The majority of the data analysed here was collected on an expedition to Montagne d’Ambre carried out from November 2017 to January 2018 by Mark D. Scherz, Andolalao Rakotoarison, Safidy M. Rasolonjatovo, Jary H. Razafindraibe, Ricky T. Rakotonindrina, Onja Randraimalala, and Angeluc Razafimanantsoa; other results from this expedition have been reported elsewhere (Rasolonjatovo et al. 2020, 2022; Rakotoarison et al. 2022; Scherz et al. 2023). The expedition collected numerous Stumpffia specimens and tissue samples from across the mountain, including material from the herpetologically unexplored west slope of the mountain. Other samples included here were collected over numerous previous visits to the mountain over the past 25 years by various researchers.

Specimens were captured on surveys consisting of opportunistic and targeted searching by day and night (with the aid of hand-held and head torches), often guided by the calling activity of male specimens. Collected specimens were photographed in life, and voucher specimens were anaesthetised and subsequently euthanised using tricaine mesylate solution (MS222). A tissue sample of each individual was taken from limbs (either a muscle sample or, in small individuals, a whole part of the limb) and stored in 99% ethanol for molecular analysis. Voucher specimens were subsequently fixed in 90% ethanol, before permanent conservation in 70% ethanol. Field number labels were tied to specimens (or, in the case of diminutive individuals, affixed to Eppendorf tubes into which the frogs were placed). Field numbers used herein include ACZCV Angelica Crottini, DRV David R. Vieites, FAZC Franco Andreone, FGMV Frank Glaw and Miguel Vences, FGZC Frank Glaw, MSZC and MSTIS Mark D. Scherz, RAX Christopher J. Raxworthy, RJS Jasmin E. Randrianirina, and ZCMV and MV Miguel Vences. Specimens were deposited at the ZSM Zoologische Staatssammlung München, Munich, Germany, UADBA collections of the Mention Zoologie et Biodiversité Animale, Université d’Antananarivo, Antananarivo, Madagascar, and the NHMD/ZMUC Natural History Museum Denmark, Copenhagen, Denmark.

Montagne d’Ambre mostly consists of basalt, basanite, ankaratrite, covered by acidic Haplic Acrisols, Haplic Ferrasols and Haplic Cambisols (Goodman et al. 2018). It is entirely surrounded by sandstone, marl and shaley limestone (Donné et al. 2021). Despite plant growth limitation by high levels of aluminium and iron in the soil, the area is mostly covered by moist, evergreen forest with a canopy height of 20 to 25 m, whilst forests closer to the summit display a canopy usually below five metres (Goodman et al. 2018). Additionally, there is rather dry vegetation (i.e., dry, deciduous forests) in a series of craters west of Lac Maudit, and along the west slope, largely due to porous underlying volcanic rock (Goodman et al. 2018). Montagne d’Ambre also contains several lakes, e.g., Lac Maudit and Grand Lac, which constitute noteworthy study sites for anuran ecology (Rasolonjatovo et al. 2022). The local climate is dominated by the North-Malagasy subhumid climate, receiving annual average rainfall levels of about 1424 mm (in the years 1981–2017), the vast majority of which occurs between November and April as well as temperatures between 12.3°C (cold season: June–August) and 30.4°C, although temperatures decrease at the upper elevations (Goodman et al. 2018). Also, vegetation and the local climate (and thus, the fieldwork) are influenced by the regular occurrence of cyclones (Goodman et al. 2018; Rasolonjatovo et al. 2020, 2022). Fieldwork in 2017–2018 was carried out along an elevational transect, with a satellite camp established on the West Slope (see Scherz et al. 2023 for further details). However, sampling gaps remain especially at the south slopes, the far east slopes, and the lowest elevations of the north slope.

