From May to July 2012, ten experimental plots arranged across two blocks (ca. 100 m apart from each other) were established in each of the eight forests. Each block consisted of five 50 x 50-cm plots that were fenced with metal flashing buried 10 cm into the soil to minimize horizontal migrations of animals. To prevent macrofauna from recolonizing or leaving the plots after our manipulations, a 1.2 mm-mesh fiberglass net was located on top and another at the bottom (see Appendix S1 and Appendix S3: Fig. S1). In total 80 experimental plots were established across the eight beech forests. Due to unanticipated events during the experiment, including some replicates being destroyed by bears, we ended up with 62 plots from seven forests for the analysis of final responses (see Appendix S1 for details). The average MAP for the three dry forests was 1095 mm compared with 1472 mm for the four wet forests (extracted from the digital atlas of the Iberian Peninsula Ninyerola, Pons, & Roure, 2005, with average precipitation values based on at least twenty years measures since 1950). Thus, dry forests were ~ 25% drier than wet forests, a difference that is slightly above the predictions of rainfall reduction for the end of the 21st century in southern Europe (Stocker et al., 2013).
	We conducted a factorial randomized-block experiment. Within each block, two plots were randomly assigned to a 'Predator removal' treatment (0X), where all large predators were removed (see details below); another two plots were assigned to a 'Predator addition' treatment (2X), where the density of large predators was doubled, and one 'Control' plot (1X), which included natural densities of large predators (see Appendix S1 for details on predator manipulation). Predator manipulations based on natural densities present at each site are more realistic than imposing an arbitrary set of different densities across all sites, since each site may have a characteristic equilibrium density of top predators. Thus, the “Predator addition” treatment can be interpreted as a doubling of the equilibrium density reflective of local conditions at the time of the manipulation. 
	Before installing the fenced plots, we carefully collected five 28-L leaf-litter samples from each block (by filling, with fallen beech leaves, identical plastic containers of 28-L volume, one from each of the five locations where an experimental plot would be established) and brought them to the laboratory. These five samples were the basic units from which we sorted and counted all the meso- and macrofauna within blocks. To minimize damage to the animals, we did not sift the litter, but instead carefully scattered all litter in a tray and closely examined every leaf to separate the large predators from the rest of the fauna. All specimens were identified to the lowest taxonomic affiliation possible (Appendix S1), and their length was measured to the nearest 0.5 mm. Large (apex) predators were kept in separate vials until their reassignment to the experimental plots, while the leaf litter and remaining animals were released into their respective container. Within each block all the large predators were randomly assigned to treatments and released in the containers, such that 2/5 were placed in each of the two 'Predator addition' plots (double density – 2X), 1/5 in the control plot (natural density – 1X), and none in 'Predator removal' plots (0X). Thus, 2X plots had two times more apex predators  than 1X plots, being the natural densities block-specific. The containers with the litter and fauna were then returned to the field and each was emptied into its corresponding fenced plot. 
	In autumn of 2012, the experimental manipulations were reinforced by removing all the large predators found in the 'Predator removal' plots and randomly transferring them to each of the two 'Predator addition' plots (Appendix S1). At the end of the experiment (April to June 2013) all animals found in the plots were counted, measured and identified to the same taxonomic level as when setting up the experiment.

To study litter decomposition, three 10 x 10-cm litterbags each containing 1 g of hazel (Corylus avellana) leaf litter were placed in each plot (see details in Appendix S1). We used hazel leaves instead of beech leaves, as hazel trees occur in or near all our forests as understory trees and its leaves decompose at a much faster rate than beech leaves (Sanpera-Calbet, Lecerf, & Chauvet, 2009), thus ensuring sufficient decomposition during the course of the experiment to enable assessment of the experimental manipulations on decomposition rate. 
One bag was collected after 6 months, and after 12 months the remaining two bags were collected to increase the precision of the estimate. The bags were dried at 60 ºC for 24 h and the remaining litter weighed to the nearest 0.01 g. 
	We estimated the decomposition constant (k) for each plot by first fitting the following equation (Olson, 1963) through the three collected bags: 
		Mt = M0 e–kt 
with Mt the mass of litter at time t, M0 the initial mass of litter, k the decomposition constant and t the amount of time elapsed since the initial measurement. We fitted a least-squares line to the model [–log (Mt / M0) = kt] and then used the fitted k to characterize the rate of decomposition in each plot. 

All analyses were performed using the statistical software R 3.4.4 (R Core Team, 2018).
Decomposition rates were analyzed with a mixed-effects linear model ("lmer" function of the "lme4" package) with a square root transformation.
Abundances of trophic and taxonomic groups were analyzed with Generalized Linear Mixed Models (GLMM) using the "glmer.nb" function in the "lme4" package of R