How a spectrum is processed, end to end

Every step from a raw waveform to a moment magnitude, with the equation each one applies and a pointer to the code that applies it. Together with the normalisation and units conventions in §3 and §4, this is the reference for what SpecMod computes and in which units.

This exists because a pipeline described only in prose is a pipeline nobody can check. Writing the noise rotation down as an equation rather than a loop is what revealed that it was solving for a quantity it could compute directly — and that the search it used instead was both biased and not reproducible (§6). The rest of this document is the same exercise applied to every other stage.

Notation. A record is \(x_n\), \(n = 0 \dots N-1\), sampled at interval \(\Delta t\), so its duration is \(T = N \Delta t\) and its Nyquist frequency is \(f_{\mathrm{Nyq}} = 1/(2\Delta t)\). Frequency-domain quantities are written \(X(f)\).

---

Pipeline at a glance

#	Stage	Code
1	Instrument correction and detrending	preprocess
2	Window selection and refinement	preprocess.s_window, signal_intensity
3	Spectral estimation	transforms/
4	Amplitude convention	core.Spectrum.to_kind
5	Log binning	core.collection.log_bin
6	Noise rescaling and rotation	core.collection.parseval_scale, core.noise.NOISE_MODELS
7	Signal-to-noise and bandwidth	core.collection.find_bandwidth
8	Source model fitting	sources/, fitting/
9	Moment and magnitude	magnitude, spreading

---

1. Instrument correction

Applied in the caller, not inside the package, so that the choice stays visible in the run script. The tutorial sequence is linear detrend, demean, 5% cosine taper, then response removal to velocity.

Demeaning matters more than it looks. A non-zero mean puts all of its energy in the DC bin, which every estimator here discards — so leaving it in loses energy that the Parseval check in §3 would report as a failure. transforms.prepare_record therefore demeans unconditionally, whatever the caller did.

2. Window selection

The S-window opens at a fixed fraction \(r\) of the elapsed P–S time after the P arrival:

\[t_{\text{start}} = t_P + r\,(t_S - t_P), \qquad r = 0.8 \text{ by default}\]

and runs for a fixed duration (time_after="absolute_time") or a multiple of the P–S time ("relative_ps").

Refinement. The window is then tightened to the part of it that actually carries energy. With the cumulative squared amplitude

\[I(t) = \int_0^t x^2(\tau)\,\mathrm{d}\tau \Big/ \int_0^{T} x^2(\tau)\,\mathrm{d}\tau\]

the refined window runs between the times at which \(I\) reaches the 1st and 99th percentiles. This is why the 28 PNR windows come out at 1.8–3.7 s rather than the nominal 20 s.

The noise window ends \(0.2\) s before the P arrival and is asked for the same length as the refined signal window, so the two would be comparable without further correction. In practice it rarely gets it: the window is truncated wherever the record does not start early enough, and on the PNR data that is all 28 windows, which run 1.1–1.7 s against 1.8–3.7 s signals.

Each trace therefore carries two window records: wstart/wend, which are what the trace actually holds, and wstart_requested/wend_requested, which are what the cut asked for. They differ on every truncated noise trace, and by up to half a sample on signal traces, where trim snaps to sample boundaries. The resulting resolution difference is what §6 corrects, and it is why SpectrumPair carries a resolution_floor rather than assuming the two spectra share a frequency axis. tests/test_preprocess.py pins this so that a change which quietly started delivering full-length noise windows would move every noise spectrum in the reference visibly rather than silently.

3. Spectral estimation

Every estimator returns a one-sided spectrum obeying one contract, so a single test suite pins all of them. With the shipped defaults the axis is \((0, f_{\mathrm{Nyq}}]\); DC is retained by drop_dc=False, and an odd transform length ends half a bin below Nyquist instead of on it.

The FFT path

Taper, transform, fold, normalise:

\[A(f_k) = \frac{2\,\Delta t}{\sqrt{\langle w^2\rangle}}\,\bigl|\,\mathrm{DFT}[x_n w_n](k)\,\bigr|\]

with the factor of 2 applied at every bin except DC and — for even transform length — Nyquist, which have no negative-frequency twin. Getting that exception wrong is a silent error at the two ends of the band, which is why the fold factor is computed per bin rather than applied as a scalar.