We obtained DNA sequences from the 3’ and 5’ fragments of the mitochondrial rRNA gene 16S and the nuclear-encoded Recombination Activation Gene (Rag1) of 45 specimens assigned to six nominal species in the course of this study. For this purpose, genomic DNA was extracted from muscle tissue or digits stored in 99% ethanol using either the standard salt extraction protocol by Bruford et al. (1992), or an SPRI Bead DNA Extraction Protocol (Phyletica Lab 2024) using Proteinase K (20 mg/ml) for digestion. Amplification via standard polymerase chain reaction (PCR) was performed according to protocols of Belluardo et al. (2022), using primers and PCR conditions as stated in Table SS1 to a final volume of 25 µl per sample. PCR products were then purified by centrifugation through Sephadex G-50 Superfine (Cytiva, 17-0041-01) incubated in HPLC-H₂O (Merck, 1.15333.2500) in MultiScreen-HV plates (Millipore, 0.45 µm, MAHVN4510) according to the manufacturer’s instructions, for five minutes at 910 × g. A volume of 5 µl of the resulting elute was added to each 2 µl 5× green GoTaq buffer (Promega GmbH), 0.5 µl Big Dye Terminator (v.3.1, ThermoFisher Scientific) 0.5 µl primer and 5.5 µl HPLC-H₂O (Merck, 1.15333.2500) to a total volume of 10 µl. Subsequently, 5 µl of this sequencing reaction was added to 5 µl HiDi Formamide (ThermoFisher Scientific, 4311320) on the Sanger sequencer (3500 Genetic Analyzer, ThermoFisher Scientific). The resulting forward and reverse sequence chromatograms were checked by eye and combined into a single contig for each sample and each marker. Sequences with low confidence due to deficient chromatogram quality were discarded. A further 117 sequences of the 3’ 16S fragments were generated using the Illumina amplicon approach (Vences et al. 2016). Published 16S and Rag1 sequences were downloaded from GenBank and included in the matrix in order to complete our sampling of Stumpffia, and we also incorporated a number of previously unpublished 16S and Rag1 sequences generated in the labs of A. Crottini and MV over the last fifteen years. We uploaded all novel sequences to GenBank (accession numbers: PX527161–PX527194 and PX600750–PX600957; Table S2).

We produced four alignments with MAFFT online (Katoh et al. 2019) from newly generated and available sequences. All alignments were quality checked by eye with subsequent manual adjustments before testing for the optimal model for molecular evolution with the software jModelTest 2 version 2.1.10 (Darriba et al. 2012). GTR+I+G was determined as the optimal model for each individual alignment. Anilany helenae (Vallan, 2000) served as outgroup, since Anilany is phylogenetically sister to the genus Stumpffia (Scherz et al. 2016; Tu et al. 2018).

· Alignment A (length = 684 nt), consisting of 3’ 16S sequences only, for a total of 315 Stumpffia specimens as well as a sequence from Anilany helenae.

· Alignment B (length = 680 nt), consisting of 5’ 16S sequences only, for a total of 348 Stumpffia specimens as well as a sequence from Anilany helenae.

· Alignment C (length = 1238 nt), consisting of manually concatenated 3’ and 5’ 16S sequences, for a total of 59 Stumpffia specimens, aligned against a full 16S sequence from Anilany helenae (GenBank accession number MZ751042.1).

· Alignment D (length = 531 nt), consisting of Rag1 sequences only, for a total of 239 Stumpffia specimens as well as a sequence from Anilany helenae. Note that there were no sequences available for the Montagne d’Ambre endemic species S. bishopi.

For all four alignments, maximum likelihood trees were calculated using the software raxmlGUI version 2.0.7 (Edler et al. 2021), setting ‘Analysis’ to ‘ML + rapid bootstrap’ and ‘Reps.’ to 1000. For Alignment C, an additional Bayesian tree was calculated using MrBayes version 3.2.7 (Ronquist et al. 2012) using Markov chain Monte Carlo algorithm (MCMC), four chains, two runs and 2,000,000 generations. Trees for Alignments A, B, and C were displayed in FigTree version 1.4.4 (Rambaut 2018). Within the genus Stumpffia, four major clades are evident sensu Rakotoarison et al. (2017), clades A–D, which help to inform biogeographic analyses of the genus. All species of Stumpffia were included in our analyses, but only members of the A clade (sensu Rakotoarison et al. 2017) are depicted in figures for space reasons, as the other clades do not occur on Montagne d’Ambre. Non-focal clades are also collapsed in our figures. Full trees are given in the supplementary material (Figs S1, S2). Full mean pairwise uncorrected p-distances were calculated based on the Alignment A, B, and C using the TaxI2 tool in iTaxoTools version 0.1 (Vences et al. 2021b).

