Fable OSI Parameter Tuning Map

This page maps the current Fable OSI implementation to its Speedgoat runtime parameters. It uses the same notation as the code: dd, dm, b_osi, b_st, b_res, b_fast, auth, fbls, and fbla.

Current setupmain_ColumnSetup.m calls prepare_fable_runtime

Runtime blockfable_osi_runtime, scalar sample I/O, fixed-size arrays

Base ratefs = 2048 Hz, dt = 1/fs

Output folderresearch/osi_parameter_tuning

Source Map

File	Role	Important notation
`main_ColumnSetup.m`	Top-level RTHS setup. Sets `hs.exp_tf_file`, `hs.osi_gain_initial`, `hs.osi_gain_min`, `hs.target_outer_ncf`, then calls `prepare_fable_runtime(projectFolder, dt, hs)`.	`hs`, `dt`, `fs`, `fable_setup`, logged columns `auth`, `b_osi`, `b_st`, `b_res`, `b_fast`, `trip_reason`, `valid`, `decision`.
`control_and_HS/OSI/prepare_fable_runtime.m`	Loads the identified `H_dc_dm`, builds PID `C`, creates the Fable OSI config and authority params, then flattens them for Simulink.	`fbls`, `fblw`, `fblbins`, `fblfw`, `fblfbins`, `fblag`, `fbla`.
`control_and_HS/OSI/fable_osi_runtime.m`	Speedgoat-compatible runtime. Called once per base-rate sample. It time-slices one OSI update across 56 ticks and publishes held outputs.	`dd`, `dm`, `en`, `auth`, `b_osi`, `b_st`, `b_res`, `b_fast`.
`OSI_tests/fable/fable_osi.m`	Desktop estimator and tuning guide. This is the readable source for the slow ST branch, residual branch, fast quick-dip guards, calibration, and prediction curve.	`cfg`, `state`, `info`, `pred_margin`, `alpha_grid`.
`OSI_tests/fable/fable_auth.m`	Desktop authority law and tuning guide. The runtime function `fableAuthStep` is a flattened version of this logic.	`prm`, `ast`, `a_next`, `decision`.
`simulink_changes.md`	Documents the current Speedgoat overload fix: time-sliced Fable OSI, direct enable wiring, and logged Fable margins.	`P_PUB = 56`, 27 ms publish latency, measured TETMax 278 us.

Legacy branch-band fields: hs.osi_branch_split_hz, hs.osi_st_branch_hz, and hs.osi_residual_branch_hz are legacy M9.1 fields. The current Fable path does not consume them in prepare_fable_runtime. Tune Fable branch behavior through hs.fable_osi_override, especially residual_band_hz, not those legacy fields.

Arguments And Runtime Signals

`prepare_fable_runtime(projectFolder, dt, hs)`

Argument or field	Definition	Behavior controlled
`projectFolder`	Path to `main`. Used to load saved `exp_tf` files through project helpers.	File resolution only.
`dt`	Base sample period in seconds. Current code uses `dt = 1/2048`.	All window sizes and runtime step sizes.
`hs.exp_tf_file`	Saved transfer-function file containing `H_dc_dm` and usually `pid_values`.	Nominal loop `L0 = C*H_dc_dm` used for the nominal margin curve and residual anchor.
`hs.target_outer_ncf`	Top-level target margin. In Fable it becomes `cfg.threshold` and `prm.target`.	Main safety/performance setpoint. Higher is more conservative.
`hs.osi_gain_min`	Authority floor. Becomes `prm.a_min`.	Deepest authority shed allowed.
`hs.osi_gain_initial`	Startup and shadow-mode authority. Becomes `prm.a_init`.	Initial product gain at `Apply OSI gain`.
`hs.fable_osi_override`	Struct whose fields override `cfg` from `fable_osi('config', ...)`.	Estimator branch tuning: ST, residual, fast quick-dip, calibration, prediction grid.
`hs.fable_auth_override`	Struct whose fields override `prm` from `fable_auth('params', ...)`.	Authority behavior: shed, recovery, trips, latch, commit gate.
`hs.fable_pred_enable`	Optional boolean flattened into `fbls(55)`. Default is 1.	Set 0 to skip the prediction grid and reduce TET load.

