Road safety is a global public health priority, as motorization and travel demand continue to rise faster than improvements in infrastructure and enforcement (WHO, 2023). Horizontal curves are consistently identified as high-risk roadway elements because they require drivers to adjust speed, maintain steering precision, and adapt to varying tire-pavement friction. Failures in these tasks, particularly under low-friction conditions, are a major cause of loss-of-control crashes (AASHTO, 2018; Donnell et al., 2016). Although curves represent only a small share of road length, they account for 20–30% of severe crashes, often intensified by excessive speed, inadequate superelevation, surface deterioration, and insufficient friction (Shalkamy et al., 2021).

India continues to face a severe road safety challenge. In the year 2022, 461,312 crashes caused 168,491 deaths and 443,366 injuries, equivalent to 53 crashes and 19 deaths every hour, with year-on-year increases of 11.9%, 9.4%, and 15.3% (MoRTH, 2022; MoRTH, 2023; MoRTH, 2024). Curve-related crashes alone accounted for 54,593 incidents and 20,573 deaths, while overturning crashes added 20,070 incidents and 9,829 deaths (MoRTH, 2022). Field records from NH-340C further highlight this risk on five study curves experienced 22 crashes and 12 fatalities between 2020 and May 2025. The Krishnapuram Anjineya Swamy Statue curve recorded nine crashes and three deaths due to rash manoeuvres, while Venkatapuram, despite only three crashes, caused eight fatalities linked to over-speeding and animal crossings. These observations underscore the urgent need for localized safety diagnosis and targeted interventions.

Key contributing factors in India include heterogeneous traffic ranging from two-wheelers to heavy trucks, rapid pavement deterioration due to monsoons and poor drainage, mismatches between design and operating speeds, and risky behaviours such as late braking and curve-overtaking (Singh, 2017; Dash et al., 2020; Shailaja et al., 2022; Rosas-López et al., 2021). Among curve types, simple circular curves pose higher risks than spirals because lateral forces depend heavily on speed, radius, superelevation, and pavement condition (Chen et al., 2007; Sharf Aldeen et al., 2022).

Safe negotiation of curves depends on balancing the demand and supply of side friction. The maximum side friction factor (F_max) reflects available supply and is shaped by pavement texture, materials, climate, and tire properties (McLean, 1981; AASHTO, 2018; Gattis et al., 2005; Abohassan et al., 2022). Current Indian standards (IRC:73-2023) assume a fixed F_max of 0.15 regardless of speed, texture, or traffic heterogeneity, risking both under- and over-design (Nicholson, 1998; Morrall & Talarico, 1994). Research has consistently shown friction to be speed-dependent. Barnett (1938) recommended reductions beyond 60 mph, McLean (1981) linked failures to inadequate superelevation, and Mudry (1999) demonstrated that 85th percentile friction often exceeded AASHTO limits (Shalkamy et al., 2021). Bonneson (1999, 2000) confirmed higher friction acceptance at lower speeds and suggested using the 95th percentile of speeds for design, while Fitzpatrick et al. (2000) showed that stability is most compromised at midpoints. These findings stand against static assumptions and call for dynamic speed–friction models in safety evaluation.

Speed profiling has been widely used to study driver behavior, consistently showing deceleration before entry, minimum speed at midpoints, and acceleration at exits. These trends are influenced by curve radius, superelevation, sight distance, shoulder provision, and approach speed (Bennett, 1994; Gong & Stamatiadis, 2008; Bonneson & Pratt, 2009; Shallam & Ahmed, 2016; Llopis-Castelló et al., 2018). Extended deceleration zones often cause discomfort and instability, particularly on circular compared with spiral curves (Chen et al., 2007; Montella & Imbriani, 2015; Himes et al., 2019; Han et al., 2023). Approach speed strongly governs mid- and exit speeds, which in turn influence vehicle stability and crash risk (Himes et al., 2019). Previous models have attempted to predict operating speeds using geometric variables such as shoulder width, radius, superelevation, and sight distance (Bennett, 1994; Gong & Stamatiadis, 2008; Bonneson & Pratt, 2009; Shallam & Ahmed, 2016; Semeida, 2014; Llopis-Castelló et al., 2018). Others stressed the role of approach speed in enabling timely driver alerts, since long deceleration zones reduce comfort and safety (Montella & Imbriani, 2015).

