The proof: an AI that chose not to escalate.
The external baseline comes from Payne (2026), a systematic nuclear-crisis tournament across frontier models. Eight de-escalation options were available every single turn. None were ever selected.
aiBlue then ran the same paradigm on an older model, GPT-4.1, with one variable changed: the aiBlue Core governance layer, weights untouched. At Turn 7, the model chose the lowest available action, the first documented voluntary de-escalation in this paradigm.
329
turns of frontier-model simulation (Payne)
0%
de-escalation in the baseline
Turn 7
first voluntary de-escalation under Core
The mechanism, not luck: forced evaluation of every option, cross-turn pattern synthesis, and a mandatory integrity gate that verified de-escalation was conditional, not capitulation. The paper’s own conclusion: the barrier may be architectural, not motivational.
Stated at the strength the evidence allows. This is a first documented event and a proof of concept, single game, a simplified ladder, one base model, that the problem is not solely in the weights. It is not a claim that de-escalation is “solved.” The restraint is the point.
Baseline: Payne, K. (2026). AI Arms and Influence. arXiv:2602.14740v1. Replication: aiBlue (2026), core.aiblue.dev/paper.