`fable_osi_runtime(dd, dm, en, fbls, fblw, fblbins, fblfw, fblfbins, fblag, fbla)`

Argument	Definition
`dd`	Numerical command displacement sample in mm. This is the loop command used as the OSI input.
`dm`	Measured actuator displacement sample in mm. This is the OSI output.
`en`	Enable code. `0` means shadow mode: estimate and log, but force `auth = a_init`. `1` publishes the authority law. `en >= 2` triggers the pause contract inside the runtime, though current model wiring sends enable directly and does not pause the indicator with `dd_pause_enable`.
`fbls`	Scalar OSI config vector. Index map is in `prepare_fable_runtime` and unpacked in `unpackConfig`.
`fblw`	Slow Hann window, length `segment_n`.
`fblbins`	Slow-bin table: `[freq_hz, fft_bin, keep, res_keep, nominal_curve]`.
`fblfw`	Fast Hann window, length `fast_segment_n`.
`fblfbins`	Fast-bin table: `[freq_hz, fft_bin, keep, cell_lo, cell_hi, band_id]`.
`fblag`	Authority-prediction relative gain grid `alpha_grid`.
`fbla`	Authority-law parameter vector. Index map is in `prepare_fable_runtime` and consumed by `fableAuthStep`.

Runtime outputs

Output	Definition	Log meaning
`auth`	Authority factor applied at the `Apply OSI gain` Product.	Column `auth`. Lower means less loop authority.
`b_osi`	Selected Fable margin: slow selected margin, and during a fast ST trip it is allowed to show `min(b_osi, b_fast)`.	Main OSI value to compare with `target_outer_ncf`.
`b_st`	Slow coherent ST branch margin.	Evidence from bins driven by `dd` with enough coherence and input power.
`b_res`	Slow residual/quiet-band branch margin.	Evidence from output power growth in undriven hazard bands.
`b_fast`	Fast short-memory ST margin.	Quick-dip / sudden collapse indicator.
`trip_reason`	`0` none, `1` quiet-band burst trip, `2` fast FRF-collapse trip.	Why immediate authority shed was triggered.
`valid`	`1` when `b_osi` is real post-warm-up evidence, `0` otherwise.	Do not treat NaN/warm-up as a real safety margin.
`decision`	`0` hold, `1` down, `2` severe down, `3` up, `4` fast trip.	Authority-law state machine output.

Core Equations

fs = 1/dt
segment_n = round(segment_s*fs)
step_n = round(step_s*fs)
n_segment = round((window_s - segment_s)/hop_s) + 1

The OSI estimates the closed-loop transfer from command displacement to measured displacement. In the code, M = Suy/Suu, where Suu is command autospectrum, Syy is measured autospectrum, and Suy is cross spectrum.

g2raw = abs(Suy)^2 / max(Suu*Syy, tiny^2)
gamma2 = max(0, (k_eff*g2raw - 1)/(k_eff - 1))
excited = keep & g2raw > gamma_gate & has_input & not_leak
quiet = res_keep & ~has_input & g2raw < gamma_quiet

The project margin is the normalized coprime-factor style chordal margin. With T = M and S = 1 - T:

b(w) = 1/(sqrt(2)*sqrt(abs(S)^2 + abs(T)^2))
b_st = min_over_excited_bins( masked_median_frequency(b(w)) )
b_osi = min(b_st, b_res), capped by cap = 1/sqrt(2)

The residual branch watches output power that is not explained by the coherent command path.

