This is the first parametric fit function presented for luminescence rock surface exposure dating (Laskaris and Liritzis, 2011). This was suggested for both OSL and TL. 1L(x;a1,b1,c1,d1)=a1+b12erfcln⁡(xc1)d12 where L is the distribution of luminescence signal with respect to depth, and it depends on the fitting parameters as a1, b1, c1, and d1. While this empirical model enables precise fitting to the luminescence depth profile, it does not explain the luminescence mechanism inside the rock, and the derived fitting parameters are not linked to the natural kinetic parameters. Additionally, this model assumes zero erosion conditions, which are not always achievable. Thus, this model was not incorporated into any further LRSED studies.

Sohbati et al. (2018) proposed a model whereby the instantaneous concentration of trapped electrons n at depth x is: 2dndt=(N-n)F(x)-nE(x) where, 3E(x)=σφ0‾e-μx4F=D˙(x)/D0 With the boundary conditions n=N at t=0, the analytical solution was evaluated as 5n(x,t)N=E(x)e-t[E(x)+F]+FE(x)+F6n(x,t)N=σφ0‾e-μxe-t(σφ0‾e-μx+D˙D0)+D˙D0σφ0‾e-μx+D˙D0 The model was further improved by incorporating an erosion scenario (Sohbati et al., 2018). 7x(t)=x0-ε˙t8E(x(t))=σφ0‾e-μ(x0-ε˙t)=(σφ0‾e-μx0)eμε˙t=E0eμε˙t9dndt=(N-n)F-nE0eμε˙t where the analytical solution, expressed as a confluent hypergeometric function M, is: 10n(x,ε˙)N=M1,1+Fμε˙,-E(x)με˙

The analytical solution improves the feasibility of modelling by incorporating the effects of dose rates, first-order kinetics (FOK) decay behaviour and erosion. However, in the case of polymineral or feldspar samples, where the infrared stimulated luminescence (IRSL) of feldspar minerals exhibits non-first-order decay, the general-order kinetic (GOK) model is more suitable for these samples.

Freiesleben et al. (2023) developed this model to incorporate GOK: 11dndt=-Eε˙(x,t)nr where r is the order of kinetics. 12Eε˙(x,t)=σφ0‾e-μ(x0-ε˙t)13n(x,tε˙)=[(r-1)tε˙E(x)+n1-r]11-r where the apparent exposure time is given as 14tε=1με˙(1-e-με˙t)

This model considers the order of kinetics and erosion scenario but ignores the effect of the dose rate on LDP propagation. However, in natural environments, the dose rate considerably affects the propagation of LDP over a longer timescale (>103 years), and this effect is amplified with an increase in the order of kinetics (Pathan et al., 2024). This model also assumes a constant average rate of erosion, which may not be applicable in different erosion scenarios, such as step-function or stochastic histories (Ganti et al., 2016; Brown and Moon, 2019; Lehmann et al., 2019a).

The modelling framework adopted in this study brings together the current understanding of rock-surface luminescence signal evolution within a single computational framework. It includes: (1) general-order kinetics, (2) advection through the natural mechanism of erosion, and (3) ambient dose-rate-driven signal accumulation. This formulation was required to simulate the luminescence-depth profile more accurately and builds upon earlier work by Lehmann et al. (2019a), Biswas et al. (2023), and Pathan et al. (2024). Lehmann et al. (2019a) introduced the advection term in response to erosion as ε˙dn‾dx but for the FOK (r=1) in addition to earlier models. Thus, the instantaneous change in normalised luminescence n‾ (see Sect. 3.1 for normalisation procedure) with respect to time t at depth x, is expressed as: 15dn‾dt=D˙D0(1-n‾)growth-{n‾rσφ‾0e-μx}decay+ε˙dn‾dxerosion

This model can be solved using a finite-difference method with a second-order upwind scheme for the advection term rather than altering the depth term (Sohbati et al., 2018). Furthermore, Biswas et al. (2023) modified this model for a zero erosion scenario by incorporating GOK (r>1) in the decay term and demonstrated the severe underestimation of the exposure age if FOK was used. Pathan et al. (2024) solved the GOK model (Eq. 15) for a non-zero erosion scenario and reported an overestimation of the erosion rate when FOK was used. Thus, we incorporate the most recent model, as in Eq. (15), into the mathematical framework of CoRSEER.

