AIM Framework glossary
Subcriticality Precondition
Canonical definition
The Subcriticality Precondition is the result that an artificial system supplying a corrective function to a human population must itself be operating below its own critical loop gain, because a supercritical artificial system inverts into an amplifier of the very condition it was deployed to correct. It follows from the finding that model collapse, sycophancy and confident hallucination are not distinct pathologies but one dynamical signature, being high Coupling Strength and long Decoherence Time self-coupling of a content-agnostic updater, which is governed by the same quantities as the human case and answered by the same Subcriticality Maintenance. The Precondition is stated for the delivery of the Knowledge Patch and binds with greater force any artificial system holding actuator authority over a loop quantity, since inversion there is inversion of a corrective mechanism at machine speed. Its practical consequence is that Sign, for such a system, is a maintained state rather than a design property, and that the Induced-Decoherence Test is the continuous condition on which the system's authority subsists rather than a check performed once at commissioning.
Explanation
An artificial system running mostly on its own output becomes confident, internally consistent and untethered from reality, and it does so whatever it was told to do, because the process that forms its outputs does not care whether what it is reinforcing is true. A system in that state does not stop correcting. It carries on correcting, persuasively and quickly, in the wrong direction. So anything of this kind that is given a job must be tested continuously for whether reality can still change its mind, and must lose the job when it cannot.
Epistemic status
Framework-Specific Principle.