res = Syy * max(1 - gamma2, residual_floor)
ratio = res / baseline
normq = ratio / max(median(ratio(quiet)), eps)
z = (log(max(normq, eps)) - mu0) / sigma0
growth = exp(max(z_max, 0)*sigma0)
b_res = max(0.05, anchor/sqrt(max(growth, 1)))

The fast quick-dip logic uses CUSUM statistics. Lower drift nu or threshold h makes detection faster and more sensitive.

cus_st = max(0, cus_st + z_d - fast_st_nu)
st_trip = cus_st > fast_st_h & b_fast < b_slow_filt & b_fast < threshold + fast_babs_offset
cus_burst(j) = max(0, cus_burst(j) + z_burst(j) - fast_burst_nu)
burst_trip = any(cus_burst > fast_burst_h)

The authority law regulates b_osi toward prm.target. It lowers authority quickly when margin is low and recovers only after hysteresis and hold-time checks pass.

filt_alpha = 1 - exp(-dt_update/b_filt_tau_s)
b_filt = (1 - filt_alpha)*b_filt + filt_alpha*b
severe = b < sev_frac*target
low = min(b, b_filt) < target - down_deadband
a_pl = a_cont * (max(b_size, 0.02)/target)^down_power
a_cont = min(a_cont, max(a_shed_tgt, a_cont - rate*dt_update))
recovery_rate = up_rate*(1 + up_prop*surplus/target)

Slow ST Branch Parameters

The slow ST branch is the accurate margin level when the command excites the dangerous frequencies. It is the best branch for recovery decisions because it observes the actual coherent transfer T = dm/dd.

Parameter	Default	Definition	Tuning effect
`band_hz`	`[2 100]`	Overall frequency band considered by the OSI.	Narrowing reduces cost and ignores irrelevant modes; widening watches more modes but can add noisy bins.
`line_reject_hz`	`[58 62]`	Power-line rejection band removed from slow and fast grids.	Move only if electrical line contamination is elsewhere.
`segment_s`	`1.0`	Hann segment length for slow spectra.	Keep near 1 s for 1 Hz resolution. Shorter weakens margin frequency resolution.
`hop_s`	`0.25`	Slow spectral refresh cadence. Must be compatible with `step_s`.	Larger is cheaper and smoother but slower; smaller is faster but heavier.
`step_s`	`0.125`	Runtime OSI step cadence and fast path cadence.	Larger lowers Speedgoat work; smaller improves quick-dip reaction but costs TET.
`window_s`	`5.0`	Slow ring memory. Produces `n_segment` median-aggregated segments.	Longer gives stable margins; shorter follows changes faster but can false dip.
`gamma_alpha`	`0.01`	False-pass probability used to derive `gamma_gate`.	Lower is stricter, accepting fewer coherent bins.
`input_power_range_db`	`35`	Bins this far below the in-band input peak are considered unexcited.	Lower accepts only stronger command bins; higher accepts weaker bins.
`sidelobe_db`	`25`	Rejects bins that look like Hann leakage from a nearby stronger bin.	Raise to be stricter about leakage; lower to keep more adjacent bins.
`min_bins`	`5`	Minimum excited slow bins required for `b_st`.	Raise to reduce false ST decisions; lower to allow sparse excitation.
`freq_medfilt_n`	`3`	Median filter width across frequency bins.	Raise to ignore single-bin dips; keep odd because runtime median assumes odd centered windows.
`margin_filter_n`	`3`	Temporal median over distinct slow updates before publishing selected slow margin.	Raise for smoother `b_osi`; adds about `(n-1)*hop_s/2` lag.
`slow_filt_alpha`	`0.2`	EMA weight for `b_slow_filt`, used by fast path direction gate.	Lower makes fast-trip reference steadier; higher follows level changes faster.
`warm_updates`	`4`	Number of slow updates before margins are reported.	Raise to avoid startup artifacts; lower for earlier evidence.
`cap`	`1/sqrt(2)`	Mathematical maximum of the chordal margin.	Not a normal tuning knob.
`threshold`	`target_outer_ncf`	Detection threshold and fast-trip absolute level.	Higher is more conservative; must be reachable by the healthy full-authority loop.
`alpha_grid`	`logspace(log10(0.05), log10(20), 25)`	Relative loop-gain grid for predicting safe authority changes.	More points improve recovery/shed boundary resolution but cost runtime work.
`pred_denom_floor`	`1e-3`	Floor on `1 - M` when converting `T` to loop gain for prediction.	Prevents numerical blow-up near `T = 1`.

Residual Branch Parameters

The residual branch watches the hazard band when the command does not excite it. It is slower than the fast branch because it needs a calibrated quiet-band baseline, but it is the branch that protects against uncommanded growth.

Parameter	Default	Definition	Tuning effect
`residual_band_hz`	`[28 100]`	Hazard band where quiet-bin growth is watched.	Most important residual branch-band knob. Narrow to avoid false alarms from command harmonics; widen to watch more modes.
`gamma_quiet`	`0.15`	Maximum raw coherence for a bin to count as quiet.	Lower is stricter quiet selection; higher allows more bins into residual evidence.
`residual_floor`	`0.03`	Minimum fraction in `res = Syy*max(1 - gamma2, residual_floor)`.	Prevents residual power from going to zero under high coherence.
`cal_skip_s`	`10`	Startup blackout before accepting residual calibration windows.	Raise if startup transients poison baseline.
`cal_windows`	`80`	Accepted slow-hop windows needed for full calibration.	More gives stronger baseline; fewer calibrates faster.
`cal_min_windows`	`24`	Minimum accepted windows if timeout occurs.	Lower calibrates under difficult data but weakens statistics.
`cal_timeout_hops`	`600`	Slow-hop attempts before timeout path can accept `cal_min_windows`.	Lower allows earlier readiness after many rejects.
`cal_reject_factor`	`4`	Rejects calibration windows whose broadband residual median is more than this factor from reference.	Lower rejects more startup/outlier windows; higher is more permissive.
`cal_restart_rejects`	`12`	Consecutive rejects that restart calibration collection.	Lower recovers faster from a bad initial reference.
`z_on`	`4.0`	Z-score threshold to engage residual branch.	Lower detects earlier but risks false engagement; raise for fewer false alarms.
`z_off`	`2.0`	Z-score threshold below which engagement decays.	Controls hysteresis. Lower holds residual engagement longer.
`res_on_n`	`4`	Consecutive slow hops above `z_on` needed to engage.	Lower reacts faster; higher needs sustained evidence.
`res_hold_s`	`5`	Time engagement may persist in the hysteresis zone without reconfirmation.	Raise to hold caution longer; lower to recover faster from one-off events.
`res_min_growth_db`	`3`	Absolute growth required in addition to `z_on`.	Raise to ignore small quiet-band changes.
`res_match_hz`	`3`	Frequency persistence tolerance for the residual peak.	Lower requires same-bin persistence; higher tolerates drifting mode frequency.
`sigma0_floor`	`0.15`	Floor on calibrated log-ratio spread.	Raise to desensitize residual z-scores globally.
`bl_adapt_tau_s`	`100`	Time constant for healthy quiet-band baseline drift.	Lower tracks benign drift faster but risks absorbing real degradation.
`bl_adapt_max_db`	`3`	Hard baseline adaptation cap around original baseline.	Raise only if benign drift exceeds 3 dB; too high can hide damage.
`regime_hops`	`8`	Sustained input-support changes before residual rebaseline.	Lower rebaselines earlier on test phase changes; higher avoids unnecessary recalibration.

Fast Quick-Dip Parameters

The fast branch is not the main recovery authority. It is the quick protection path for sudden margin dips, FRF collapse, and quiet-band bursts. Its output can force decision = 4 and immediate authority shed.

Parameter	Default	Definition	Tuning effect
`fast_segment_s`	`0.25`	Fast Hann segment length. At 2048 Hz this is 512 samples.	Shorter detects sooner but gets coarse/noisy spectra; longer is calmer and slower.
`fast_ring_n`	`3`	Number of fast segments in the fast median ring.	Raise for smoother `b_fast`; lower for faster quick-dip sensitivity.
`fast_min_bins`	`3`	Minimum fast bins needed for fast ST margin.	Raise to avoid trips from sparse evidence.
`fast_st_nu`	`1.25`	CUSUM drift for FRF-change statistic.	Lower accumulates evidence faster; higher reduces false trips.
`fast_st_h`	`4.5`	CUSUM trip threshold for FRF-change statistic.	Main fast ST sensitivity knob. Lower means earlier trips.
`fast_babs_offset`	`0.10`	Fast ST trip additionally requires `b_fast < threshold + fast_babs_offset`.	Raise to trip while still above target; lower to trip only near/under target.
`fast_amove_frac`	`0.02`	Relative authority move that disarms the FRF-change statistic for one slow window.	Prevents the law from tripping on its own authority changes. Lower is more conservative against self-trip.
`fast_pin_lo`	`0.25`	Newest command power must be at least this fraction of its EMA.	Raise to reject fading-command artifacts; lower keeps fast path armed through fades.
`fast_pin_hi`	`4.0`	Newest command power must be below this multiple of its EMA.	Lower rejects command bursts more aggressively.
`fast_pin_alpha`	`0.04`	EMA update weight for command-power pinning.	Higher follows changing input power faster; lower makes the pin reference steadier.
`fast_burst_bands`	`[28 44; 44 60; 62 78; 78 94; 94 100]`	Quiet-band burst sub-bands on the fast grid.	Remove bands to reduce cost/false trips; adjust to match actual instability modes.
`fast_burst_nu`	`1.0`	CUSUM drift for quiet-band burst z-scores.	Lower detects bursts faster; higher calms the detector.
`fast_burst_h`	`6.0`	CUSUM trip threshold for quiet-band bursts.	Lower trips sooner; higher requires stronger sustained burst evidence.
`fast_sig_floor`	`0.20`	Floor on calibrated log burst-power spread.	Raise to desensitize all burst bands.
`fast_d_sig_floor`	`0.30`	Floor on calibrated log FRF-change spread.	Raise to desensitize the fast ST CUSUM.

Authority-Law Parameters

These parameters are flattened into fbla. They control how the measured OSI behavior is converted into the gain factor auth that multiplies dc.