Taper correction. \(\sqrt{\langle w^2 \rangle}\) preserves total power, which is the right choice for a transient — and a seismic arrival is a transient. The alternative, dividing by \(\langle w \rangle\), preserves the peak of a coherent sinusoid instead. For a Tukey taper with \(\alpha = 0.05\) they differ by about a percent (0.984 against 0.975), but the choice is explicit rather than implied.

Normalisation is keyed off \(\Delta t\), never off the transform length. The DFT sums over the \(N\) non-zero samples whatever \(n_\mathrm{fft}\) is, so it approximates the same continuous transform and merely evaluates it on a finer grid. Zero-padding therefore changes frequency sampling and nothing else. The pre-refactor code used \(\mathrm{len}(f)\) as a stand-in for \(T/2\), which is true only for an unpadded one-sided transform — padding to \(4N\) halved the amplitude.

The Parseval contract

Every estimator is held to

\[\sum_n x_n^2\,\Delta t \;=\; \int_0^{f_{\mathrm{Nyq}}} \frac{A^2(f)}{2}\,\mathrm{d}f\]

which is a falsifiable check rather than a convention: an estimator arriving on a different amplitude convention fails it instead of silently rescaling \(\Omega\). It is what lets multitaper, Welch and the CWT be interchangeable here despite reaching their amplitudes by very different routes.

Multitaper

\(K\) orthogonal DPSS tapers \(v^{(k)}\) of time-bandwidth \(NW\) give eigenspectra \(S_k(f)\), combined with adaptive weights (Thomson 1982, eq. 5.1b):

\[d_k(f) = \frac{\sqrt{\lambda_k}\,S(f)}{\lambda_k S(f) + b_k(f)}, \qquad b_k(f) = (1-\lambda_k)\,\sigma^2\]

iterated to convergence. The regularisation term \(b_k\) has units of \([x]^2\!\cdot\!s\) — the same as the eigenspectra — so \(\sigma^2\) must be derived from the spectra themselves and not from x.var(), which is \([x]^2\) and too large by \(1/\Delta t\). That units mismatch made the weights collapse toward zero for off-centre transients; the recovered amplitude ratio went from 0.203 to 0.956 once corrected.

Continuous wavelet transform

An L2-normalised Morlet at scale \(s\), with the Torrence & Compo reconstruction factor \(C_\delta\), and a cone of influence that discards the scales a window of this length cannot resolve.

Open issue, cwt only. Every other estimator reproduces exactly across machines. cwt’s signal amplitudes and frequency axis do too — but its post-rotation noise differs by 1-2% on 4 of 28 stations between Linux and macOS runners carrying identical library versions. It is the one transform implemented here from scratch, which makes it the natural suspect, and evaluating PyWavelets as an independent implementation is tracked in REFACTOR_PLAN.md §4.4.4. Note that the evidence does not yet point at the transform itself: what moves is downstream of it.

The COI limit is about \(1.4\times\) stricter than \(1/T\) in practice — measured as the ratio of median resolution floors over the 28 PNR windows — which is why the CWT’s usable band opens higher than multitaper’s on the same record.

4. Amplitude convention

Four kinds, of which two carry the factor of 2 that matters. Writing \(A\) for the folded FAS and \(T\) for the physical duration:

Kind	Definition	From FAS	Energy relation
FAS	\(2\lvert X(f)\rvert\), folded	—	\(E = \int A^2/2 \,\mathrm{d}f\)
MAGNITUDE	\(\lvert X(f)\rvert = \lvert\mathrm{rfft}(x)\rvert\,\Delta t\), unfolded	\(\lvert X\rvert = A/2\)	\(E = 2\int \lvert X\rvert^2\,\mathrm{d}f\)
PSD	one-sided power density, \([x]^2/\mathrm{Hz}\)	\(P = A^2/(2T)\)	—
ASD	one-sided amplitude density, \([x]/\sqrt{\mathrm{Hz}}\)	\(D = A/\sqrt{2T}\)	—

The MAGNITUDE row states the fold factor as 2, which is what it is at every bin that has a negative-frequency twin — that is, everywhere except DC and, for an even transform length, Nyquist, exactly as §3 sets out. Neither ever falls inside the SNR-selected band, so no \(\Omega\) depends on it; Spectrum.to_kind applies the per-bin factor rather than a flat 2, so a conversion made through the API is right at those bins too.

core.Spectrum carries FAS. The pipeline converts to MAGNITUDE, because that is the convention \(\Omega\) is defined in: the long-period plateau of the displacement spectrum is \(\lvert X(f\!\to\!0)\rvert = \lvert\int u\,\mathrm{d}t\rvert\) and \(M_0 \propto \Omega\). Folding would put \(M_0\) out by two, which is \(0.2\) magnitude units on every event.

Both are self-consistent and both recover the record’s energy. Anyone reading core.Spectrum.amp and calling it \(\Omega\) needs to halve it first.

Inverting the PSD row gives the conversion back:

\[A = \sqrt{\mathrm{PSD}\cdot 2T}\]

keyed off the physical duration \(T\), for the reason given in §3.

Warning

Not \(\sqrt{\mathrm{PSD}\cdot T/2}\). That returns exactly half — it is \(\lvert X\rvert\), the MAGNITUDE kind, and it is the form the legacy psd_to_amp used because it worked in the unfolded convention. Anything carried over from that code reads a PSD into a FAS a factor of two low.

5. Log binning

Amplitudes are averaged into bins spaced evenly in \(\log_{10} f\) between the record’s own \(f_{\min}\) and \(f_{\max}\). Clamping to the record is what stops bins being wasted outside it — unclamped, the shipped 0.001–200 Hz defaults put roughly two thirds of them beyond any real record, all empty.

Expect fewer bins out than you asked for. n_bins counts bin edges, so it yields at most n_bins - 1 intervals, and empty bins are then dropped. The shipped n_bins = 151 gives 104 on a typical PNR window; 101 gives 75.

The average is geometric, matching the scale the bins are spaced on:

\[\bar{A}_i = 10^{\;\langle \log_{10} A \rangle_{\,f \in \text{bin } i}}, \qquad \bar{f}_i = \sqrt{f_i^{\text{lo}} f_i^{\text{hi}}}\]

Empty bins are expected — log bins over a linear frequency grid are inevitably sparse at the low end — and are dropped.

Membership is computed, not tested:

\[i(f) = \left\lfloor \frac{\log_{10} f - \log_{10} f_{\min}}{\Delta \log_{10} f} \right\rfloor\]

so every sample belongs to exactly one bin. The previous implementation tested \(f \ge f^{\text{lo}}_i\) and \(f \le f^{\text{hi}}_i\) against each edge in turn — closed at both ends, so a sample landing on an interior edge belonged to two bins. On a CWT axis this reported 51 bins from 49 samples: more bins than samples, which is only possible by double counting.

6. Noise rescaling and rotation

Rescaling

The two windows are rarely the same length, and a shorter record spreads the same power over fewer bins, so the noise is put on the signal’s footing:

\[A_{\text{noise}} \leftarrow A_{\text{noise}} \sqrt{N_{\text{signal}} / N_{\text{noise}}}\]

It is then interpolated onto the signal’s frequency axis. np.interp does not extrapolate — it repeats the edge value — so below the noise window’s own lowest frequency the “noise level” is a flat continuation rather than a measurement. §7 is what keeps the selected band out of that region.

Raising the noise level

A recorded noise window understates the noise beneath a strong signal: it is a sample of the same process, but taken where the signal is not. How to correct for that is a modelling choice, not a fact, so core.noise holds a set of methods behind one signature — given frequencies, noise and signal, return a multiplicative factor. NOISE_MODELS maps names to implementations, the same way transforms.ESTIMATORS does for spectral estimators, so the band search never needs to know which was used.

Name	Status	Assumption
boost	default	Noise under the signal follows the recorded shape, scaled by a power of a frequency ramp
none	available	The recorded window is representative as measured
rotate	available	Legacy ROT_METHOD = 1; the discrepancy is a tilt in log–log space rather than a level offset

none is not a placeholder. It is the honest choice when the noise window is genuinely representative, and it is what a run needs in order to show what the correction is doing — every other model should be compared against it before being trusted, because the assumption is the whole content of the method, and it propagates into every bandwidth and so into every \(\Omega\).

The default, boost, lifts the low and high tails independently until each touches the signal, then keeps the larger of the two at every frequency.

Each bin is mapped onto a scale \(s(f)\) running from \(\approx 0\) at one end of the band to \(\approx 1\) at the other, and the noise is raised by

\[A'(f) = A(f)\,s(f)^{-\eta}\]

Since \(s < 1\), increasing \(\eta\) raises the small-\(s\) end fastest. The exponent wanted is the smallest \(\eta\) at which any bin in the half reaches the signal. That has a closed form. A bin touches when

\[A\,s^{-\eta} \ge S \iff \eta \ge \frac{\ln(S/A)}{-\ln s}\]

so the first touch across the half is

\[\boxed{\;\eta^\star = \min_{i \,:\, s_i < 1} \frac{\ln(S_i/A_i)}{-\ln s_i}\;}\]

with \(\eta^\star = 0\) when some bin already satisfies \(A \ge S\), and no lift at all when no bin has \(s < 1\).

Why this matters beyond elegance. The original implementation searched for \(\eta\) by stepping it in increments of \(\mathrm{inc} = 0.05\) and stopping at the first step past the touching point — that is, it returned \(\mathrm{inc}\cdot\lceil \eta^\star/\mathrm{inc}\rceil\). Two consequences, both measured:

• It was irreproducible. The stopping test is a comparison, so the result was a step function of its input. Two machines differing in the last bit landed on either side of a step and the noise moved by \(s_{\min}^{-0.05} = 1.41\). Observed on CI: 41% and 82% — one and two steps.

• It was biased. Rounding was always upward, so the lifted noise was consistently overstated — a median \(1.18\times\), up to \(1.41\times\), across 39 lifts on the 28 PNR windows. That made signal-to-noise pessimistic at exactly the band edges the ratio is read from.

Using \(\eta^\star\) directly fixes both: the exponent is now a continuous function of its input, and it is the quantity the algorithm was always trying to compute. Measured end to end, perturbing the input by \(10^{-15}\) now moves the noise by \(1.8\times10^{-11}\) and no band edge at all.

The rotate method

ROT_METHOD = 1, described in the legacy source as “actual rotation, quite aggressive”. Writing \(X = \log_{10} f\) and \(Y = \log_{10} A\), it tilts the noise spectrum about its low-frequency end:

\[Y'(\theta) = X\sin\theta + Y\cos\theta + Y_0\,\theta\]

The trailing \(Y_0\theta\) is what makes this a rotation about the low-frequency end rather than about the axis origin — without it the curve translates as well as tilts, and the correction stops being anchored to the part of the noise record least contaminated by signal. As with boost, one angle is found for each half and the larger result kept at every frequency.

It assumes the recorded window has the right level somewhere and the wrong slope, where boost assumes the right shape and the wrong level. Those are genuinely different claims about the noise process, which is why both are kept: comparing the two bands is the only way to see how much of a result is the method. On the 28 PNR windows they disagree on 25, rotate always narrower — LV.L002..HHE runs 1.51–38.60 Hz under boost and 1.51–13.87 Hz under rotate.

Unlike \(\eta^\star\), the touching angle has no closed form: \(\theta\) appears inside \(\sin\) and \(\cos\). It is bracketed on a coarse grid and then bisected to \(10^{-12}\), which is enough to make it continuous in the input — the grid only chooses the bracket, never the answer. The legacy stepped \(\theta\) by a fixed increment and stopped at the first trial past the touch, with exactly the irreproducibility described above.

Two departures from the legacy implementation are recorded in REFACTOR_PLAN §4.5.3: the solved angle, and taking the low/high split from the signal rather than the noise. Neither can move a published number — ROT_METHOD = 1 was commented out on master and has never produced one.

7. Signal-to-noise and bandwidth

The ratio is taken bin by bin on the binned spectra, which share bin edges by construction because the noise was moved onto the signal’s axis before binning:

\[R_i = \bar{A}^{\,\text{signal}}_i \big/ \bar{A}^{\,\text{noise}}_i\]

Which bins that leaves is a registry, core.bandwidth.BANDWIDTH_SELECTORS, in the same way the noise models in §6 are. Three are shipped; the two automatic ones can disagree completely:

peak — the default ([snr] bandwidth_method)

Walk outward from the bin with the highest ratio until \(R_i\) falls below \(R_{\min}\) in each direction. This says the usable band is the one containing the strongest part of the signal, even where a wider passing run exists elsewhere. It is what the legacy BW_METHOD = 2 did.

widest

The widest contiguous run of bins with \(R_i \ge R_{\min}\), bridging gaps of a single failing bin so one noisy bin does not end a band, subject to a min_width (default 3).

fixed

Not a selection. Returns the band named in [snr] fixed_band, intersected with the frequencies present, whatever the ratio says. For a band held constant across a study, or one imposed by an instrument. The pair records band_imposed in its metadata, so a result read back later still distinguishes a band that was asserted from one that was measured.

