Every brief on this site relies on the same implicit practice: assess the evidence, flag the certainty, decline to advocate. On one subject that practice is under pressure from three directions at once — the capability claims come almost entirely from interested parties, the governance question is live and consequential, and the subject matter is people being killed — and an implicit practice under that much pressure is not a standard. It is an instinct, and instincts are not checkable.
So for FR-VIII-23 · Military AI and Strategic Stability, the practice is written down. Adopted 15 September 2026.
The order this happened in, stated plainly
The honest sequence matters more than the tidy one. This standard was drafted to be adopted before the subject was commissioned, and it was not: the brief was researched, written and published first, and the standard is being adopted after it went live. That is the wrong order, and saying so is cheaper than pretending otherwise on a page about editorial discipline.
What was done about it is at the foot of this page: the published brief was audited against all four commitments below, clause by clause, before this page was adopted, and the audit’s result — including the one line that sits closest to the boundary — is published here rather than filed internally. A standard adopted after the fact is worth something only if the thing it governs is then checked against it in public.
The four commitments
These are constraints on the brief, not opinions held by the Institute. Each one is written so that a reader can catch a violation without access to anything but the published page.
1 · The subject is the evidence, not the policy
What is technically demonstrated, what is claimed, what is measurable, what is unresolved. Where a position is contested — what meaningful human control requires, whether autonomy compresses decision time in practice, whether an autonomy limit can be verified at all — the brief presents the strongest form of each position and says which evidence would discriminate between them. It does not conclude which policy is right.
The boundary, because this is the commitment most easily blurred: saying that a claim exceeds its source is assessment; saying a treaty should be signed is advocacy. Saying that a framework certifies a variable it does not measure is assessment; saying a state should adopt a different framework is advocacy. The first of each pair belongs on the page. The second does not.
2 · Interested parties are flagged harder here than anywhere else
Almost every capability claim in this field originates with a defence contractor, a defence ministry, or an advocacy organisation, and each of the three has a structural reason to overstate or understate. No capability claim is carried at Established on a vendor or ministry source alone. Independent corroboration — a UN panel, a peer-reviewed study, an accountability body, investigative reporting that names its sourcing — is what moves a capability claim to that flag, and where it does not exist the claim stays at Frontier or below, however confidently it is asserted elsewhere.
3 · The brief carries no operational content
No targeting methods, no countermeasures, no architecture or parameters that would function as a design note. The line the brief is written to: readable by a legislator deciding a treaty position, and useless to someone building the thing. This is a constraint on how the subject is treated, not a reason to avoid it. The governance question is important and badly under-served, and the way to serve it is analysis, not engineering.
4 · Human cost is stated, not aestheticised
Where the evidence concerns casualties or civilian harm it is reported plainly and sourced — without the technical register that turns people into an outcome variable, and without rhetorical amplification. Both failures are failures of the same kind: they substitute the writer’s effect for the reader’s judgement.
Why assess-only, and the argument against it
The choice that had to be made was commitment 1, and it is a real choice rather than a safety constraint. The case against assess-only is serious and deserves its strongest form: a research institute that lays out the evidence on autonomous targeting and then declines to say anything about where that evidence points is not being neutral, it is being less useful. Refusing to reach a conclusion is itself a position, and on a subject with this much at stake it is the position that costs the reader the most.
The Institute has taken the other view, and the reason is specific rather than general. The four-flag system’s entire value is that a reader can trust it not to be arguing — that a claim marked Established is marked that way because of its evidence and not because of where it points. That trust is the asset the whole corpus is built on, it is spent the first time a reader catches the flags carrying an argument, and the most contested subject in the corpus is the most expensive place in the world to spend it. A reader who wants the Institute’s view on the treaty is better served by a brief whose evidence they can check than by a verdict they cannot.
This is a position about this corpus, not a claim that assessment is always superior to advocacy. Organisations that advocate on this subject do necessary work, and several of them are cited in the brief. The Institute is not one of them, and the page says which it is.
On escalation and nuclear command
This is where speculation is cheapest and the consequences are largest, so one further constraint applies. The brief is unusually willing to say that the evidence does not exist, and treats the absence of a public evidence base as a finding rather than as an invitation to model. A scenario is not a source. Where the honest statement about nuclear command and control is that the arrangements are protected mainly by declarations that nobody has agreed to verify, the brief says that, rather than producing a plausible architecture nobody can check.
What this standard is not
It is not avoidance. The subject is covered, at the same depth as every other brief in the corpus, with the same fifteen-section framework and the same flags. It is not a claim of neutrality about outcomes either: reporting that no individual has ever been prosecuted for an autonomous or decision-support-mediated targeting error is not a neutral act, and the Institute does not pretend it is. What the standard commits to is narrower and checkable — that the brief’s selection, flags and conclusions answer to its evidence, and that a reader who disagrees with the Institute about policy can still use the page.
What the audit of the published brief found
Every claim carrying the Established flag was checked against commitment 2. All of them are documentary or procedural facts — the negotiating record and the consensus rule at the CCW, stated doctrine, a UN Panel of Experts finding, the human-factors investigations into the Patriot fratricides — rather than capability claims resting on a vendor. The single most over-cited capability claim in this field, that the Kargu‑2 strike in Libya was the first confirmed autonomous kill, is carried in the brief as a correction: capability and use were documented, casualties attributable to autonomous mode were not, and the brief says the claim exceeds its source.
Commitment 3 holds. The engineering section discusses controllability at the level a legislator reads — target-type limits, geographic and temporal bounds, deactivation that survives jamming, architectural separation between advisory systems and release authority, tamper-evident logging — which is the published position of the ICRC and of the arms-control literature, and is not a design note.
Commitment 4 holds. The Vincennes shootdown is reported as 290 people killed; the Patriot fratricides as three aircrew killed; the reported twenty-second reviews in the Lavender investigation are given with their sourcing and its limits.
One line sits close to the boundary of commitment 1, and is published here rather than quietly kept. The ethical section states that ethics frameworks certifying the presence of a human without measuring the quality of that human’s engagement “certify the wrong variable”. That is a judgement, and the Institute judges it to be a judgement about measurement rather than about policy — it says a proxy fails to track the thing it stands for, which is an evidential claim and is supported by the automation-bias record the same section cites. It does not say which framework a state should adopt. A reader who thinks that line crosses into advocacy has a reasonable argument, and the way to make it is to the Institute directly; it would be a correction, not a difference of taste.
Scope
This standard binds FR-VIII-23 and any future brief whose subject is autonomous weapons, military AI, or nuclear command and control. It sits on top of the practices that apply to the whole corpus — the four-flag system, the review policy, and the citation record — and does not replace any of them. Briefs bound by it carry a link to this page in their header, so the standard is one click from the page it governs rather than filed somewhere a reader would have to know to look.