On Friday, the United States government banned every federal agency from using Anthropic’s AI technology. The Department of War designated Anthropic a supply chain risk, a classification historically reserved for foreign adversaries. Hours later, OpenAI CEO Sam Altman announced his company had signed a deal with the Pentagon to deploy its models on classified networks.
The public reaction split immediately along predictable lines. Anthropic stood on principle. OpenAI capitulated. Or the inverse: Anthropic was naive about how defense contracting works, and OpenAI was pragmatic about how to serve the mission while maintaining its values.
I want to set both framings aside entirely, because both of them are wrong. Not wrong in emphasis or degree. Wrong in kind. They are arguing about the furniture while the foundation is cracking.
The Facts as They Stand
Anthropic had been the first commercial AI lab to deploy its models across the Pentagon’s classified network, through a partnership with Palantir. Over months of negotiation, the company sought two specific restrictions on how its technology could be used: no integration into fully autonomous weapons systems, and no deployment for mass domestic surveillance of American citizens. The Department of War insisted that AI models must be available for “all lawful purposes” and would not agree to contractual carve-outs.
When talks collapsed, the administration moved with extraordinary speed. President Trump directed every federal agency to cease using Anthropic’s technology, with a six-month phase-out period. Secretary of War Hegseth designated Anthropic a supply chain risk, which bars any military contractor or supplier from doing business with the company. That same evening, Altman announced OpenAI’s deal, claiming it had secured the same two restrictions Anthropic had demanded.
There is a genuine irony here that most commentary has noted: OpenAI appears to have gotten the exact terms Anthropic was asking for. An OpenAI employee publicly questioned whether the safeguards amounted to anything more than window dressing. Altman spent his weekend on social media defending the deal. Claude, Anthropic’s chatbot, surged past ChatGPT to become the most downloaded free app in the Apple Store. Chalk graffiti covered the sidewalks outside both companies’ offices in San Francisco.
None of this is what matters.
The Wrong Layer of the Stack
What matters is that both companies offered the Pentagon fundamentally the same thing: a probabilistic system with policy constraints layered on top.
Anthropic said: we will not allow these use cases. OpenAI said: we secured guarantees against these use cases. In both cases, the underlying models are opaque. They are stochastic. They are unverifiable in the mathematical sense that matters for high-stakes deployment. The restrictions are contractual. The compliance is aspirational. The architecture guarantees nothing.
Altman wrote on X that OpenAI will build “technical safeguards to ensure our models behave as they should.” That is the most important sentence anyone has written about this situation, and I want to take it with the seriousness it deserves.
What technical safeguards? On what architecture? With what mathematical guarantees? What does “behave as they should” mean for a system whose outputs are drawn from probability distributions? How do you verify that a large language model will not perform mass surveillance when the model itself has no formal mechanism for constraining its own behavior at the architectural level?
These are not rhetorical questions. They are engineering questions. They have answers. But the answers require a fundamentally different approach to how AI governance infrastructure is built.
Determinant Judgment and Its Limits
In a previous piece, I wrote about Kant’s distinction between determinant and reflective judgment, and why I believe this distinction will become the most consequential intellectual framework for understanding what AI governance actually requires.
Determinant judgment takes a known rule and applies it to a specific case. You have the concept. You have the phenomenon. You classify accordingly. The Pentagon debate, at its surface level, is a debate about determinant constraints: no mass surveillance, no autonomous weapons without human oversight. These are known rules being applied to known categories. They are the right rules. I support them without reservation.
Both Anthropic and OpenAI are attempting to formalize determinant judgment. GOVERN, the platform Dr. Ben Harvey launched last week, is doing the same at the policy layer: converting laws and regulations into mathematically enforceable logic. This work matters. It is necessary.
But it is not sufficient.
