Accountability Architecture

No fielded system needs to pull a trigger to matter legally. Decision-support software chooses which sensor data is fused, classifies what a camera sees, ranks possible targets, recommends weapon pairings, plans routes and guides munitions in their final seconds. Each function shapes what a human decides, how fast, and on what picture of the world.

Independent analysis of current practice finds exactly this: the locus of lethal decision-making is moving into software, while target recognition and navigation assistance are already routine and decision cycles compress to tens of seconds.

The chain below follows a strike from training data to investigation. Select a stage to see the decision being made, who actually controls it, the evidence that should exist, the law engaged — and where accountability typically fails.

The problem is not an empty statute book. International humanitarian law governs the conduct of hostilities. Article 36 of Additional Protocol I requires legal review of new weapons, means and methods of warfare. Export licensing can halt transfers where risk thresholds are met. International criminal law reaches individuals; domestic law reaches companies in defined ways; contracts can require almost anything a buyer insists on.

The difficulty is that these regimes were built for different objects — a shell, a shipment, a soldier, a company — and a continuous, software-mediated decision system is all of them at once, changing monthly under commercial confidentiality.

The explorer below maps eight legal layers against the system lifecycle, from design to remedy. Each cell answers five questions: who owes the duty, what triggers it, what evidence it needs, what consequence follows, and where the gap is.

Every layer of the architecture is real. The fragmentation happens between them. The state owes IHL duties but relies on a supplier's undisclosed model. The weapons review approved version 3; version 11 is deployed. The export licence assessed hardware; behaviour lives in software updates. Criminal law needs a person with knowledge and contribution; the knowledge is distributed across a company and the contribution across a supply chain. The contract could have required logs; it did not.

No single actor designed this outcome, and no single reform fixes it. Duties are split across bodies of law with different objects, thresholds and investigators — and the technical evidence that would connect them is precisely what current arrangements fail to preserve.

That is the accountability gap: not a missing law, but missing architecture between laws.

The problem is not that no law exists. Duties are split across different bodies of law, responsibility is easily fragmented, and technical evidence may not survive.

In 2026, n-Space staged Trial Trial at Somerset House Studios: a public-interest simulated criminal trial of a fictionalised case assembled from documented reporting about AI-assisted targeting. Practising lawyers argued it; expert witnesses were cross-examined; a citizen jury deliberated.

A simulation cannot convict anyone, and this one did not try. Its purpose was adversarial stress-testing: when a legal framework must survive cross-examination — with a named defendant, specific evidence and a jury asking plain questions — its gaps stop being abstract.

The exercise asked when a supplier's contribution becomes legally relevant; what happens when the state controls the strike but the supplier controls parts of the decision environment; what credible notice should oblige; whether the system's role could be reconstructed after civilian deaths; and whether nominal human approval survives realistic tempo. The findings below are research findings, not judgments.

Four recurring factual patterns do most of the damage to the current framework. Each is presented the same way: the pattern, the law engaged, the evidence that goes missing, what currently happens, and the safeguard that would change the answer.

They are deliberately mundane. None involves a rogue robot; all involve ordinary institutional behaviour — throughput pressure, software updates, commercial awkwardness, incomplete records — meeting rules written for different objects.

Nine kinds of actor sit around a single strike, and several forms of responsibility can attach to the same event without cancelling one another: state responsibility under international law, command responsibility, individual criminal liability, corporate liability, contractual and regulatory consequences.

The map below shows, for each actor, the core responsibility, the knowledge-or-control question an investigator would ask, the evidence that answers it, and the form accountability could take.

Two errors to avoid. Responsibility is not a single parcel passed down the chain until it reaches whoever touched the interface last. And ordinary engineers and employees are not liable merely because they worked on a product — individual liability follows personal knowledge, conduct and authority.

The hard part is not drafting duties; it is the structure of the relationship. Governments depend on strategic suppliers for capability and technical knowledge. Procurement, national-security and industrial policy all reward preserving those relationships. Classification and commercial confidentiality limit what outsiders can inspect. Rapid procurement can lock a supplier in before assurance capacity matures — and when harm occurs, liability drifts toward the operator, the actor with the least upstream control.

The counterargument deserves stating fairly: duplicative, poorly designed regulation delays urgent capability, favours incumbents who can afford compliance, and excludes smaller firms. Maximal paperwork is not the goal.

The answer this site argues for is lean, technically literate assurance tied to risk and function — designed so that clear rules reduce uncertainty for government and responsible industry alike, and make lawful innovation easier to demonstrate, not harder.

Eight practical measures, built on the levers government already holds. Procurement is the first enforcement layer; criminal law is the backstop, not the plan.

Together they make one change: they convert assertions — about human control, testing, updates and supplier conduct — into evidence that exists before anyone needs it.

National assurance needs an institution: a technically competent function, controlled by neither the procuring team nor the supplier, with access to design records, versions and test evidence, and authority to require re-review when material changes occur.

Beyond it, this site outlines an International Autonomous Weapons Agency — an IAEA for high-risk military-AI functions, but not a copy of it. Its remit would be narrow: target selection and prioritisation, engagement without a new human decision, lethal swarms, machine-speed coupling, and material changes to certified systems. Its powers would be bounded: standards, confidential registration, review of national processes, laboratory accreditation, inspections, incident investigation, export verification, sanitised findings, referral.

Its limits must be stated as plainly as its promise: non-parties, secrecy, proxy use, covert software, unequal capacity — and the standing risk that certification becomes political legitimation. It does not exist, and nothing here presents it as inevitable.

The committee-facing recommendations are listed in full on the recommendations page; the eight measures above are their substance. What they amount to is a single test that procurement, ministers and Parliament can each apply.