A reduced version of Alignment D (length = 323 bp), containing no missing data and only including individuals of Clade A sensu Rakotoarison et al. (2017), was used to visualize the relationship among alleles (haplotypes) of Rag1 using a network approach in the program Hapsolutely (Vences et al. 2024). Haplotypes were inferred using the PHASE algorithm (Stephens et al. 2001), using a phase (-p) and allele (-q) threshold of 0.5 with 1000 MCMC iterations. Based on the phased sequences, a network was constructed under the Median-Joining algorithm (Bandelt et al. 1999) and edited in Adobe Illustrator v29.4 (Adobe Inc., San Jose, CA, USA). The input files for Hapsolutely, along with the datasets for other analyses, are available from a dataset uploaded at the Zenodo repository at https://doi.org/10.5281/zenodo.16385861.

We analysed advertisement calls of various Stumpffia specimens from Montagne d’Ambre recorded over the last 20 years. Call series of different lengths were recorded from Stumpffia angeluci (MSZC 0531), Stumpffia bishopi (MSZC 0730), Stumpffia huwei (MSZC 0405, MSZC 0660, ZCMV 13618, ZCMV 13619, MSZC 0744, two series of MSZC 0769), Stumpffia madagascariensis (MSZC 0707, ZCMV 12185), Stumpffia maledicta (MSZC 0666, MSZC 0724, ZCMV 13504), Stumpffia mamitika (MSZC 0793), and Stumpffia megsoni (MSZC 0545, MSZC 0764, MSZC 0765, MSZC 0768, MSZC 0777). Calls from other locations were recorded for S. angeluci (ZCMV 13608, Joffreville) and S. mamitika (ZCMV 13616, Ankarana; one not-collected specimen, Andapa). Temperature data was unavailable for many recordings and is, thus, rarely given here. Calls are deposited in the Zenodo repository at https://doi.org/10.5281/zenodo.16385861. Note that recordings from ZCMV specimens were made with a damaged field recorder, which however mostly affected their amplitude profiles only, and does not appear to have had detrimental effects on the fidelity of call frequencies.

Recordings were resampled at 22.05 kilohertz (kHz) and analysed in Audacity (Audacity Team 2014). Call duration was calculated by identifying the start and end of calls in waveform view. Duration of the inter-call interval (from the end of one call to the start of the next) was calculated subsequently. Dominant frequencies were obtained from plotted spectra of single calls. Results were displayed as ‘minimum – maximum (mean ± standard deviation)’ in milliseconds (ms) or respectively in Hertz (Hz), according to Rakotoarison et al. (2017). We define a call of Stumpffia herein as a single, high-pitched tonal note (Rakotoarison et al. 2017) and use the terminology for call descriptions by Köhler et al. (2017), using the call-centered terminological scheme. Inter-call intervals varied quite substantially even within a call series, raising the question of whether well-motivated calling specimens were recorded and thus, do not seem to be a suitable character for species identification in this assemblage. Plots of call duration vs. dominant frequency were created using the ggplot2 package (Wickham 2016) in R (R Core Team 2024) with RStudio (RStudio Team 2019).

Basic morphometric measurements were taken on a total of 64 specimens from seven nominal species and one potential candidate species: Stumpffia angeluci (N = 13), Stumpffia bishopi (N = 3), Stumpffia huwei (N = 14), Stumpffia madagascariensis (N = 4), Stumpffia maledicta (N = 12), Stumpffia mamitika (N = 9), Stumpffia megsoni (N = 6), and Stumpffia sp. aff. angeluci (N = 3). All specimens were collected on Montagne d’Ambre, except for three individuals of S. mamitika (ZSM 3237/2012, ZSM 0307/2004, ZSM 0228/2016) and five specimens of S. angeluci (ZSM 300–303/2004, ZSM 1671/2008).

