Auditability Requires Openness

By the Founders, Department of Technology | San Diego, California |

By the Founders, Department of Technology | San Diego, California | November 20, 2026

Machine intelligence systems are becoming critical infrastructure. The more powerful and autonomous they grow, the more important it becomes that their behavior can be examined, constrained, and accounted for. Yet most of the software stack on which these systems run remains closed. That closedness is not a minor implementation detail. It is a structural obstacle to meaningful auditability.

If safety constraints, resource limits, and audit trails are to be real rather than aspirational, the systems that enforce them must themselves be open to scrutiny. Auditability requires openness.

The core point

Kernel-level safety properties, resource accounting, and audit mechanisms cannot be independently verified if the code and design that implement them are proprietary. Openness is a necessary condition for trust at the foundation of machine intelligence systems.

Why closed stacks fall short

Today’s dominant approach places advanced machine intelligence on top of general-purpose operating systems and proprietary accelerator software. Safety features, monitoring, and resource controls are typically added later—in user space, in libraries, or in external policy layers. These mechanisms can be useful, but they sit outside the core control plane of the system. When the underlying platform is closed, independent parties cannot fully examine how those controls are implemented, whether they can be bypassed, or what happens under failure or adversarial conditions.

This creates a credibility gap. Claims about safety and accountability become difficult to evaluate from the outside. Governments, researchers, and the public are asked to trust assurances that cannot be independently checked at the lowest levels of the stack.

What kernel-level design makes possible

A purpose-built Machine Intelligence Operating System (MIOS) starts from a different premise: neural and tensor workloads are primary, and the operating system should be designed around them. In principle, such a system could treat safety constraints, resource accounting, and auditability as first-class kernel properties rather than after-the-fact additions.

That architectural shift only delivers its full value if the design and implementation are open to examination. Formal verification, independent security review, and public confidence all depend on the ability of qualified third parties to inspect the mechanisms that claim to enforce limits and record behavior. Closed implementations can still be useful inside a single organization. They cannot, by themselves, support the broader claim that the foundation is trustworthy.

Openness is necessary, not sufficient

Openness does not automatically prevent misuse. Source code that is public can be forked, modified, or deployed without the safety mechanisms intact. An open MIOS would make strong controls possible and auditable; it would not guarantee that every deployment uses them.

The correct claim is therefore modest and precise: openness is a necessary condition for meaningful auditability at the foundational layer. It is not a sufficient condition for safety. Treating it as the former strengthens the case. Treating it as the latter weakens credibility.

An open, multi-stakeholder path

For these reasons, MIOS should be pursued as an open, internationally collaborative research and development effort. No single vendor or nation should control the foundational software layer on which advanced machine intelligence systems will increasingly run. An open approach allows:

  • Independent examination of safety and accounting mechanisms
  • Broader participation by academic and public-sector researchers
  • Reduced risk of lock-in around proprietary stacks
  • Clearer pathways for governments to require verifiable properties

This is not a call for a single global project under one bureaucracy. It is a call for the core architecture and reference implementations to be developed in the open, under governance that permits scrutiny and contribution from multiple stakeholders.

Shared responsibility

Academia can stress-test the technical claims and advance formal methods. Industry can evaluate performance, practicality, and economic viability. Governments can assess whether public-interest requirements—security, energy accountability, resilience, and auditability—are being designed into the foundation while that foundation is still open to influence.

None of these communities needs to accept every detail of any particular proposal. All of them have reason to insist that the lowest software layer for machine intelligence be capable of supporting real auditability. That capability is difficult to achieve, and impossible to verify, if the layer itself remains closed.

Auditability requires openness. The sooner that principle shapes the foundations of machine intelligence systems, the less likely we are to discover later that the most important controls were never truly inspectable.

The foundation is still open to design

Machine intelligence is being deployed faster than the software foundations beneath it are being re-examined. The conversation about those foundations should be public, technical, and serious.

Explore more articles

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *