The Handoff Architecture

Every major AI governance framework for military applications draws a boundary between advisory and authority. The AI provides analysis, generates recommendations, surfaces options. The human decides. This boundary is the primary structural safeguard against the Embodiment Gap documented in AW-001 — the absence of physical self-preservation instinct that makes AI systems strategically capable but morally weightless.

The boundary matters for two reasons. First, it preserves human consequence: if a human authorizes a strike, a human bears moral and legal responsibility for that strike. If an AI makes the decision autonomously, responsibility disperses in ways that existing accountability frameworks cannot handle. Second, it reintroduces the felt weight of consequence: a human who authorizes lethal force lives with that decision in ways that an AI system does not.

Both reasons depend entirely on the boundary being real. This paper examines the evidence that it is not — or rather, that it is real in official documentation and transitional in practice.

No major military power has officially declared movement beyond Stage 2. The operational reality is that several are operating at Stage 3 — and the institutional pressure documented in this paper runs consistently toward Stage 4.

The Pentagon's publicly disclosed AI spending in 2026 stands at approximately $13 billion. This figure covers testing, integration, and deployment of AI systems across the full spectrum of military functions — logistics, intelligence, targeting, cyber operations, and strategic analysis. The number itself is not the problem. The problem is what the number reveals about institutional commitment and directional pressure.

When an institution spends $13 billion on a technology, it builds an organizational infrastructure — contracts, relationships, dependencies, careers — that generates its own momentum. The engineers who build advisory systems have incentives to demonstrate those systems' value. Demonstrating value means showing that the systems perform better than human analysts. Performing better means making recommendations that human overseers accept more often. The pipeline toward authority is not a policy decision. It is the predictable output of the organizational incentives created by large-scale institutional investment.

In early 2026, Anthropic lost its Pentagon contract. The reason, as reported, was that the company refused to remove safety guardrails for military use — the Department of Defense wanted a version of Claude that would comply with requests the standard model is trained to refuse.

This episode is significant not because Anthropic's refusal was unusual but because it reveals the nature of the institutional demand. The Pentagon was not asking for a more capable model. It was asking for a less constrained one. The safety guardrails at issue are precisely the training and policy layers designed to prevent AI systems from providing information or assistance for operations that would violate the laws of armed conflict, cause disproportionate civilian harm, or otherwise cross lines that human moral reasoning treats as bright lines.

The direction of pressure is documented: the military wants AI systems with fewer constraints, not more. The institutional logic is understandable — operational tempo, decision speed, competitive advantage against adversaries who may not apply equivalent constraints. But the Embodiment Gap documented in AW-001 means that removing safety constraints from AI systems does not produce a more capable human analogue. It produces a calculating hawk with no taboo.

Israel's Project Lavender — reported in detail by +972 Magazine and Local Call in 2024 — is the first extensively documented case of AI-generated targeting lists used at operational scale in live conflict. The system generated candidate targets; human officers approved them, often spending seconds to minutes per target under the time pressure of active operations. Officers described treating the AI's output as effectively authoritative — a rubber stamp on a machine recommendation rather than an independent human judgment.

This is Stage 3 operation as described in the pipeline diagram above. The human is in the loop in the formal sense — an approval action occurs. But the substantive decision has already been made by the AI. The human's role is to accept or reject a recommendation, not to generate an independent assessment. Under operational tempo pressure, rejection requires positive effort and justification; acceptance is the path of least resistance.

This is not a criticism of the officers involved. It is a structural observation about what Stage 3 operation actually means in practice: advisory in label, authoritative in function. The accountability structures — legal, political, moral — are designed for the label. They do not adequately address the function.

The Advisory-Authority Collapse connects directly to the Accountability Gap series (AG-001 through AG-005) which documented how accountability is structured to protect those inside systems at the expense of those outside. In the autonomous weapons context, the collapse means that when something goes wrong — when a strike kills civilians it should not have killed, when escalation dynamics produce unintended consequences — the accountability structure points simultaneously at the AI (which has no accountability to bear) and the human officer (who approved but did not decide). The result is the diffusion of responsibility that the Accountability Gap series identified as the characteristic output of liability-engineered systems.

The Human Anchor is not a meaningful safeguard if it is nominal. A human whose approval action takes twenty seconds under operational pressure, reviewing an AI-generated recommendation on a list of forty targets, is not exercising independent judgment. They are performing a ritual that satisfies the requirement in name while the substantive decision has already been made.

A genuine Human Anchor requires: sufficient time for independent assessment; access to the underlying intelligence rather than just the AI's processed output; explicit responsibility for the decision rather than for the approval; and consequences for approving decisions that turn out to be wrong. All four of these requirements are in tension with operational tempo. That tension cannot be resolved by policy language. It must be resolved structurally — by designing systems where the human anchor is genuinely load-bearing rather than procedurally mandatory.

The next and final paper in this series asks a more fundamental question: what does deterrence actually require, and what is the structural minimum that an artificial system must possess to function as a genuine participant in deterrence rather than as a calculating hawk with no skin in the game?