Morphological measurements followed Rakotoarison et al. (2017), and were the following: SVL snout–vent length, HW maximum head width, HL head length, TD horizontal tympanum diameter, ED horizontal eye diameter, END eye–nostril distance, NSD nostril–snout distance, NND nostril–nostril distance, FORL forelimb length, HAL hand length, HIL hindlimb length, measured straightened, FOTL foot length including tarsus, FOL foot length without tarsus, and TIBL tibia length. Additionally, UAL upper arm length, LAL lower arm length, THIL thigh length, and TARL tarsus length were taken. We did not examine the relative position of the tibiotarsal articulation position when stretched forward along the body to avoid damaging specimens. A measurement scheme is given in Figure 1. All measurements were performed using digital callipers (Mitutoyo Corp., CD-15DAX, accuracy ± 0.02 mm; Mitutoyo Corp., CD-15CP, accuracy ± 0.03 mm; Hogetex, Digital Caliper, 6M05.3.42, accuracy ± 0.03 mm). Measurements were subsequently compared to previous measurements from Rakotoarison et al. (2017, 2022), but to avoid inconsistencies due to measurer effect, only newly obtained measurements obtained by a single measurer were retained for analyses.

Morphometrics were used to perform a principal component analysis (PCA) in RStudio (RStudio Team 2019) with R (R Core Team 2024), using the ggplot2 (Wickham 2016) and tidyverse (Wickham et al. 2019) packages. To avoid losing specimens due to missing data, three hind limb measurements were imputed for four specimens of Stumpffia huwei from the west slope (MSTIS 00930, 00931, 00964, and 00967) by linear regression against SVL. One specimen, ZSM 0228/2016 (ZCMV 13616), was omitted due to a missing measurement of FORL. PCA was performed for the following traits: log₁₀(SVL), HW, HL, TD, ED, END, NSD, NND, FORL, HAL, HIL, FOTL, FOL, TIBL, UAL, LAL, THIL, TARL. Values other than SVL were corrected for relative size by taking their residuals from a linear model against SVL.

A total of 91 Stumpffia specimens and tissue samples were collected on Montagne d’Ambre (Fig. 2) between October 2017 and January 2018, including nine call vouchers (i.e., males observed and recorded calling). Sequences of the mitochondrial 16S rRNA markers we analysed, successfully assigned all sequenced individuals to species level (Figs 3S1–S4), and confirmed the presence of the following species: Stumpffia angeluci (seven specimens), S. bishopi (three specimens), S. huwei (22 specimens), S. madagascariensis (14 specimens), S. maledicta (30 specimens), S. mamitika (one specimen), S. megsoni (13 specimens), S. sp. aff. angeluci (three specimens) (Figs 3, 4, S4). Stumpffia sp. aff. angeluci represents a previously unknown lineage, and Stumpffia mamitika is here reported from Montagne d’Ambre for the first time.

We recovered remarkably high infraspecific variation in Stumpffia huwei, notably between the populations from the west and north slopes; they differ by uncorrected pairwise distances (p-distances) of 1.68–2.09% in the 5’ 16S fragment of Alignment B (Table S3). However, two specimens from near the Gîte d’Étape (ZCMV 3996 collected in 2009 and MSZC 0643) on the north slope also differed by 1.17–2.17% in the 5’ 16S fragment from other individuals from the same locality (Table S3). Nevertheless, uncorrected p-distances from the 3’ 16S fragment (Alignment A) that is more frequently used for species delimitation in Malagasy anurans (Vieites et al. 2009) are only high between the west and north slope specimens (up to 1.65%) but do not differ strongly between specimens of the same location. These west slope specimens are clearly differentiated from other S. huwei in the phylogenetic trees calculated, with strong support (Fig. 3; Table S4).

Concurrently with their genetic differentiation, west slope specimens of S. huwei show differences in bioacoustics and morphometrics compared to those from the north slope (Fig. 5; Table S5). Dominant call frequency shows zero overlap in the measured parameters between west slope specimens and other specimens of S. huwei (5120–5515 Hz from west slope specimens MSZC 0744 and MSZC 0769 vs. 4615–5057 Hz from north slope specimens; Table 1), despite being highly similar in body size on average (Fig. 5). However, call duration and inter-call intervals are not substantially different, partly due to high standard deviations within call series of single individuals. Morphometrically, west slope specimens and other members of S. huwei differ in only few traits (Table S5): West slope specimens have on average longer upper arms (UAL) without overlapping values with other specimens of S. huwei and, thus, longer forelimbs (FORL), with or without scaling to body size (SVL). Although there is substantial overlap in PCs 1 and 2 of our morphometric PCA (Fig. 6), they have almost no overlap in PC3, which is weighted most strongly by ED, THIL, UAL, and FORL (Fig. S5).

