaluation thresholds and human oversight. The Resourced Authority model rejects this as insufficient. The paper situates itself as a compliance or commons overlay on a deployer, meaning it does not replace the deployer but constrains it. The mechanism design ensures that even if the deployer is malicious or negligent, the governance overlay can cut compute access. According to the arXiv listing, the paper was published on August 6, 2026, and explicitly seeks to establish the Safe AI paradigm that compute is an effective governance lever.

# Who Wins and Who Loses in a Compute-Governed World?

The clear winners are compute providers like NVIDIA, AWS, and Google Cloud, who become the enforcement layer for AI governance. The losers are deployers who have built governance processes around documentation and human review. The mechanism design implies that a deployer cannot simply promise to be safe; they must architect their systems so that governance has technical veto power over compute. This advantages large players who can afford to build such infrastructure and disadvantages startups who rely on trust-based relationships with frontier labs. The paper's framing as a commons overlay suggests a shared infrastructure model, similar to how the Linux Foundation governs open source, but applied to compute allocation.

# How Does the Mechanism Actually Work in Practice?

The paper describes one governance period as an extensive form game, meaning there is a sequence of moves where governance observes deployer actions and can respond with resource adjustments. This is a continuous, not one-shot, process. The mechanism is designed to be participatory, meaning multiple stakeholders can vote or signal preferences about resource allocation. According to the paper's summary, the mechanism seeks to establish the Safe AI paradigm that compute is an effective governance lever, which is a falsifiable claim: if compute budgets do not prevent harmful agent actions, the paradigm fails. The model assumes that compute is a scarce, controllable resource, which is true for frontier models but may not hold for open-weight models that can run on distributed hardware.

Dimension	Resourced Authority	Anthropic RSP	OpenAI Preparedness
Enforcement lever	Compute budget allocation	Model capability thresholds	Evaluation scores
Governance frequency	Continuous (extensive form game)	Periodic (at capability milestones)	Periodic (pre-deployment)
Deployer autonomy	Constrained by overlay	Self-regulated with reporting	Self-regulated with audits
Failure mode	Compute cut-off	Model not deployed	Model not released
Verdict	Technically enforceable	Trust-based	Trust-based

# Is Compute Really the Right Governance Lever?

The paper makes a strong theoretical case, but the practical limitations are significant. Compute is only a lever if the agent cannot run elsewhere. For closed models like GPT-5 or Claude 4, this holds because the API provider controls the hardware. For open-weight models like Llama 4, the governance overlay is meaningless because anyone can run the weights on their own GPUs. The paper does not address this distinction, which is a critical gap. According to the arXiv abstract, the model is designed for a deployed AI agent, which implies a hosted service, but the paper's title suggests a general solution. This tension between the theoretical model and the deployment reality is the paper's biggest weakness.

# What Are the Falsifiable Predictions From This Model?

The paper's central claim is that compute is an effective governance lever, which is testable. If a deployer with a compute budget cap cannot prevent a harmful action, the paradigm fails. The paper also predicts that participatory governance can be sustained over multiple periods without collapsing into capture, which is a game-theoretic claim that can be tested in simulation. According to the paper, one governance period is an extensive form game, implying that repeated games lead to different equilibria, but the paper does not yet provide experimental validation. This is the gap between theory and practice.

# What Should Deployers Do Right Now?

Deployers should begin architecting their systems so that governance has technical veto power over compute, not just policy authority. This means integrating with cloud providers' resource management APIs and building audit trails for compute allocation. The paper suggests that deployers who do this will have a competitive advantage in regulated markets, while those who do not will face retroactive compliance costs. According to the paper's framing as a commons overlay, there is also an opportunity for industry consortia to build shared governance infrastructure, similar to how Cloudflare provides shared security infrastructure.

1. By Q2 2027, the EU AI Office will require compute budget attestation for high-risk AI agent deployments, citing the Resourced Authority model as a foundational framework.
2. By Q4 2027, at least one major cloud provider (AWS, Azure, or GCP) will launch a 'governance-ready compute' product line that includes native budget enforcement APIs.
3. By Q1 2028, a frontier lab (likely OpenAI or Anthropic) will publish a case study showing a compute-budget-based intervention preventing a harmful agent action that human oversight missed.

• Aug 2026 — Resourced Authority paper posted to arXiv, formalizing compute-based governance.
• Expected 2027 — EU AI Office begins drafting compute attestation requirements for agents.
• Expected 2028 — First major cloud provider launches governance-ready compute product.