shifted-onset disturbances localized, with no pre-onset actions
Adaptive Early-warning Governor for Internal Stability
Read the transition. Govern the response.
AEGIS is a training-state governance system built around three separable functions: observe model-internal change, decide whether the evidence warrants action, and route a qualifying event into a bounded response. The current response policy demonstrates utility and controllability; it is not claimed to be optimal.
negative-control runs authorized an action
5 / 5 Qwen, 5 / 5 Llama, and 3 / 3 Qwen3-4B pairs benefited at their registered endpoints
What the evidence now supports
Observe. Authorize. Respond.
The program tests these functions separately so a signal is not mistaken for permission and a successful response is not mistaken for an optimal one. AEGIS remains operator-controlled at every layer.
Sense a state transition
Prospective shifted-onset studies show that the internal monitoring layer can localize a qualifying transition while the run is active.
Keep evidence and permission separate
Negative controls and transient watch-and-clear studies test whether unusual telemetry should lead to action. Observation alone is not authority.
Demonstrate bounded utility
Matched studies show that one predeclared response can reduce downstream damage. Other contracts—alert, checkpoint, quarantine, halt, or staged response—remain operator choices.
Evidence boundary.
The confirmed end-to-end result covers one abrupt high-loss disturbance class across Qwen and Llama from approximately 1B through 4B parameters. A registered timing audit did not establish optimal intervention timing, an observational forecasting audit did not support promotion as an efficiency controller, and distributed efficacy remains incomplete. These are limits of the claim, not footnotes.
External validation
Test the ontology in the workload that matters.
The design-partner path begins in shadow mode: measure transition behavior, watch-and-clear behavior, monitoring cost, and qualifying-incident exposure. Only then should the training team authorize the response contract that fits its stack.