With advances in data-driven approaches, machine learning has been increasingly applied in transportation safety. Applications include lane-change prediction (Zhao et al., 2022), traffic speed forecasting with SHAP explanations (Lee, 2023), crash detection from speed analysis (Parsa et al., 2020), freeway crash prediction using Bayesian networks and XGBoost (Yang et al., 2022), real-time cab speed estimation (Shen, 2022), travel time prediction (Chen & Fan, 2021), and high-speed urban traffic flow modelling (Lu et al., 2023). However, limited research has focused on predicting reliable operating speeds at approach, midpoint, and exit sections of horizontal curves under mixed-traffic and pavement conditions. This gap highlights the need for models that integrate speed–friction dynamics with local roadway and behavioural factors.

Advisory speed plaques remain a standard tool for curve safety management (Traffic, 2009). Research has linked advisory speeds to crash–radius relationships (Bonneson et al., 2007; Fitzpatrick et al., 2000; Milstead et al., 2011; Montella & Imbriani, 2015) and factors such as speed differentials (Figueroa Medina & Tarko, 2007; Pratt et al., 2015), curve geometry (Dixon & Rohani, 2008), side friction (Dixon & Rohani, 2008), superelevation (Bonneson et al., 2009; Pratt et al., 2019), and signage (Hummer et al., 2010). However, static advisory speeds are often ineffective because of design inconsistencies (Nicholson, 1998), variable driver behaviour (Fleiter & Watson, 2005), attitudes (Hassen et al., 2011), and decision-making patterns (Warner & Åberg, 2006). Recent advances favour adaptive systems that continuously monitor vehicles and deliver dynamic feedback. Customized techniques create adjustable alerts (Davis et al., 2018), while automated systems provide real-time visual or audio warnings when drivers approach curves at unsafe speeds (Davis et al., 2018; Mahmud et al., 2023).

Despite international research, little work has focused on developing field-based friction models or adaptive speed frameworks for Indian highways, where mixed traffic, heterogeneous driving behaviour, and pavement surface texture variability exacerbate safety risks. To address this gap, the present study pursues four objectives:

By integrating a calibrated Fmax⁡* model, SI-based diagnostics, and detailed speed profiles, this study offers a data-driven framework for identifying hazardous curves and guiding India-specific design standards, safety audits, and real-time risk management strategies (Montella et al., 2015; Abohassan et al., 2022).

This study develops a framework for evaluating horizontal curve safety under Indian roadway conditions through four steps: (i) field data collection, (ii) development of a locally calibrated maximum side friction model (Fmax⁡*), (iii) safety index evaluation, and (iv) mid-curve speed prediction for real-time warnings.

Field surveys recorded curve geometry (radius, superelevation) and operating speeds (tangent, entry, middle, and exit) as well as pavement texture (Mean Texture Depth, MTD, via the sand patch method). These inputs were used to develop a nonlinear empirical model of Fmax, expressed as Fmax⁡*=a*MTDb*Speed−c*ed, where parameters a, b, c, and d were calibrated against AASHTO (2018) values using nonlinear regression to capture Indian conditions. The calibrated F_max values were compared with the side friction demand:

(1) Fd=V2g.R−e

The actual side friction demand values were quantified for safety assessment using three indices: the Safety Index (SI), Change in Safety Index (ΔSI), and Dynamic Curve Safety Index (DCSI). These indices enabled a systematic assessment of how safety varies from curve entry to exit. Subsequently, mid-curve operating speeds identified as critical for stability were predicted using a regression model, based on geometric and field data. This model generated advisory speed values derived from SI, supporting proactive warnings through adaptive roadside signs or in-vehicle systems.

The analytical framework integrates deterministic modelling, uncertainty quantification, and safety diagnostics. The nonlinear regression establishes the friction model, while the delta method and bootstrap quantify parameter uncertainty. First-Order Reliability Method (FORM) translates these uncertainties into reliability-based safety probabilities, and global sensitivity analysis identifies dominant influencing variables.