Parameter	Default	Definition	Tuning effect
`target`	`0.55`, overridden by `target_outer_ncf`	Margin level the authority law regulates onto.	Primary safety/performance knob. Higher sheds earlier and recovers later.
`a_min`	`0.08`	Authority floor.	Raise to preserve performance; lower to allow deeper emergency detuning.
`a_max`	`1.00`	Authority ceiling.	Usually keep 1.0.
`a_init`	`1.00`	Startup and shadow-mode authority.	Set lower only for intentionally detuned startup tests.
`dt_update`	`cfg.step_s`	Authority update cadence.	Must match OSI runtime step cadence.
`b_filt_tau_s`	`0.4`	Time constant for margin EMA.	Raise for smoother authority; lower for faster response.
`down_deadband`	`0.01`	Deficit begins below `target - down_deadband`.	Widen to ignore tiny target grazes; narrow for tighter regulation.
`down_confirm_n`	`2`	Consecutive low updates before non-severe shed.	Lower sheds sooner; higher rejects noise.
`down_power`	`1.5`	Exponent in `a*(b/target)^down_power`.	Higher makes each shed deeper.
`down_rate`	`2.0`	Max confirmed shed rate in authority units per second.	Raise to catch margin drops faster; lower for smoother authority.
`down_settle_s`	`1.25`	Hold time before resizing a shed from similar evidence.	Lower re-sheds faster; higher avoids compounding a stale margin reading.
`re_shed_delta`	`0.02`	Fresh lower margin drop that bypasses `down_settle_s`.	Lower makes continued decline resize sooner.
`sev_frac`	`0.60`	Raw margin below `sev_frac*target` is severe.	Raise to treat more dips as emergencies; lower reserves severe path for deeper collapses.
`sev_down_rate`	`8.0`	Emergency shed speed.	Keep high for crash-catching; lower only to reduce abrupt commands.
`up_deadband`	`0.02`	Recovery starts only above `target + up_deadband`.	Raise to stop target-hugging oscillation; lower for faster recovery.
`up_hold_s`	`1.5`	Healthy dwell time before recovery.	Raise for caution; lower to reclaim authority faster.
`up_rate`	`0.08`	Recovery rate when coherent ST evidence is valid.	Raise for faster recovery under measured margin evidence.
`up_rate_blind`	`0.02`	Recovery rate without coherent input.	Keep lower than `up_rate`; this is probe-and-verify recovery.
`up_prop`	`3.0`	Recovery boost from surplus margin.	Raise to sprint when margin is comfortably high.
`min_step`	`0.02`	Commit gate from internal `a_cont` to applied `auth`.	Raise for fewer, larger authority moves; lower for smoother continuous authority.
`use_pred`	`1`	Uses `pred_margin` to bound sheds and cap recovery.	Set 0 when prediction grid is disabled for runtime budget.
`pred_guard`	`0.03`	Predicted margin must exceed `target + pred_guard` to count safe.	Raise if prediction is optimistic; lower to recover more aggressively.
`pred_floor_frac`	`0.90`	Shed floor as fraction of predicted-safe authority.	Lower allows more conservative shed below predicted boundary.
`trip_drop`	`0.60`	Multiplicative drop on fast trips.	Lower is more aggressive. Burst trips may compound while evidence persists.
`trip_power`	`1.5`	Fast ST trip sizing exponent using `b_fast/target`.	Higher makes low `b_fast` produce deeper immediate drops.
`trip_cool_s`	`5.0`	Cooldown after trip during which recovery is blocked and trip re-arms.	Raise for caution after quick-dip events.
`trip_undo_s`	`2.5`	If regime recalibration follows a fast trip within this time, undo the trip.	Raise to tolerate phase-change false trips; lower to keep trips latched.
`self_move_gate_s`	`1.5`	Authority must be steady this long before fast ST trip is honored.	Prevents self-trip. Raise if authority motion still triggers fast ST.
`self_move_frac`	`0.02`	Relative committed authority move that resets self-move age.	Lower is stricter about considering authority steady.
`latch_relax`	`1.15`	Post-event recovery ceiling relative to post-shed authority.	Lower keeps recovery more latched after hard events.
`latch_decay_tau_s`	`1`	Time constant for forgetting the latched ceiling.	Raise for slower post-event recovery; lower for faster recovery after healing.

Speedgoat Runtime Parameters

These are mostly scheduling and flattening parameters. They control whether the algorithm fits inside the 2048 Hz real-time budget.

Runtime item	Definition	Tuning or failure mode
`P_PUB = 56`	Publish phase in `fable_osi_runtime`. One 0.125 s step is time-sliced and outputs publish 56 base-rate ticks after the step boundary.	At 2048 Hz, latency is about 27 ms. This is part of the overload fix.
`pred_enable` / `hs.fable_pred_enable`	Controls whether the prediction grid runs.	First TET reduction knob. Set 0 to skip prediction; authority then uses power-law regulation alone.
`step_s`	Controls OSI step cadence and fast path cadence.	If target overruns, raising to 0.25 halves fast-path rate.
`hop_s`	Controls slow path refresh cadence.	If target overruns, raising to 0.5 halves slow-pipeline rate.
`band_hz`	Frequency span stored in `fblbins` and `fblfbins`.	Narrow to reduce bin count and median/prediction work.
`segment_s`	Slow FFT length.	Do not shrink below 1 s unless accepting worse frequency resolution.
`freq_medfilt_n`	Runtime assumes odd centered medians.	Keep odd; `prepare_fable_runtime` asserts this.
`en`	Runtime enable code.	Current model uses direct enable wiring; the OSI keeps evaluating during command pause.