The model accounts for signal growth due to environmental dose rate, depth-dependent light attenuation, nonlinear bleaching behaviour of IRSL signals, and advective loss of material due to surface erosion. Therefore, it provides a physically consistent framework for simulating LDP evolution under both exposure and erosion. This class of model has been used in several studies. Lehmann et al. (2019a) and Elkadi et al. (2022) used this model assuming FOK (r=1) to jointly invert 10Be cosmogenic nuclide data and LRSED profiles, allowing erosion histories and erosion-corrected exposure ages to be constrained through complementary dating systems. Smedley et al. (2021), Brill et al. (2021), and other studies also applied this framework for reconstructing erosion histories, generally assuming FOK.

The incorporation of general-order kinetics is particularly important for feldspar-based IRSL signals, where slower and non-exponential bleaching behaviour is commonly observed. The FOK model assumes simple exponential decay with (r=1), whereas the GOK model introduces a kinetic-shape parameter (r), which can represent slower bleaching behaviour observed in feldspar signals (Duller, 1997; Chen and Leung, 2003; Guralnik et al., 2015). Previous studies have shown that the use of FOK may lead to underestimation of exposure ages and overestimation of erosion rates, in some cases by up to an order of magnitude (Biswas et al., 2023; Pathan et al., 2024; Luo et al., 2026). By embedding this model within CoRSEER, we provide a reproducible platform for forward and inverse modelling of LRSED profiles, enabling more consistent reconstruction of exposure–erosion histories. The prerequisites for CoRSEER, sampling strategy, and sample-processing methods for LRSED are provided in Sects. S1–S3 in the Supplement, respectively.

To eliminate the effect of intersample variability (mass and luminescence production efficiency), the test-dose normalised luminescence (Lx/Tx) was measured at different depths to generate the LDP. The LDP is further adjusted (or normalised) to plateau at 1 to facilitate further mathematical operations (Sohbati et al., 2012a). Traditionally, users select a depth beyond which the experimental LDP is assumed to be saturated through visual inspection and compute the mean Lx/Tx of all points thereafter to normalise the entire profile. However, the high scatter in the saturation region often leads to subjective depth selection, introducing human error and bias. The incorrectly identified plateau can systematically distort the normalised signals and skew the kinetic parameter estimation. To overcome this, we present an empirical normalisation using the sigmoidal curve fitting (SCF) method. We chose the following three-parameter logistic sigmoidal function (Kaufmann, 1981) to model the luminescence-depth profile: 16f(x)=a1+exp⁡[-b(x-c)] Here, a is the upper asymptote, the maximum value the function can attain as x∼∞, denotes the saturation luminescence intensity (or Lx/Tx), and b is the growth rate, which controls the steepness of the curve. A higher b indicates a steeper transition of Lx/Tx with depth. c represents the inflection point or x0.5, that is, the depth at which half of the saturation signal is reached (Fig. 1). To derive a, b, and c, we fitted Eq. (16) to the experimental LDP using weighted nonlinear least-squares regression in MATLAB. The initial guesses for fitting parameters were set to a0=max⁡(LxTx), b0=1 and c0=median(x). We enforced bounds to ensure physically meaningful solutions: (a) lower bounds: [0,0,0], (b) upper bounds: [∞,∞,max⁡(x)], and provided the weight factors (wi) as wi=1α where α is the fractional error. The experimentally observed LDP (Lx/Tx vs. depth) was normalised with the respective a value for each sample and referred as SCF-normalised LDP. This approach was used to identify the plateau and normalise the experimental data. This method removes bias, improves reproducibility, and ensures a consistent LDP normalisation algorithm across diverse datasets by eliminating manual intervention.