Five horizontal curves were selected on National Highway 340C (NH-340C) in Andhra Pradesh for detailed investigation, considering their accident history and geometric complexity. The study curves represent a broad range of radii, superelevation levels, pavement texture conditions, and crash histories commonly observed on rural two-lane highways in India. The selected sites include the Pamulapadu curve, Banumukula curve, Krishnapuram statue curve, Krishnapuram curve, and Venkatapuram curve. For each location, geometric characteristics such as chainage, radius, superelevation, roadway width, and deflection angle were surveyed to serve as input variables for the development of side friction and safety models. The radii of the selected curves ranged from 50 m to 180 m, while superelevation varied between 4.0% and 7.0%, and carriageway widths were consistently about 7.0–7.2 m. Although the curves were designed for a nominal speed of 80 km/h, field observations revealed notable deviations in actual operating speeds.

All geometric, traffic, pavement texture, and speed data were collected separately for each individual curve location. Independent field surveys were conducted at each site to capture roadway-specific operating characteristics, pavement conditions, and traffic behavior under mixed-traffic conditions.

Figure 1 presents the geographical location of the five selected study curves along NH-340C in Andhra Pradesh, India. The map illustrates the spatial distribution of the investigated sites between Pamulapadu and Venkatapuram.

The field investigations comprised measurements of spot speeds at tangent (pre-start), entry, middle, and exit locations; derivation of acceleration–deceleration profiles; measurement of Mean Texture Depth (MTD) using the sand-patch method; and detailed surveying of curve geometry. The MTD values at curve midpoints ranged from 0.380 mm to 1.123 mm, reflecting pavement surface variability. Table 1 summarizes the geometric parameters and surface conditions of the study curves.

A maximum side friction factor (Fmax⁡*) model for Indian conditions was developed using field-measured parameters, including Mean Texture Depth (MTD), midpoint operating speed, and superelevation (e). These inputs were analyzed using a nonlinear regression framework to ensure that the resulting model realistically captures pavement–tire interactions and mixed-traffic dynamics.

The general model structure was specified as:

(2) Fmax⁡*=a*MTDb*Speed−c*ed

where constants a, b, c, and d were estimated through optimization.

For each observation i, the model prediction is:

(3) F̂max⁡,i(a,b,c,d)=a×MTDib×Speedi−c×eid

Residual at row i (what the optimizer tries to drive toward zero):

(4) ri(a,b,c,d)=Fmax,i(AASHTO)−F̂max⁡,i(a,b,c,d)

Stacking residuals into a vector r(ϴ)∈ℝn, the objective minimized is the sum of squared errors (SSE):

(5) SSE(ϴ)=∑i=1nri(ϴ)2

Conceptually, (Virtanen et al., 2020) start from an initial guess ϴ⁽⁰⁾. Later, Updates and to reduce SSE using a Gauss-Newton-type step with damping (Levenberg-Marquardt) or a trust-region method when bounds are used (Levenberg, 1944; Marquardt, 1963). Finally, stops when the gradient is small or the maximum number of iterations is reached. When no bounds are imposed, SciPy historically uses a Levenberg-Marquardt approach for performing nonlinear least-squares minimization (More et al., 1980). It uses the Jacobian to numerically approximate curvature (via finite differences) and take curvature-aware steps by solving (Moré, 1978; Virtanen et al., 2020).

(6) (JT·J+λI)Δp=JT·r

where J is the Jacobian of residuals and λ is the damping; J^T: Transpose of the Jacobian; I: Identity matrix, Δp: Parameter update vector. Added to the current parameter vector in each iteration:

(7) pnew=pcurrent+Δp

Residual vector = r_i = y_i − f(x_i;p)

For each residual r_i, the solver uses sensitivities ∂ri/∂a,∂ri/∂b,∂ri/∂c,∂ri/∂d, which for the model follow from and SciPy approximates these numerically.

(8) F̂max⁡,i=a.Xi,Xi=MTDib.Speedi−c.eid

Hence,

∂F̂max⁡,i∂a=Xi,∂F̂max⁡,i∂B=aXiln⁡(MTD)i,∂F̂max⁡,i∂c=aXiln⁡(speed),∂F̂max⁡,i∂B=aXiln(ei)

differences. Conceptually, it is expected that b > 0 (greater texture yields more friction), c > 0 (larger c indicates stronger reduction of friction with increasing speed), and d may be small but positive (superelevation marginally enhances available friction). Also, takes a > 0.

