# Simulink Changes ## Change 6 (2026-07-28): online TSE matches specimen identification The `Online TSE Estimator` now calls `control_and_HS/a_specimen_id/online_tse_estimator_runtime.m`. The runtime uses the official 30 Hz filtered, noise-whitened TLS estimator with 0.1 s windows and non-overlapping KF assimilation. It no longer differentiates raw displacement, routes estimates to LSE, clamps the parameters, or initializes stiffness at 7.028 MN/m. MATLAB Coder library generation passes. Saved-record replay on `20260727_1647_testing_OSI.mat` matches the offline `windowed_noise_whitened_tls.m` result at all 8,284 candidate endpoints, with maximum relative error `3.30e-10`. At 177.975 s, the corrected continuous online stiffness is 5.915 MN/m, compared with 5.941 MN/m from the selected specimen-ID window. A 20 s Speedgoat timing verification activated BLWN at 3.064 s. The run completed in the target's IDLE state with no reported error or overload. TETMin/TETAvg/TETMax were 47.40/49.31/71.31 us against the 488.28 us base-rate budget, leaving 85.40% worst-case timing margin. The worst tick was at 3.733 s (pipeline phase 221), after the estimator's SVD path became active. The file logger did not return an importable run, so no signal MAT file was saved from this timing-only verification. ## Change 5 (2026-07-27): simplified V6 becomes canonical SOL-OSI The accepted simplified V6 estimator is now the canonical `control_and_HS/OSI/sol_osi.m`. The former production estimator and its scheduled runtime are retained as `sol_osi_v1.m` and `sol_osi_v1_runtime.m`. The active Speedgoat path uses `sol_osi_runtime.m`, configured by `prepare_sol_osi_runtime.m`. The runtime preserves the 32-tick schedule: the two FFTs remain on separate ticks, and each of the 25 per-frequency residual-fit candidates has its own tick. The Simulink MATLAB Function block is named `SOL-OSI + Auth`, calls `sol_osi_runtime`, and uses neutral `osi*` parameter names. The OSI tap feedthrough gain now uses `osi_dfm`. ## Change 4 (2026-07-12): SOL_OSI promoted without an SLX edit `gpt5_6_runtime.m` now implements the canonical `sol_osi.m` healthy reference, recovery hysteresis, residual-memory gain, and high-level fit. `prepare_gpt5_6_runtime.m` packs the added scalar parameters into `gpts`. Legacy runtime and workspace names are retained, so the existing MATLAB Function block and `.slx` interface are unchanged. Desktop/runtime replay agrees to floating-point roundoff. ## Change 3 (2026-07-10): GPT-5.6 replaces FABLE on Speedgoat The `Fable OSI + Auth` MATLAB Function block is now `GPT-5.6 OSI + Auth` and calls `control_and_HS/OSI/gpt5_6_runtime.m`. Its seven compile-time parameters come from `prepare_gpt5_6_runtime.m`. The existing scalar signal inputs, nine diagnostic outputs, authority-gain connection, and File Log width are unchanged. The logging mux is named `Mux_gpt5_6_runtime`. GPT-5.6 is time-sliced over 32 base-rate ticks per 0.125 s update. The two 1024-point FFTs run on separate ticks, and each of the 25 quiet-band fit candidates has its own tick. A two-window circular buffer preserves the exact update-boundary snapshot while new 2048 Hz samples arrive. Offline comparison against `gpt_osi` + `gpt_auth` was exact for `b_osi`, `b_st`, and authority; `b_res` agreed to floating-point roundoff. The OSI tap feedthrough gain now uses `gpt_dfm`, calculated by the new setup from the numerical-substructure Newmark effective stiffness. The three changed Simulink blocks (`GPT-5.6 OSI + Auth`, `Mux_gpt5_6_runtime`, and `OSI tap D-term gain`) are yellow. Target verification ran for 125 s with GPT-5.6 enabled. The 2048 Hz budget was 488.28 us; measured TETMax was 78.50 us and TETAvg was 53.79 us. This is 83.9% worst-case timing margin, with no overload or early stop. ## Change 2 (2026-07-03): Speedgoat overloading fix Reproduced and fixed the CPU overload that stopped the target when the pause switch was flipped to resume (`Overload limit (0) exceeded in 0.000488281s (tid=0)`, TETMax 780 us against the 488 us base-rate budget). The overload fired ~5 s after resume, when the 5 s slow-spectral ring first fills and the whole OSI slow pipeline (2048-pt FFTs, ring medians, prediction grid) landed in a single base-rate tick. Model + runtime changes: 1. **Time-sliced OSI runtime** (`control_and_HS/OSI/fable_osi_runtime.m`). Same arithmetic (bitwise-equal per `test_fable_runtime_equivalence` on the synthetic stream and all three saved records); the per-step work is now spread over one bounded phase per tick and published at a fixed 56-tick (27 ms) latency after each 0.125 s step. The heaviest rare event, the calibration-finalize pooled-MAD median (nf x cal_windows sort), is split into 16 chunk sorts + 4 merge passes. Measured on target over the full 145 s verification protocol: TETMax 278 us (57% of budget; 275-278 across three runs), worst tick = the single 2048-pt slow FFT; the calibration finalize no longer peaks (was 780 us overload). 2. **OSI enable wiring**: the `Fable OSI + Auth` block's `en` input now comes straight from `osi_gain_enable_switch value`. The old `OSI Runtime Mode` Sum (deleted, along with `OSI Pause Code Gain` and `dd_pause_enable1`) computed `en = 2*dd_pause_enable - osi_gain_enable` (the enable was subtracted - a sign bug), and pausing dd also paused the indicator. **The indicator now always runs**; the pause switch only pauses the command signal. 3. **Online TSE Estimator** decimated: `update_samples` 64 -> 256 (8 Hz stiffness updates) and the first update offset to tick 128 of the 256-sample OSI step, so the 4x4 SVD + 3x3 pinv never share a tick with an OSI pipeline phase. Estimates converge ~4x slower; accuracy per update unchanged. 4. **`OSI Margins Scope`** added next to the fable block, showing `b_st` (slow), `b_fast` (fast) and `b_osi` (selected). These are also logged in File-Log columns 13, 15, 12 as before. Regression tooling: `OSI_tests/overloadin_test/run_overloading_test.m` predicts target fit from a host benchmark (calibrated against the measured 780 us overload; pre-fix host measurement 793 us, k = 0.91) - run it before any target run after changing `fable_osi_runtime.m`. Comparison report: `OSI_tests/overloadin_test/report.html`. ## Change 1: OSI Authority Immediate-Drop Threshold The OSI authority gain law now sets the severe/immediate-drop threshold as: ```matlab down_immediate_margin = 0.25 * target_outer_ncf; ``` For the current target: ```matlab target_outer_ncf = 0.3; ``` this gives: ```matlab down_immediate_margin = 0.075; ``` This means normal below-target NCF values, such as `0.28`, are no longer treated as severe. They must persist for the configured confirmation count before the authority factor drops. Only very low NCF values trigger the immediate-drop path. Clarification on where the authority factor is applied: If the authority factor `a` is constant, multiplying the PID output by `a` is effectively the same linear loop as using a PID with gains `a*P`, `a*I`, and `a*D`. They become different mainly when `a` changes while the test is running. Multiplying the PID output immediately scales everything the PID has already built up, including the integrator contribution. Scaling the PID gains changes how strongly the PID reacts from that point onward. So with adaptive authority, they can give different transients even if the steady linear effect is similar.