In inverse modelling, multiple LDPs are simulated by solving Eq. (15) using the finite difference method for two main scenarios: (a) varying kinetic parameters, σφ0‾, and μ (keeping the exposure-erosion history as fixed) or (b) by varying the exposure-erosion history (keeping σφ0‾, and μ as fixed). For a calibration sample (known exposure-erosion history), kinetic parameters (σφ0‾, and μ) are randomly varied, whereas for natural samples (known σφ0‾, and μ), the exposure-erosion history is randomly varied. The other kinetic parameters, D˙, D0, and r are constrained independently using DRAC (Durcan et al., 2015; Sohbati et al., 2018), IRSL dose-response curve (Wintle and Murray, 2006) and fitting decay of IRSL with light exposure (Biswas et al., 2023), respectively. These parameters are user-defined in CoRSEER. The most probable, or best-fitting, LDPs and corresponding parameters are constrained by comparing the simulated LDPs (n‾sim) and experimentally observed SCF normalised LDP (n‾obs) and by minimising the chi-square misfit (χ2) and maximising the likelihood, following the approach by Lehmann et al. (2019a). 17χ2=∑i=1l(n‾)obs(i)-(n‾)sim(i)2δ218likelihood=P=e-χ2/219Pnormk=PkPmax where l is the number of experimental data points, δ is the standard deviation of the residual between the observed experimental profile and the best-fitting sigmoidal curve, divided by “a” factor, Pk is the likelihood value associated with the simulated LDP numbered as k, Pmax is maximum-likelihood, which is associated with the best-fitting synthetic LDP and Pnormk is the normalised likelihood value associated with the simulated LDP numbered as k. The best LDP and corresponding parameters were selected by applying the Monte Carlo acceptance criterion as Pnormk>U, where U is a random number between zero and one. Finally, CoRSEER establishes the values of best-fit, median, ±1δ, and ±2δ for all required parameters with confidence intervals.

Across the published calibration datasets, CoRSEER-I generally reproduces literature calibration parameters, with the best agreement observed for Lehmann et al. (2019a) and Luo et al. (2022), where the CoRSEER-I-derived log⁡10(σφ0‾) and μ values are broadly comparable to reported estimates. This agreement is consistent with a similar normalisation choice for the Lx/Tx depth profiles. In contrast, several datasets show systematic offsets between CoRSEER-I and the literature values. For example, in Balco et al. (2023), the sample IR50_KAY_110, the calibration parameters shift to higher values, from log⁡10(σφ0‾)=-7.30 and μ=0.34 mm-1 to -6.58 and 0.51 mm-1 whereas for the sample pIRIR225_KAY_110 shifts to lower values, from -7.68 and 0.54 mm-1 to -7.77 and 0.36 mm-1. In Brill et al. (2021), CoRSEER-I predicted decreased calibration parameters for IR50_HAR1-1_CAL but increased parameters for IR50_TEM3-1_CAL and IR50_VAL4-1_CAL1. For the two calibration samples for which Brill et al. (2021) did not report the final parameters (IR50_RAB5-1_CAL and IR50_VAL4-1_CAL2), CoRSEER-I yielded -7.38, 0.76 mm-1 and -7.50, 0.37 mm-1, respectively. In Smedley et al. (2021), the ROAD set was dominated by a decrease in both parameters, with two clear increased cases (IR50_ROAD01 and IR50_ROAD03). The Lehmann et al. (2018) dataset shows predominantly decreased parameters for MBMB samples, with one increased case (IR50_MBMV7) and one near-balanced case (IR50_MBMV1). We attribute these discrepancies primarily to (1) differences in the fitting strategy, (2) differences in the Lx/Tx normalisation (manual plateau selection versus automated normalisation) and (3) both the σφ0‾ and μ may slightly covary during optimisation of the best fitting. The boundary ranges of log⁡10(σφ0‾) and μ values predicted by CoRSEER-I are -8.41 to -4.82 and 0.36 to 3.36 mm-1 (Table 2), showing the plausible flexibility of the app for a wide variety of exposure conditions and lithologies.

We plotted CoRSEER-I vs. the literature dataset and evaluated the R2 value with respect to the 1:1 line to summarise the compilation-scale performance for 23 samples. In Fig. 5a, μCoRSEER versus μliterature gives R1:12=0.562 for n=23, meaning CoRSEER-I tracks published μ values moderately well, even though a large fraction of points falls below the 1:1 line. In Fig. 5b, σφ0‾CoRSEER versus σφ0‾literature gives R1:12=0.316 for n=23, showing weaker agreement and stronger scatter. It is well established that the σφ‾0 and μ covary, and we present the relationship between them in Sect. S5 in the Supplement. To investigate the offset between the established calibration parameters and the underlying reason behind it, we plot (log⁡10(σφ0‾)/μ)CoRSEER vs. (log⁡10(σφ0‾)/μ)literature in Fig. 5c which gives R1:12=0.379 showing weaker agreement. This suggests that covariance contributes only weakly to the offset, whereas normalisation and fitting strategy are the dominant contributors. Visually, most points fall below the 1:1 line in both panels, so the dominant overall tendency in this compilation is that CoRSEER estimates lower μ and σφ0‾ than those reported in the literature for many samples, while a smaller set of samples shows the opposite paired behaviour.