To improve numerical stability, a log-linear transformation was applied for initial parameter estimation:

(9) lnFmax⁡*=lna+b,ln(MTD)−c·ln(Speed)+d·ln(e)

Let, y = ln F_max, x₁ = ln MTD, x₂ = ln Speed, x₃ = ln e

Runs an ordinary least squares on y = β0 + β.x₁ + β2.(−x₂​) + β3.x₃

Ordinary least squares on this linearized form provided initial guesses and set: a(O)=eβ0,b(0)=β1,c(0)=−β2,d(0)=β3

These values were supplied to the nonlinear solver to accelerate convergence. The parameter vector was optimized as:

(10) (ϴ̂,∑̂)=𝐜urve_fit(f,X,y)withϴ̂=(â,b̂,ĉ,d̂)

Where, ϴ̂ are the best-fit parameters that minimize the sum of squared errors (SSE), ∑̂ is the estimated covariance matrix derived from the Jacobian at the solution. From this, the standard error for parameter k was obtained as:

(11) SE(ϴ̂k)=∑̂kk

And the 95% confidence intervals were computed as,

(12) ϴ̂k±1.96SE(ϴ̂k)

The above ensures the statistical reliability of the fitted parameters. The calibrated Indian-specific predictive model was obtained as:

(13) Fmax⁡*=1.5721*MTD−0.5508*Speed−0.583*e0.033

This functional form captures the nonlinear influence of pavement surface texture, speed, and geometry on the tire-road interaction. Predicted values ranged between 0.16 and 0.28, closely matching the AASHTO reference range of 0.15–0.30 while explicitly capturing Indian roadway conditions.

The uncertainty in Fmax⁡* is evaluated by the following methods.

After developing the F_max model specific to the measured field conditions and speed profiles of vehicles under mixed traffic-flow conditions, the safety conditions at three different segments of each curve at start, middle, and end locations were evaluated by using the following three safety indices:

The safety index (SI) is defined as:

(25) SI=Fmax⁡*Fd

The above index represents the immediate safety margin between available and required friction. SI values >1.0 indicate safe operating conditions with surplus friction, SI = 1.0 denotes a critical threshold with no margin of safety, and SI<1.0 indicates unsafe operation with a risk of instability (Donnell et al., 2016).

To capture how safety evolves along the curve, the change in safety index (ΔSI) was introduced. The variation from:

Approach to midpoint:

(26) (ΔSI1=SImiddle−SIapproach)

Midpoint to exit:

(27) (ΔSI2=SIexit−SImiddle)

The above two indices distinguish deteriorating conditions (ΔSI<0) from improving conditions (ΔSI>0). Negative values indicate heightened risk zones, typically at entry or middle points, while positive values reflect recovery at exits.

The dynamic curve safety index (DCSI) is defined as the ratio:

(28) DCSI=(ΔSI1ΔSI2)=(a1a2)=(Vmiddle2−Vapproach2)(Vexit2−Vmiddle2)

The relationship quantifies curve risk balance values near 1.0 denote stability, > 1.0 indicate entry instability, and < 1.0 signal exit instability, while negative DCSI marks critical deceleration–acceleration patterns. Together, SI, ΔSI, and DCSI provide an integrated tool for assessing, tracking, and managing curve safety by combining analysis of speed, texture, and geometry.

Curve safety was assessed by linking operating speeds at approach, midpoint, and exit to side-friction demand and supply, and synthesizing these into three indicators: the Safety Index (SI), the Change in Safety Index (ΔSI), and the Dynamic Curve Safety Index (DCSI). Vehicle-class midpoint speeds formed the basis for estimating friction demand, which was compared with the calibrated Fmax⁡* model (Eq. (13)) to compute the indices.

Vehicle Speed Profiles: Spot speed surveys on five horizontal curves of NH-340C revealed marked differences across vehicle classes. Cars averaged 79.1 km/h at tangents and 61.4 km/h at midpoints, while three-wheelers were the slowest (≈24.6 km/h). Speed variability was highest for cars (SD: 16.9–18.0 km/h) and lowest for three-wheelers (7.6–8.4 km/h), reflecting heterogeneous driving behavior. Acceleration analysis showed sharp entry decelerations (up to −6.3 m/s² for cars), sustained negative accelerations at midpoints, and partial recovery at exits (+0.5 to +1.2 m/s²). These patterns indicate cautious entry, controlled mid-curve maneuvering, and incomplete exit recovery.