The specimen of S. mamitika found on Montagne d’Ambre (ZSM 116/2018, field number MSZC 0793) also differs in bioacoustics from conspecifics from other locations (Fig. 7), without overlap in dominant frequency (5369–5781 Hz in ZSM 116/2018 vs. 4417–4626 Hz in other calling males of S. mamitika; Table 1). The spectrograms reveal that ZSM 116/2018 displays not just a harmonic of the other specimens’ dominant frequencies (which seems to be the case for the S. madagascariensis specimens; Fig. 8) but emits calls with a different fundamental frequency (Figs 7, 8). As in S. huwei, call duration and inter-call intervals are not substantially different, partly due to high standard deviations within call series of single individuals. Morphometrically, ZSM 116/2018 differs only very slightly in tympanum diameter (TD/SVL) and eye–nostril distance (END/SVL) from other S. mamitika but displays otherwise values within the range for the species in the remaining traits (Table S5). It does, however, differ strongly from conspecifics in the 5’ 16S sequences by unpaired p-distances of 0.51–2.95; it was not sequenced for the 3’ 16S fragment (Table S3).

Phylogenetic analysis revealed a highly divergent lineage containing MSZC 0436, MSZC 0710, and ZCMV 13048, which we now tentatively label as Stumpffia sp. aff. angeluci (Figs 3S4). All three specimens are from the vicinity of Lac Maudit. They are closest to Stumpffia angeluci, with p-distances of 3.00–3.19% in the 3’ 16S sequences (Table S4) and 3.05–3.86% to them in the 5’ 16S sequences (Table S3). In the short Illumina-sequenced 3’ 16S fragment, they exhibit lower p-distances towards Stumpffia maledicta (1.50% to all), but the longer 5’ fragment contains more informative sites. They overlap in morphospace with S. angeluci (Figs 6, S5), but a more detailed morphological analysis may reveal further differences; these specimens warrant recognition as an unconfirmed candidate species (Vieites et al. 2009).

In the phylogenetic tree calculated based on Alignment C (3’+5’ 16S; Figs 9, S3), some of the Montagne d’Ambre species together form a monophyletic group, containing S. maledicta, S. huwei, S. mamitika, S. angeluci, and S. sp. aff. angeluci. We refer to this clade herein as the ‘Ambre-endemic clade’ (AE-clade); S. maledicta, S. huwei, and S. sp. aff. angeluci are, according to current knowledge, microendemic to Montagne d’Ambre, whereas S. mamitika is also found in other parts of northern Madagascar as much as 240 km away (Vohemar, Ankarana, Andapa) and S. angeluci has been recorded on Montagne des Français, 35 km away (Rakotoarison et al. 2017). This latter species also occurs at the lowest elevations on Montagne d’Ambre (Fig. 4H). In addition to this clade, the phylogenetically distant Stumpffia madagascariensis and S. bishopi are also microendemic to Montagne d’Ambre. The diminutive Stumpffia madagascariensis (9.7–11.9 mm SVL; Rakotoarison et al. 2017; note, fig. 26a, b in that paper shows a specimen incorrectly assigned to this species that is substantially larger; it is probably S. angeluci) is sister to the candidate species S. sp. Ca25 from Montagne des Français, about which very little is known. Stumpffia bishopi and S. megsoni, the latter also occurring in dry forests further north, form a clade with S. be Köhler et al., 2010, S. hara Köhler et al., 2010, and S. staffordi Köhler et al., 2010 (Figs 3, 9, S4). Whether S. bishopi or S. staffordi is the sister taxon of the remaining lineages in this clade differs between the 16S 3’ tree (Fig. 3) and both the 16S 5’ tree (Figs S1, S4) and the concatenated tree based on Alignment C (Figs 9, S3). However, since S. bishopi is a lineage diverging from a very basal node whilst S. megsoni is a later-diverging lineage, they most likely have different origins on the mountain.

Montagne d’Ambre Stumpffia species show multiple cases of haplotype sharing in the nuclear Rag1 marker (Fig. 10): Stumpffia huwei and S. angeluci; S. maledicta and S. angeluci; S. huwei and S. maledicta; S. mamitika and S. sorata; and S. mamitika and S. gimmeli. Stumpffia gimmeli and S. sorata are not found on Montagne d’Ambre, but both species are rather widespread and contain significant intraspecific genetic diversity (Rakotoarison et al. 2017, 2019b). Stumpffia megsoni and S. madagascariensis are each separated from their genetically closest species by multiple mutation steps and show no signs of haplotype sharing.

