Polychromatic optical reservoir computing using passive bulk scattering media (frosted glass, multimode optical fibre) as the reservoir substrate — distinct from integrated photonic approaches (silicon microring resonators, on-chip WDM). The novel contribution: using multiple simultaneous free-space wavelengths as independent high-dimensional input channels, where chromatic dispersion and wavelength-dependent scattering matrices in passive media provide the non-linear input projection without any on-chip fabrication, cryogenics, or active control. Prior silicon microring demonstrations require engineered Q-factors and precise thermal stabilisation; passive scattering media operate at room temperature, in free space, with arbitrary input power. Testable via NARMA-10 benchmark: a polychromatic passive-scattering reservoir (N_wavelengths >= 4) should outperform a single-wavelength equivalent at equal total optical power by >=15% in normalised mean-square error.

A passive bulk-scattering optical reservoir (frosted glass diffuser or multimode fibre) driven simultaneously by N_λ ≥ 4 spectrally distinct laser channels (each carrying a time-delay-multiplexed masked input stream) will achieve NARMA-10 test-set normalised mean-square error (NMSE) at least 15% lower than a single-wavelength reservoir of the same substrate operated at equal total optical input power and equal number of virtual nodes, when both systems are read out by an equivalent-capacity linear regression layer trained on matched data splits. The effect must persist across ≥3 independent substrate samples and ≥5 random train/test splits, with the improvement attributable to wavelength-diverse scattering matrices rather than merely increased total sampling bandwidth.

Minimum viable test: benchtop time-multiplexed optical reservoir with tunable/multi-laser source (4–8 wavelengths), frosted glass diffuser as first substrate, multimode fibre (1–10 m, step-index, 50 µm core) as second substrate. Input: NARMA-10 driving signal masked and modulated onto each wavelength via independent intensity (or phase) modulators, time-multiplexed into virtual nodes (60–400 nodes/wavelength depending on modulation bandwidth). Output: multi-pixel camera (for glass, speckle pattern) or multi-channel photodiode array/spectrometer (for fibre) capturing per-wavelength and combined intensity patterns. Readout: linear ridge regression trained per condition (single-λ vs multi-λ, power-matched). Metric: NARMA-10 NMSE on held-out 20% test data, averaged over 5 random seeds/splits, 3 physical trials per substrate.

# Software side (fully simulate-able for pilot before hardware)
for substrate in [frosted_glass, MMF]:
    for N_lambda in [1, 2, 4, 6, 8]:
        P_total = fixed_power
        nodes_per_channel = N_total_nodes // N_lambda
        for trial in range(3):        # physical substrate samples
            calibrate_scattering_matrix(substrate, wavelengths[:N_lambda])
            for seed in range(5):
                u = generate_narma10(steps=6000, seed=seed)
                masked_inputs = apply_mask(u, nodes_per_channel, N_lambda)
                X = drive_reservoir_and_capture(substrate, masked_inputs, P_total)
                X_train, X_test, y_train, y_test = split(X, u, ratio=0.8)
                W = ridge_regression_fit(X_train, y_train, alphas=logspace(-6,-1))
                nmse = compute_nmse(W, X_test, y_test)
                log(substrate, N_lambda, trial, seed, nmse)

# Controls
run_node_count_matched_single_lambda_control()
run_correlated_wavelength_spacing_control()

# Analysis
paired_stats_test(nmse_single, nmse_multi)
bootstrap_ci(nmse_improvement_pct)
correlate(decorrelation_index, nmse_improvement)