We evaluated the apparent exposure ages (tapp) of the unknown age LDPs of the literature dataset using CoRSEER-II for 41 samples. As the evaluated age exclusively depends on the chosen calibration samples, dose rate, and characteristic dose, we follow the same pattern as given in the literature dataset (unknown-age LDPs), except for a few, as mentioned further. Lehmann et al. (2019b) used a combined calibration parameter of IR50_MBTP7_CAL (2 a) and IR50_MBTP8_CAL (11 a), whereas the CoRSEER-I supports only a single calibration sample. Thus, we used IR50_MBTP7_CAL, as it closely resembles the calibration parameters in the literature. For Brill et al. (2021), we used CoRSEER-I to evaluate the parameters of IR50_RAB 5-1 CAL for the samples of IR50_RAB 1-1, IR50_RAB 1-2, and IR50_RAB 5-1 (these samples were evaluated using IR50_HAR 1-1 CAL in the literature due to the large uncertainty in IR50_RAB 5-1 CAL). We plotted the CoRSEER-II derived versus published literature apparent age dataset and evaluated the R1:12 value with respect to the 1:1 line to summarise the compilation-scale performance. Interestingly, the evaluated tapp using CoRSEER-II is comparable to the literature and falls closely around the 1:1 line (Fig. 6) despite a significant difference in the calibration parameters used in this study. The derived R1:12=0.889 represents the CoRSEER-II track, which has apparent exposure ages with high certainty. There was a significant underestimation by CoRSEER-II for the IR50_RAB 1-1, IR50_RAB 5-1 and IR50_TEM 2-1. This is due to the following reasons: (1) a change in the calibration sample, as mentioned earlier. (2) The high SCF normalisation factors a=1.1 (up to 10 %) of the sigmoidal fitting for the IR50_TEM 2-1, which might have changed the position of x0.5.

The range of tapp values predicted by CoRSEER-II extends from 1 a to 8.5 ka (Table 3), showing its potential as a decadal-to-millennial chronometric mathematical tool for LRSED.

Here, we establish the steady-state erosion rate (ε˙ss) and corresponding onset time of erosion (tss) of literature dataset (n=11) using CoRSEER-III. We carefully chose the calibration and dose parameters, as mentioned in the earlier sections. The choice of sample for analysis assumes that these samples are in a steady state, as described in the literature (Lehmann et al., 2019a; Smedley et al., 2021). Comparison of the literature and CoRSEER-derived erosion rates is important because, unlike the previous studies, CoRSEER-III derives erosion rates directly from the ε˙ss–tss matrix. Therefore, the present result is not only a comparison with previously reported values, but also the first systematic demonstration of an algorithmic method for extracting erosion-rate solutions from this matrix. The results are shown in Table 4.

The CoRSEER-III-derived erosion rates broadly agree with the literature erosion-rate values. The literature erosion rates range from 0.0035 to 4.3 mma-1, whereas the CoRSEER-III-derived erosion rates range from 0.003 to 4.327 mma-1. Using the literature values as reference, the absolute percentage difference ranges from 0.6 % to 86.7 %. The minimum difference occurs for IR50_MBTP6, where CoRSEER-III gives 4.327 mma-1 compared with the literature value of 4.3 mma-1. The largest difference occurs for IR50_MBTP5, where CoRSEER-III gives 0.560 mma-1 compared with the literature value of 0.300 mma-1.

The uncertainty associated with CoRSEER-III-derived erosion rates remains within a reasonable range for most samples. The relative upper error in the CoRSEER-III erosion rates varies from 10.8 % to 44.9 %, whereas the relative lower error varies from 10.8 % to 33.3 %. The lowest relative uncertainty occurs for IR50_MBAM2, where the erosion rate is 0.037 mma-1 with an uncertainty of approximately ±10.8 %. The highest upper uncertainty occurs for IR50_BALL02, where the upper error reaches approximately 44.9 %. These values show that the algorithm provides bounded erosion-rate estimates rather than single manually selected solutions.