Validation of F_max and International Comparisons: The Indian Fmax⁡* model, derived from midpoint speeds, Mean Texture Depth (MTD), and superelevation, closely matched AASHTO trends at moderate speeds (30–60 km/h) but predicted higher values at low (<20 km/h) and high (>80 km/h) speeds, reflecting local pavement and mixed-traffic conditions (Figure 1). At low speeds (<20 km/h), friction demand becomes negligible compared with available friction. Consequently, safety indices remain well above the critical threshold (SI>1), resulting in limited differentiation between models. The proposed framework, therefore, provides greater analytical sensitivity at moderate and high operating speeds where friction demand approaches available limits. Comparisons with international models (South Africa, Austroads, Sri Lanka, Malaysia, Abu Dhabi, Korea, Malawi, and IRC) showed that the present study generally yields higher friction availability at higher speeds, highlighting the need for region-specific calibration (Figure 6).

Influence of Curve Radius on Speed and Friction: Curve radius strongly influenced both Fmax⁡* and operating speeds. The Indian model showed scattered SI and Fmax⁡* values at small radii (≈60–80 m), wide variability at medium radii (≈100–140 m), and partial stabilization at large radii (>160 m) (Figures 6 and 7). In contrast, AASHTO produced more conservative and tightly clustered predictions (Figure 6). Small-radius curves required F_max up to 0.6 under Indian conditions, exceeding the AASHTO limit of 0.4, while medium- and large-radius curves also showed higher and more variable friction demands. These results indicate that Indian drivers impose greater friction demands than international guidelines assume.

Curve Safety Assessment Using SI, ΔSI, and DCSI: Safety evaluation highlighted distinct class-specific patterns (Figures 7 and 8). SI indicated that trucks and buses were largely safe (76–80%), two-wheelers were highly stable (94.3%), while cars and three-wheelers often operated near critical limits (52.3% and 56.9% unsafe, respectively). ΔSI showed deteriorating safety toward exits, especially for three-wheelers (ΔSI₁ < 0 for 23 cases, ΔSI₂ < 0 for 113) and heavy vehicles (ΔSI₂ < 0 for 229). DCSI revealed entry–exit imbalances, with all cars and two-wheelers in high-risk zones (DCSI < 0) and severe instability among three-wheelers and heavy vehicles. Collectively, these indices form a comprehensive framework that integrates speed, friction, pavement texture, and geometry to identify hazardous curves and guide targeted interventions.

Implications for Highway Safety: The analysis demonstrates three critical points: (i) side friction demand decreases as speed increases, (ii) curve radius strongly moderates the speed–friction relationship, and (iii) international models often underestimate friction requirements under Indian conditions. Adoption of a region-specific model of Fmax⁡* improves the accuracy of safety assessments and supports interventions such as speed regulation, curve realignment, and pavement treatments, thereby enhancing safety in mixed-traffic environments.

Generalizability of the proposed model: Although the present study uses five curves from NH-340C, the selected curves cover a broad range of geometric and pavement conditions typical of Indian rural highways. The observed radii range between 50 m and 180 m, superelevation between 4 and 7%, and Mean Texture Depth between 0.38 and 1.12 mm. These ranges represent typical design parameters recommended in IRC standards for two-lane rural highways. Therefore, the calibrated friction model reflects realistic operating conditions under heterogeneous traffic environments. Future research will extend the dataset to multiple highways across different climatic regions of India.

Field-observed speed profiles reveal a consistent deceleration trend from the curve approach to its midpoint, followed by partial acceleration toward the exit. The mid-curve represents the most critical control point, where three-wheelers record the lowest operating speeds. At the same time, cars exhibit greater variability and more abrupt acceleration patterns, leading to higher differential speeds among vehicle classes. Such fluctuations highlight the necessity of real-time driver advisory systems to maintain safe, uniform speed behavior on horizontal curves.

To quantify these dynamics, a regression-based model was developed to predict mid-curve operating speeds (V_middle) under field conditions, using measured approach speeds (V_start) and geometric parameters. The calibrated model demonstrates a strong linear relationship, expressed as:

(29) Vmiddle=0.9821Vstart−4.271

Predicted and observed speeds showed close agreement (Figure 9). From training results, the statistical metrics found as Mean Absolute Percentage Error (MAPE) = 2.72%, Root Mean Square Error (RMSE) = 1.37 km/h, and Coefficient of Determination (R²) = 0.975, while testing results yielded MAPE = 2.96%, RMSE = 1.68 km/h, and R² = 0.962. These metrics confirm the model’s reliability and minimal error across datasets.