Montagne d’Ambre specimens contain medium to large-sized Stumpffia, ranging in body size (SVL) from 9.7 mm in S. madagascariensis to 24.3 mm in S. megsoni (Figs 4, 6). Stumpffia angeluci seems to be larger on average on Montagne d’Ambre compared to the other specimens of the same species that were collected on Montagne des Français (Table S6). Morphometrically, all species are rather similar, and only few species lack overlap in a plot of PC1 and PC2 of our size-corrected PCA (Fig. 6): Stumpffia madagascariensis is the most distinct, overlapping in morphospace only with S. huwei. Stumpffia sp. aff. angeluci is also moderately distinct, overlapping in morphospace only with S. angeluci and S. maledicta. Although S. megsoni is much larger than all other species, this is evidently not associated with shape changes beyond standard allometry, as it overlaps in morphospace with most other taxa. Body size differs to only a limited extent among species, with S. megsoni being much larger, and S. madagascariensis much smaller, than all other species (Fig. 6). Stumpffia angeluci shows particularly great variation in body size, from 10.8 to 18.4 mm; we cannot discount that some of the smallest measured specimens may be juveniles or subadults. While qualitative characters were not measured in this study, it is noteworthy that slightly expanded finger discs are only known from S. megsoni (Köhler et al., 2010).

We analysed calls of 23 individuals belonging to seven species (Table 1; only S. sp. aff. angeluci was not recorded calling), which consist of single short tonal notes resembling a high-pitched ‘beep’, and are repeated in a call series of variable duration. Intraspecific variation was moderate, except in inter-call intervals, which varied substantially even intra-individually (i.e., within single call series). Dominant frequencies showed only little variation within a call series, but in some cases varied substantially within a species (Table 1), e.g., in S. huwei (4615–5515 Hz). Most extreme are the differences in dominant frequencies in the two analysed calls of S. madagascariensis (3994 Hz for ZCMV 12185 vs. 5901 Hz for MSZC 0707 on average; Fig. 8H), but this appears to be due to the dominant frequency being on a different harmonic of calls with a shared fundamental frequency around 2000 Hz (harmonics of the call of MSZC 0707 can be seen in Fig. 8C), which might be linked to different recording distances (Köhler et al. 2017). Frequency modulation is either absent or little expressed in Montagne d’Ambre Stumpffia (Fig. 8A–G). No frequency modulation was found in S. madagascariensis, S. bishopi, and S. huwei; slight upward modulation was present in Montagne d’Ambre S. mamitika (absent in Ankarana specimen, Fig. 7), S. maledicta, and S. angeluci; slight downward (or no) modulation was detected in S. megsoni.

The community of Stumpffia on Montagne d’Ambre—comprising a mixture of a micro-endemically radiated species, micro-endemic species belonging to more widespread clades, and at least one more-widespread species—exhibits strong call differentiation among species (Fig. 8; Table 1), especially among species co-occurring at the same elevation. Different species occurring at the same elevation emit calls that differ drastically interspecifically (compare Fig. 4H to Fig. 8H). For example, Stumpffia mamitika, S. megsoni, and S. angeluci are all found (though not exclusively) at 785 m a.s.l. and emit calls with dominant frequencies of 4428–5781 Hz (S. mamitika), 3098–3523 Hz (S. megsoni) and 4084–4551 Hz (S. angeluci). The small overlap between S. mamitika and S. angeluci is actually due to inclusion of specimens of S. mamitika from Ankarana; the only specimen of S. mamitika recorded on Montagne d’Ambre (MSZC 0793) called at dominant frequencies of 5369–5781 Hz, i.e., >800 Hz higher than the maximum dominant frequency recorded from S. angeluci. More importantly, call duration ranges of those three species also do not overlap, which are—in contrast to frequency—independent of body size (Köhler et al. 2017).

Calls of Montagne d’Ambre Stumpffia are so distinct that we have been able to provide a dichotomous key to their identification (Appendix). Note that this key only applies to reasonably motivated individuals calling on Montagne d’Ambre and may else fail.