Observable-reduced neural world model MVP

observable_reduced_experiment.py implements the linear special case of the proposed observable reduction.

For the aggregate observation y=a+q and g(a)=r*a, the exact identity is

y'' = (r-alpha-beta)y' + beta*r*y.

The script:

1. Integrates the hidden two-compartment system exactly with matrix exponentials.
2. Fits the reduced second-order coefficients from finite-difference derivatives.
3. Sweeps alpha,beta to test c_y=beta*r and c_v=r-alpha-beta.
4. Sweeps sample interval relative to tau=1/(alpha+beta).
5. Compares long-horizon RK4 rollout against a generic first-order model y'=k*y.

Run:

/home/maxwelhelp/main/bin/python3 observable_reduced_experiment.py

Outputs are written to results.json and the console. This MVP does not train a neural residual network or estimate all three rates from coefficients alone: two observable coefficients cannot identify three rates without additional information. It tests the exact structural reduction and its numerical forecasting consequence.