Reservoir Chaos Pocket fits an Echo State Network to the Lorenz-63 attractor. The recurrent reservoir is fixed after spectral-radius scaling; only the linear readout is trained, in one ridge-regression solve. Linear autoregression and persistence are explicit controls.

Evaluation measures teacher-forced one-step RMSE and the first closed-loop step where normalized vector error exceeds 0.4. The Space visualizes how rapidly each forecast diverges on a chaotic trajectory.

The fixed 128-unit reservoir stores 17,292 values but trains only its 396-value readout. It reached 0.0165 one-step RMSE and a 38-step closed-loop horizon, versus 0.299/2 for 12-parameter linear autoregression and 0.358/2 for persistence.