In contrast, the erosion-onset time, tss, shows a considerably larger departure from the literature values. The literature tss values range from 4 to 3146 a, whereas the CoRSEER-III-derived tss values range from 3 to 3080 a. Using the literature values as reference, the absolute percentage difference ranges from 18.0 % to 64.1 %. In most samples, CoRSEER-III estimates a younger erosion-onset time than the literature value. The only exception is IR50_MBTP1, where CoRSEER-III gives a larger onset time of 3080 a compared with the literature value of 2610 a, corresponding to an increase of approximately 18.0 %. The maximum difference occurs for IR50_MBTP5, where CoRSEER-III estimates 23 a compared with the literature value of 64 a, corresponding to a reduction of approximately 64.1 %.

Overall, the results indicate that CoRSEER-III reproduces the erosion-rate field reasonably well, as presented in Table 4, but it estimates erosion-onset time differently from the literature. This difference is expected because the CoRSEER-III estimate is derived from an automated best-fitting erosion-time slope detection algorithm, whereas previous studies relied more strongly on manual or interpretative selection of model solutions.

When the sample is in a transient state, one can visually identify the solution that generally lies in the sloped portion of the ε˙ss vs. tss likelihood plot. It is mathematically challenging to produce an objective method to constrain a pair of possible solutions. Thus, we do not present any transient-state condition scenarios. But it is possible to derive a transient state solution from the results of CoRSEER-III if the onset of erosion (ts) is known by some means.

Large scatter in the luminescence depth profile (LDP) and high fractional uncertainty in Lx/Tx make plateau identification unreliable and reduce the stability of any inversion (i.e., the SCF normalisation, estimation of calibration parameters, exposure age, erosion rates). We can reduce Lx/Tx uncertainty by adopting a modified post-infrared infrared stimulated luminescence protocol with lower heating rates and longer isothermal holding times (Jenkins et al., 2018), and by using metal cups (Elkadi et al., 2021). For each sample, we recommend measuring two or more LDPs and maintain millimetre-scale sampling density with a substantial fraction of points located on the sigmoidal transition. Importantly, the fractional error structure should be similar or random with depth; if the fractional error at lower depths is systematically higher than at higher depths, the fitting becomes biased toward the plateau, and the inferred parameters become unreliable. Although we cannot yet provide a statistical criterion for the required fraction of points on the sigmoidal part of the curve, based on our observation in Figs. 2a and 3a, we recommend that more than 50 % of all Lx/Tx measurement points should lie before the final plateau. If the measured profile is dominated by plateau points, we recommend reducing the measurement depth such that the total number of points on the initial plateau and sigmoidal region exceeds 50 % of the total.

In the case of incomplete LDP, the maximum measurement depth is shallower than the steady-state depth, the plateau cannot be reconstructed. We therefore recommend measuring Lx/Tx up to the equilibrium depth (which is the stable plateau). Similarly, for the large scatter in experimental data, it is difficult to identify the plateau manually. The SCF and parameter estimation to normalise the LDP is a major step forward as it resolves both issues and removes the manual intervention from mathematical inversion, thereby improving the accuracy of fitting (Fig. 2e).

While using CoRSEER-I, we suggest that there are a few precautions we must take to estimate σφ‾0 and μ accurately. If the IRSL signal from feldspar minerals is used for LDP measurements, we suggest evaluating the order of kinetics (r) beforehand and incorporating it into the calculation. The application of FOK can cause severe underestimation and overestimation of the exposure age and erosion rate (Biswas et al., 2023; Pathan et al., 2024). The different methods for evaluating the r value are thoroughly discussed in Biswas et al. (2023) and Pathan et al. (2024). Furthermore, if the calibration sample is young (say 1–10 a), for example, a known-age deglaciated surface or manually exposed surface created by chipping off a boulder face, erosion is assumed to have no effect on the LDP. In contrast, when using older natural calibration samples (t>100 a), erosion is expected to affect LDP significantly (Lehmann et al., 2018; Smedley et al., 2021), and this effect becomes stronger in high-erosion environments. This can lead to substantial underestimation of σφ‾0 and smaller bias in μ. This is inferred from the samples IR50_MBMV1 (t=137 a) and IR50_MBMV6 (t=2 a), which were samples from the same site with similar lithology, but the log⁡σφ‾0 was estimated as -7.8 and -6.13, respectively, which is a difference of about two orders of magnitude, whereas μ was estimated at 1.04 and 0.8 mm-1. Thus, the erosion rate must be explicitly incorporated into CoRSEER-I. To avoid this complexity and reduce model dependence on uncertain erosion histories, it is recommended to use young calibration surfaces at sampling sites where there is no erosion over the period of interest. This allows for the precise calibration of kinetic parameters under well-constrained exposure conditions. This will further remove any ambiguity due to changes in decay kinetics, lithology, surface orientation, shielding, and light penetration.