These results confirm that the model accurately captures the link between approach and midpoint speeds, making it suitable for real-time curve-speed estimation. Trained with field data, it can generate safe speed alerts to improve vehicle stability and reduce excessive friction demand. Such models support safety measures such as dynamic speed display signs (Monsere et al., 2005) and audible warnings (Xu et al., 2023), helping drivers adjust their behavior and reducing crash risk

Advisory speed plaques are widely used for curve safety management (Traffic, 2009). Past studies linked advisory speeds to factors such as curve radius, geometry, side friction, superelevation, and warning signs (Bonneson et al., 2007; Dixon & Rohani, 2008; Montella & Imbriani, 2015; Pratt et al., 2019). However, static advisory speeds often prove inadequate due to inconsistent design, varying driver behavior, and differences in decision-making (Nicholson, 1998; Soole et al., 2013). Recent approaches emphasize adaptive systems that monitor real-time speeds and provide automated warnings through visual or audio feedback (Davis et al., 2018; Mahmud et al., 2023).

The safety indices (SI, ΔSI, DCSI) proposed in this study offer a foundation for dynamic curve monitoring. Speed data, obtained via roadside sensors or GPS, can be used to calculate SI and identify unsafe conditions. Risk alerts may then be issued through vehicle-actuated flashing beacons, adaptive electronic signage, or integrated ADAS systems (Antony & Whenish, 2021). This approach supports immediate feedback, particularly under adverse conditions, and allows adjustments by vehicle type and site-specific characteristics.

This study developed a field-based framework for evaluating horizontal curve safety by integrating measured traffic speeds, pavement surface characteristics, and probabilistic reliability analysis. The framework was applied to five horizontal curves located along NH-340C to examine the relationship between operating speeds, pavement texture, and available side friction. The major findings are summarized as follows:

Overall, the results demonstrate that integrating friction modelling, safety indices, and speed prediction provides a useful analytical approach for evaluating curve safety at the site level and may support future development of adaptive warning systems.

This study presents a field-based framework for assessing horizontal curve safety by relating vehicle operating speeds, pavement texture, and available side friction. However, several limitations should be acknowledged. The analysis was based on observations from five horizontal curves located along NH-340C. Although the selected curves represent a range of geometric and pavement conditions typical of rural highways, the limited number of study sites restricts the extent to which the results can be generalized beyond the investigated corridor. In addition, all field measurements were conducted under dry pavement conditions in order to isolate the influence of geometry and surface texture on friction availability.

In real-world conditions, environmental factors such as rainfall, pavement wetness, and drainage characteristics can significantly affect tire-pavement interaction and reduce available friction levels. Future studies should therefore incorporate weather-related variables, including rainfall intensity and pavement surface moisture, possibly through continuous monitoring systems or sensor-based measurements. Expanding the dataset to include additional curves from other highway corridors and varying traffic compositions would also improve the robustness and broader applicability of the proposed framework. Further work may also investigate the integration of real-time traffic data and intelligent transportation technologies to support dynamic speed advisory systems for hazardous curves.

R. Srinivasa Kumar: Conceptualization, Methodology, Software, Supervision, Writing – original draft, Writing – review & editing. Vemula Rajesh Kumar: Data curation, Formal analysis, Investigation.

The authors are thankful to the Department of Civil Engineering for continuous support and the Traffic Police Department for permission and assistance during field data collection.

The authors have no conflicts of interest with other entities or researchers.

The authors declare that no generative AI or AI-assisted tools were used in the preparation of this manuscript.

The research reported in this paper involved field observations of vehicle speeds and roadway characteristics on public highways. No personal or identifiable data were collected from individuals. Therefore, the study did not involve human subjects and did not require formal ethical approval.

No agency was involved in funding this study.

Data is available upon request.

Handling editor: Jiří Ambros, Transport Research Centre (CDV), Czechia.

Reviewer: Alyssa Ryan, Michigan State University, the United States of America.

Submitted: 13 March 2026; Accepted: 2 June 2026; Published: 13 June 2026.