The LTE pipeline here uses a mixed presentation objective: (1) 31 tidal constituents are summed as cos-phase waves with the selected run’s YEAR_LEN; (2) the Ada-style drifting impulse comb from Amplify(...), using delB, bg, and year, modulates the tidal sum; (3) the bidirectional IIR integrator is seeded at IDATE = 1920.9; (4) an affine calibration aligns the reconstructed manifold to column 4; (5) MLR with regressors [1, m, sin/cos(2πk⋅m)] is fit on the years outside the shaded interval and then compared against the observed target series inside the hold-out window and over the full record. The spectrum panel is a direct check of lte_results.csv columns 5–7, so its frequency axis is in cycles per manifold-level shift rather than cycles per year.
The selected Feb2026 run supplies the displayed record. Across runs, the subdirectory and corresponding observed data file change together (for example /11 and 11.dat), while lte_results.csv, lt.exe.p, and lt.exe.resp keep the same filenames inside each run directory. Within lte_results.csv, column 1 = date, column 2 = source-model output, column 3 = observed target series from that run’s .dat file, column 4 = internal forcing manifold level, column 5 = manifold-space frequency, column 6 = model spectrum, column 7 = data spectrum. Presentation of the final fit uses the observed series from column 3, while the latent manifold calibration still targets column 4. The shaded interval is a validation / test hold-out window; the surrounding years supply the fitted MLR coefficients.