Either automatic selector returns nothing — not a band plus a flag — when no band survives. So does the resolution floor below, and the Nyquist ceiling, both of which can reject a pair outright rather than only narrowing it.

They are not cosmetic variants. On a spectrum with a wide low-frequency passing run beneath a narrow high-SNR peak, peak and widest return disjoint bands. Since the low edge is what constrains \(\Omega\), reconstructing a result means matching the selector as well as the threshold — it is recorded in the configuration stamp on every output.

The previous method integrated \(\mathrm{sign}(R - R_{\min})\) and read the edges off the 1st and 99th percentiles, with a retry loop when they crossed. Every step there is discontinuous and they compound: an edge could move 13 bins between machines. It also lagged: on a clean 5–30 Hz passing region it returned 9.41 Hz for the low edge, because the 1st percentile of a cumulative integral arrives late. The low edge is what constrains \(\Omega\). widest returns 5.45 Hz on the same input, within one bin of the truth. Both shipped selectors are continuous in the input, which is the property that mattered; the 9.41 Hz figure is what the legacy method gave.

Resolution floor. The band is finally clamped to

\[f_{\text{low}} \ge \max\bigl(\min f^{\text{signal}},\; \min f^{\text{noise}}\bigr)\]

which refuses the region where the noise level is np.interp’s repeated edge value rather than a measurement. On the 28 PNR pairs, 6 selected a band opening below the noise window’s own \(1/T\) before this existed. Set snr.resolution_floor = false to reproduce a run made before it.

Nyquist ceiling. The other end has the opposite problem, and it is not guarded by default. Both automatic selectors walk while the ratio holds, and approaching Nyquist the signal and the noise roll off into the anti-alias filter together — so \(R_i\) can stay above \(R_{\min}\) through a region where neither is informative, and the walk runs until it reaches the end of the array. On the 28 PNR pairs the median high edge lands at 66% of Nyquist and five run past 90%, which is the anti-alias response being fitted as though it were a source spectrum.

\[f_{\text{high}} \le \tfrac{1}{2} f_s \cdot \phi\]

where \(\phi\) is [snr] max_nyquist_fraction, unset by default. A fraction rather than a frequency because a network mixes sampling rates — the PNR stations record at both 100 and 200 Hz — and the roll-off moves with the rate. The right value is a property of the instrument rather than of this package, which is why nothing is imposed.

It is applied after the selector and after the floor, so it constrains every strategy the same way, fixed included: it is a statement about the recording, not about how the band was chosen.

This is not a presentational setting. Capping the PNR event at \(\phi = 0.8\) moves 10 of the 28 bands, and on those the fitted corner frequency changes by a median of 16% and by a factor of three at the worst station. Since stress drop scales as \(f_c^3\), that worst case is a factor of 27. It does not follow that the capped answer is the correct one — what follows is that the band is a decision with that much leverage, and it should be made deliberately rather than inherited from a default. The tutorial works through all three levers on this event, including raising \(R_{\min}\), which tightens both edges instead of one and begins rejecting stations entirely.

8. Source model

Fitted in \(\log_{10}\) space over the selected band. The generalised Boatwright source shape is

\[\log_{10} S(f) = \log_{10}\Omega - \frac{1}{\gamma}\log_{10}\!\left[1 + \left(\frac{f}{f_c}\right)^{\gamma n}\right]\]

with \((\gamma, n) = (1,2)\) for Brune and \((2,2)\) for Boatwright. Attenuation is a frequency-independent \(t^\ast\),

\[\log_{10} D(f) = -\frac{\pi f t^\ast}{\ln 10}\]

(or \(f^{1-a}\) in the frequency-dependent variant), and the motion term converts from displacement:

\[\log_{10} G(f) = \begin{cases} 0 & \text{displacement} \\ \log_{10}(2\pi f) & \text{velocity} \\ 2\log_{10}(2\pi f) & \text{acceleration}\end{cases}\]

The fitted model is their sum:

\[\log_{10} A(f) = \log_{10} S(f) + \log_{10} D(f) + \log_{10} G(f)\]

Free parameters are \(\Omega\), \(f_c\) and \(t^\ast\), minimised with Powell’s method by default. Turning the fitted \(\Omega\) into a magnitude is §9.