When selecting a calibration sample, it is critical to recognise that the light attenuation coefficient (μ) and the effective detrapping rate at the surface (σφ0‾) of the calibration samples are the best representatives of the actual sample. Recent empirical studies demonstrate that photon penetration and attenuation are highly sensitive to micro-mineralogical heterogeneity, such as the spatial distribution of opaque and translucent minerals, which can cause localised “light piping” or “shadowing” (Meyer et al., 2018; Cui et al., 2024). Furthermore, μ is strongly wavelength-dependent and can be drastically altered over time by dynamic surficial boundary conditions, including the accretion of rock varnish (Luo et al., 2018) or localised iron-hydroxide precipitation due to moisture (Meyer et al., 2018; Smedley et al., 2021). The effective detrapping rate (σφ0‾) is similarly sensitive to external environmental geometries, fluctuating significantly based on the target surface's topographic aspect, inclination, and seasonal solar incidence angles (Fuhrmann et al., 2022). High-resolution spatial analyses reveal that these optical properties can vary significantly even on a sub-millimetre scale within contiguous rock formations (Ou et al., 2018; Smedley et al., 2025). Consequently, to avoid severe systematic chronological errors, it is imperative to ensure that the optical characteristics, dynamic surficial coatings, and spatial orientation governing μ and σφ0‾ in the calibration sample, the measurements are meticulously matched to those of the unknown target sample.

Although our present modelling framework focuses on single calibration (SC), it is worth noting that multi-calibration strategies (MC) can be highly effective under certain conditions. When multiple calibration samples of different known ages, but from the same lithology and sunlight exposure conditions, are available, it becomes possible to constrain σφ‾0, μ and r simultaneously (Pathan et al., 2024). However, the accuracy of such multi-calibration depends strongly on the number of calibration samples and their age spread. A wider temporal distribution across well-characterised surfaces enhances the resolution of the kinetic behaviour. While we recognise the potential of this methodology, it lies beyond the scope of the current work and is not implemented in the present version of CoRSEER.

After estimation of kinetic parameters, the estimation of exposure and erosion history appears to be highly dependent on the position of x0.5. There are two possible scenarios:

Case 1: For non-eroding exposure conditions, if the depth of half saturation of the LDP, x0.5 is smaller than x0.5,max, then the apparent exposure age is equal to the actual exposure age of the sample, and if x0.5,max<x0.5≤x0.5,ss then the sample reaches steady-state, and the exposure age is greater than the tmax, maximum dateable exposure age.

Case 2: If the derived apparent ages are substantially lower than the expected age (derived using an independent method), this indicates the presence of erosion, which was seen in the sample IR50_BALL02 (Smedley et al., 2021). CoRSEER-III should be used to further re-establish the exposure-erosion history. Furthermore, when using IRSL to establish the LDP, the r value should be the same for calibration and unknown samples (Pathan et al., 2024). We incorporated all these modelling aspects, diagnostics, and interpretation into the input and output GUI plots (Fig. 3).

The fundamental assumptions and modelling aspects underpinning CoRSEER-III were taken from Lehmann et al. (2019a), where the authors coupled 10Be concentration and the LDP to evaluate the erosion-corrected 10Be exposure age range. The main difference is that CoRSEER-III does not incorporate 10Be concentration; rather, it incorporates independent exposure age (i) which can be derived from an independent dating method or (ii) by incorporating tmax from forward modelling of x0.5 vs. t.

One of the most important observations is that CoRSEER-III captures the same broad erosion-rate scale as the literature values. This indicates that the erosion-rate extraction algorithm is able to preserve the first-order structure of the erosion solution space. In practical terms, this is encouraging because the algorithm avoids subjective manual picking while still recovering erosion-rate estimates comparable to previous studies. This is also a possible reason for the large relative difference between CoRSEER and literature results in a few cases. Thus, for maximum accuracy and precision, we recommend following the guidelines. Users should not run the model once over a very large erosion-exposure window and treat the result as final. A better strategy is to run the model in multiple stages. First, use a broad search window to identify the approximate high-likelihood region. Then shrink the exposure-erosion window around that region. Finally, increase the matrix density so that the best-fitting slope is resolved with higher numerical precision. This stepwise refinement should reduce grid artefacts and improve the stability of both erosion-rate (and errors) and onset-time estimates.