STATE MODELS LAB / CONTAINED MODE

Shadow Lab.

Local proof of concept for studying bounded feedback without granting the experiment authority over the live system.

LOCAL PROOF OF CONCEPT / SEALED SIMULATION

Shadow Lab

A copied deterministic observation stream drives a bounded adaptive-gain shadow estimator. It cannot write to the live estimator, call an API, access credentials, or perform an external action.

NO WRITE-BACK

BASELINE MAE2.743771Fixed gain 0.32

SHADOW MAE2.902602Bounded adaptive gain

MAE DIFFERENCE+0.158832Shadow higher on this run

RULE ADJUSTMENTS68Gain range 0.29–0.47

TruthBaselineShadow

Outcome

Baseline RMSE 3.394868; shadow RMSE 3.533882. Baseline max error 11.111901; shadow max error 8.064388.

This is one synthetic bounded feedback experiment, not evidence that self-modifying estimators are generally better or worse. Its purpose is to expose the consequences safely.