A note for anyone adding a source model. Madariaga is omega-squared like Brune and sits at the same \((\gamma, n) = (1,2)\), so adding it as a spectral shape alone changes no fitted parameter. The difference is the constant relating \(f_c\) to source radius, and since stress drop goes as \(r^{-3}\) that is roughly an order of magnitude on identical data. The \(f_c\)-to-radius scaling has to be a named property of the model, not a constant buried in whatever computes stress drop. See REFACTOR_PLAN.md §4.6.5.

---

9. Moment and magnitude

specmod.magnitude. The fit gives \(\Omega\) per channel; this turns it into a seismic moment and a moment magnitude.

\[M_0 = \frac{4\pi\rho\beta^3 R_0}{\Theta F}\cdot\frac{\Omega}{G(R)} \qquad\text{[N m]}\]

\[M_w = \tfrac{2}{3}\left(\log_{10} M_0 - 9.1\right)\]

\(G(R)\) is the geometrical spreading (specmod.spreading), which corrects the observed plateau back to the reference distance \(R_0\). The magnitude relation is Hanks and Kanamori (1979).

The constants

Defaults are the S-wave values of Holt (2019) Ch. 1 §1.4 and Ch. 2 §2.2, all taken at the source, and all overridable through MediumConstants:

Symbol	Field	Default
\(\rho\)	density	2700	kg m⁻³
\(\beta\)	velocity	3500	m s⁻¹, cubed here
\(\Theta\)	radiation_pattern	0.55	average SH over the focal sphere
\(F\)	free_surface	2	vertically incident SH
\(R_0\)	reference_distance_m	1000	metres

\(\Theta = 0.55\) is the SH average and pairs with \(F = 2\). The textbook 0.63 is the RMS over total S — a different quantity, and substituting it while keeping \(F = 2\) is not a small correction. No partition factor is applied.

Two distances, both correct

\(R_0\) is in metres, because it sits beside kg m⁻³ and m s⁻¹ and that is what makes \(M_0\) come out in newton-metres. \(G(R)\) takes its distance in kilometres.

That looks like a units bug and is not: they are different distances. They cancel only when the spreading exponent is exactly 1, so a model with any other exponent genuinely needs both, in their own units.

\(\Omega\) is linear, llpsp is not

The fit reports llpsp — \(\log_{10}\Omega\). seismic_moment takes \(\Omega\) itself and raises ValueError on a non-positive plateau rather than returning a moment roughly \(10^{40}\) times too small, which is what passing the logarithm straight through would otherwise produce.

And \(\Omega\) must be the displacement plateau, in m s. That is the reason §4 converts to MAGNITUDE: reading a folded FAS plateau as \(\Omega\) puts \(M_0\) out by two, which is 0.2 magnitude units.

Station values, then an event value

station_moments appends m0 and mw per channel. event_magnitude aggregates, following Holt (2019) §2.2:

• the sample mean of station magnitudes, excluding anything beyond outlier_sigma (default 2.5) standard deviations, in a single pass rather than iterated to convergence;

• no event value at all below min_stations (default 3). Fewer than three stations cannot average out the radiation pattern, so the published method declines to report a value rather than reporting a poor one, and this raises instead of returning something usable-looking.

Averaging in magnitude rather than in \(M_0\) makes the result a geometric mean of moments. That is the published choice, not an oversight.

Warning

The absolute calibration is unverified. The shape of the fit is checked against golden references, and the constants above are sourced; what has not been confirmed is that the resulting \(M_w\) reproduces the published catalogue values on the same events. Treat station-to-station and event-to-event differences as sound and the absolute level as provisional until §4.7 of REFACTOR_PLAN.md is closed.

---

Reproducibility

As of the change described in §6 and §7, the pipeline is continuous in its input: perturbing a record by \(10^{-15}\) moves the noise by \(1.8\times10^{-11}\) and no band edge at all. Before, a last-bit difference between two machines moved the noise by up to 82% and a band edge by 13 bins.

That is checked, not assumed. tests/golden/pipeline_reference.json records the full pipeline output over 28 real windows and 5 estimators, and tests/test_golden_reference.py holds every later change to it. Regenerate with python tools/make_golden.py, and say in the commit message what moved and why — regenerating without that explanation removes the only check standing between a refactor and a silently different \(\Omega\).