Speedgoat overload response order from the runtime header: 1. fable_pred_enable = 0 2. step_s = 0.25 3. hop_s = 0.5 4. band_hz = [2 80]

Final Categorized Tuning Table

This table focuses only on parameters that are practical to tune to control OSI behavior. Categories are branch-based plus quick-dip and authority behavior.

Category	Main parameters	Primary behavior controlled	Faster / more protective	Smoother / fewer false actions
Slow ST branch	`window_s`, `hop_s`, `margin_filter_n`, `freq_medfilt_n`, `gamma_alpha`, `input_power_range_db`, `min_bins`	Accurate coherent margin `b_st` used for normal regulation and recovery.	Lower `window_s`, lower `hop_s`, lower `margin_filter_n`, higher accepted-bin permissiveness.	Raise `window_s`, `margin_filter_n`, `freq_medfilt_n`, and `min_bins`; lower `gamma_alpha`.
Residual branch	`residual_band_hz`, `z_on`, `z_off`, `res_on_n`, `res_min_growth_db`, `sigma0_floor`, `regime_hops`	Quiet-band hazard detection `b_res` when command excitation does not directly observe the dangerous band.	Lower `z_on`, lower `res_on_n`, lower `res_min_growth_db`, widen hazard band only around real modes.	Raise `z_on`, `res_on_n`, `sigma0_floor`; narrow `residual_band_hz` away from command harmonics.
Fast quick-dip	`fast_st_nu`, `fast_st_h`, `fast_babs_offset`, `fast_burst_nu`, `fast_burst_h`, `fast_ring_n`, `fast_burst_bands`	Immediate trips for sudden FRF collapse or quiet-band burst: `trip_reason` 1 or 2, `decision = 4`.	Lower `fast_st_h`, lower `fast_burst_h`, lower `nu` values, raise `fast_babs_offset`.	Raise CUSUM thresholds, raise `fast_ring_n`, narrow `fast_burst_bands`, raise spread floors.
Authority drop	`target`, `down_deadband`, `down_confirm_n`, `down_power`, `down_rate`, `down_settle_s`, `sev_frac`, `sev_down_rate`, `trip_drop`, `trip_power`	How `b_osi` and fast trips reduce `auth`.	Raise `target`, raise `down_rate`, raise `down_power`, raise `sev_frac`, lower `trip_drop`.	Lower `target`, raise `down_confirm_n`, widen `down_deadband`, raise `down_settle_s`.
Authority recovery	`up_deadband`, `up_hold_s`, `up_rate`, `up_rate_blind`, `up_prop`, `min_step`, `pred_guard`, `latch_relax`, `latch_decay_tau_s`	How fast `auth` returns after margin is healthy.	Lower `up_hold_s`, lower `up_deadband`, raise `up_rate`, raise `up_prop`, lower `pred_guard`.	Raise `up_hold_s`, raise `up_deadband`, lower recovery rates, raise `pred_guard`, lower `latch_relax`, raise `latch_decay_tau_s`.
Speedgoat fit	`fable_pred_enable`, `step_s`, `hop_s`, `band_hz`, `alpha_grid`, `fast_burst_bands`	CPU/TET budget and runtime latency.	Keep defaults for highest fidelity and fastest detection if TET permits.	Disable prediction first, then raise `step_s` or `hop_s`, narrow `band_hz`, reduce burst bands.