The deeper problem is that determinant constraints only work when the system they constrain is capable of honoring them. A contract that says “this model will not perform surveillance” has exactly as much force as the model’s architecture allows. If the model is a stochastic system with no formal mechanism for constraining its outputs to a defined boundary, then the contract is a statement of intent, not a guarantee of behavior.
This is where reflective judgment enters, and where the entire Pentagon debate reveals its actual stakes.
Who Judges the Oracle’s Judgments?
Reflective judgment operates in the opposite direction from determinant judgment. You encounter a situation for which no existing rule applies, no precedent exists, no framework adequately captures what the moment demands. You must find or create the appropriate framework yourself. Arendt called this “thinking without a bannister.”
The Pentagon AI crisis is a reflective judgment problem masquerading as a determinant one. Everyone involved is acting as though the question is: which company will agree to the right rules? That is the determinant framing. The actual question is: does any current AI architecture have the capacity to enforce rules at the level these decisions require?
The government is not wrong to want AI deployed in defense. The mission is right. AI will reshape national security, and the United States should lead in deploying it responsibly. ARX is built in direct alignment with U.S. national priorities around advanced compute and AI governance. I believe the government has both the authority and the responsibility to deploy these systems. That is not the issue.
The issue is that selecting AI vendors based on contractual willingness rather than architectural capability is a category error with consequences measured in national security outcomes. It is the equivalent of choosing a bridge contractor based on who will sign the liability waiver rather than whose engineering can bear the load.
The Structural Difference
I am building the Vector Translation Matrix as deterministic architecture because I believe the gap between contractual compliance and architectural compliance is the most consequential gap in AI governance today.
The difference is structural. In a probabilistic system with policy constraints, compliance is a behavior that is hoped for and monitored after the fact. In deterministic architecture, compliance is a mathematical property of the system itself. It is not something the system chooses to do or can be persuaded not to do. It is a property of the architecture in the same way that the angles of a triangle summing to 180 degrees is a property of Euclidean geometry. It holds because the structure requires it.
That is the difference between enforcing rules on top of a system that might behave unpredictably and building a system where the rules are embedded in the mathematics from the ground up.
Altman says OpenAI will build a “safety stack,” a layered system of technical, policy, and human controls that sit between the model and real-world use. If the model refuses to perform a task, the government agreed it would not force OpenAI to comply. This is better than nothing. But it is still a system in which the model’s behavior is a matter of prediction and hope, not proof.
What This Moment Requires
The convergence I wrote about last week is accelerating. Harvey’s GOVERN addresses the policy-to-logic layer. The broader industry is moving toward deterministic governance as a category. The Pentagon crisis is accelerating that movement not because anyone involved intended it, but because the crisis has made the inadequacy of the current approach impossible to ignore.
The fact that credible builders are converging on this problem from multiple directions, policy logic, mathematical architecture, formal verification, is the signal. The fact that the most powerful government on earth is currently selecting AI vendors through political dynamics rather than architectural evaluation is the urgency.
Both of those things can be true simultaneously. The mission can be right and the method can be wrong. That is the nature of reflective judgment: the capacity to hold complexity without reducing it to a false binary.
The question is not which AI company deserves the Pentagon contract. The question is whether the architecture can keep the promise the contract makes. Until that question has a provable answer, we are negotiating the terms of trust with systems that have no mathematical mechanism to honor them.
That should concern everyone on every side of this debate.
Read the previous piece on AI governance and the philosophy of judgment
Sukh Sidhu is the Founder and CEO of ARX, building the stateful runtime layer for enterprise AI. ARX aligns with U.S. national priorities around advanced compute and AI governance. He can be reached at info@arxqm.com.
Sources:
- Anthropic Banned from Federal Use — The New York Times, February 28, 2026.
- OpenAI Signs Pentagon Deal — Reuters, February 28, 2026.
#AIGovernance #NationalSecurity #DeterministicAI #Pentagon #Anthropic #OpenAI #DefenseAI #AICompliance #AIArchitecture #AIEthics #ARX