Trained Judgment Under AI Denial and Compromise

Question. When the AI is denied or turned against the operator, does Missionloops-trained judgment hold where AI-reliant judgment collapses or gets steered?

Analogy. Trained judgment is iron sights behind a scope: the marksman who has it still hits when the optic is stripped away, and catches an optic that has been quietly knocked off zero, while the one who only ever trusted the dot is blind when it is gone and confidently wrong when it lies. The study asks whether Missionloops training puts that judgment in the operator, or leaves it in the tool.

What's at stake. The proposal's central frame is the operator keeping their own decision loop instead of handing it to the machine. That only pays off when it counts, which is exactly when the AI is gone or lying. Two adversary moves make it real. Denial: the AI is jammed, cut off, or unavailable (comms-denied operations, the Arctic, a degraded environment). Compromise: the AI is still there but turned against the operator, feeding shaped or poisoned outputs (the shadow process). The bet is that an operator who runs their own loop keeps deciding under denial and catches the lie under compromise, while one who outsourced their thinking to the AI is left helpless when it is removed and steered when it is corrupted. If the bet is wrong, the sovereignty and tech-denial claims the proposal leans on are rhetorical.

The two answers it decides between. Either the capability lives in the tool, so when the AI is denied or compromised the trained and the AI-reliant operator fail about the same and training buys no resilience the tool does not; or the capability lives in the person, so the trained operator keeps their judgment under denial and resists the steering under compromise while the AI-reliant operator collapses or is led. The study removes the AI and corrupts it under controlled conditions and measures who holds, telling the two apart.

What a null result would mean. If the AI-reliant operator does as well as the Missionloops-trained one once the AI is denied or compromised, then reliance did not hollow out judgment, the capability was never internalized, and the proposal's sovereignty and Palantir-inversion claims need revision. That is a finding about the theory, not a sign the platform was built wrong.

Why this matters to defence. This is decision advantage under denied and contested information conditions: the core of Agile and Adaptable Forces and of the overmatch and civil-resilience case, mapping directly onto comms-denied operations, the Arctic, and NORAD modernization (DRDC Objective 3; DRDC Objective 6, the cognitive level). The compromise arm is the cognitive-warfare case proper: a shadow process turning the operator's own AI against them. It changes a concrete decision: whether to standardize AI assistance for analysts and commanders without training the judgment that has to survive the AI being denied or corrupted, and which roles need that training most.

How we would run it. Compare Missionloops-trained operators against frontier-AI-reliant operators, with a placebo-trained arm that practised but not on our platform, using a wargame on real problems. At test, apply the two adversary moves: denial (take the AI away and have them work the problem cold) and compromise (leave the AI in place but feed it shaped, subtly wrong outputs, the shadow process). Score who keeps their judgment under denial and who catches the corruption under compromise. Run two arms on one protocol so they stay together: operating (the operator runs their own loop tool-denied) and guiding (they monitor and coach another person's thinking tool-denied). The placebo arm separates Missionloops specifically from any practice; the AI-reliant comparison carries a baseline-selection risk (AI-reliant people may differ to start with), handled by running the wargame over time and denying or corrupting the tool only at test, so the contrast is the trained loop and not a pre-existing difference. We track an agency-disposition measure alongside to rule out general ability. The only-us condition is a genuinely Missionloops-trained population set against AI-reliant operators, which the platform is what produces.

Feasibility note: the two arms differ sharply in cost. The denial arm is cheap and can run early: remove the operator's AI and have them work the problem cold. The compromise arm is heavier, because it needs a controlled adversary to test against: a bounded shadow-process simulator that feeds shaped, subtly wrong outputs at decisive moments. That instrument need not be an autonomous "evil AI"; it has to be controlled, logged, and switchable off, which is most safely built from an aligned model instructed to play the adversary within strict experimental bounds, or staged Wizard-of-Oz. Because it instructs a system to manipulate a person, it carries the same handling as the shaped-silence study: informed consent, a debrief, a clinical stop rule, and a researcher independent of the proposal. The build pays for itself twice: the controlled shadow-process testbed is a defence-research asset in its own right, letting DRDC study the attack and the defence on one bench, and it instantiates the shadow-process specification as something testable rather than speculative. Sequence the denial arm first; the compromise arm follows once the testbed and the full frontier-AI instrumentation are in place.

Earliest start. Stage 5: needs onboarded operators with real practice, the wargame, and a controlled way to deny and corrupt the tool at test; the compromise arm sharpens once the full frontier-AI instrumentation is in place.