Post-glacial hunter-gatherer subsistence patterns in Britain: dietary reconstruction using FRUITS

The diets of 85 individuals from 21 sites were modelled using FRUITS based on their bulk bone collagen C and N isotope ratio signatures. The sites, which occur in a range of environments, group into three distinct periods corresponding to the British ‘Late Upper Palaeolithic’, ‘Early Mesolithic’ and ‘Late Mesolithic’, respectively. The FRUITS models for three LUP sites dated to the Bølling–Allerød Interstadial suggest an emphasis on terrestrial (animal and plant) resources. The FRUITS predictions for the Early and Late Mesolithic suggest there was significant variability in diet between sites and occasionally between individuals from the same site. The Late Mesolithic coastal site of Cnoc Coig in western Scotland shows the expected emphasis on marine resources. In contrast, Early and Late Mesolithic coastal sites in South Wales show greater reliance on terrestrial food sources. In several cases, our model predictions differ from the interpretations of previous authors. A surprising outcome is the lack of evidence for the consumption of freshwater resources at sites near large rivers. We add the caveat that our model predictions are likely influenced by inadequate baseline δ13C and δ15N data for wild terrestrial plant and aquatic resources, in particular.


Introduction
After more than a century of archaeological investigation, knowledge of the lifeways of post-glacial hunter-gatherers in Britain is still extremely limited. In large part, this reflects the generally poor preservation of food remains in open-air archaeological sites (especially animal bones) due to adverse soil conditions, as well as the inundation of coastal areas by relative sea-level rise since the Last Glacial Maximum-the effects of which have been exacerbated by inadequate archaeological recovery techniques and regional research biases.
In consequence, knowledge of post-glacial hunter-gatherer subsistence patterns relies heavily on stable isotope analysis of human remains found mainly in the relatively protected environment of caves. Here, we present a synthesis of the bone collagen stable isotope data for British hunter-gatherer populations, evaluate the robustness of conventional approaches to interpreting stable isotope data and offer new interpretations of hunter-gatherer diet using Bayesian mixing models.

Palaeodiet and stable isotope analysis
Stable isotope analysis of bone collagen is a long-established tool for reconstructing past diet (see Schoeninger 2010 for a review). The principles of stable isotope interpretation were established in the 1970s and 1980s (Vogel and van der Merwe 1977;Schoeninger and DeNiro 1984). Carbon (δ 13 C) and nitrogen (δ 15 N) isotope ratios in the tissues of humans and other mammals reflect those of foods consumed. Carbon is incorporated into plant tissues during photosynthesis. Carbon stable isotope ratios of plants vary depending on the environmental sources of carbon and the fixation mechanism used. Most plants that constitute a significant part of human foodwebs fix carbon through one of two photosynthetic This article is part of the Topical Collection on Post-glacial human subsistence and settlement patterns Electronic supplementary material The online version of this article (https://doi.org/10.1007/s12520-020-01087-1) contains supplementary material, which is available to authorized users. routes, either the C 3 or the C 4 pathway (DeNiro and Epstein 1978). C 4 plants were a very minor component of the Northwest European post-glacial flora (cf. Long 1983) and, although some species are edible, they are unlikely to have contributed significantly to the diets of Lateglacial and Holocene hunter-gatherers. Most marine plants and temperate region grasses (including wild and domestic varieties of cereals such as wheat, barley, oats and rye), as well as most fruits and vegetables, have C 3 cycles. Terrestrial C 3 plants, which fix carbon from atmospheric CO 2 , have lower average δ 13 C (c. − 26.5‰) than marine plants owing to distinct environmental carbon sources-marine plants fix carbon from oceanic carbonate as well as CO 2 , which is relatively 13 C-enriched (+ 7‰ in comparison to atmospheric CO 2 ) (e.g. Maberly et al. 1992). Variations in plant δ 13 C are passed on through respective food chains into the tissues of human consumers. There is an offset or fractionation of c. + 5‰ between plants and consumers, i.e. human consumers of terrestrial C 3 resources typically have δ 13 C values of c. − 21.5‰, while human consumers of marine resources typically have δ 13 C values of c. − 13.0‰ (Chisholm et al. 1982;Tykot 2004). Measurement of bone collagen δ 13 C can thus indicate the relative proportions of terrestrial versus marine food sources in diet.
Notably, the δ 13 C values of freshwater food sources may overlap with those of terrestrial C 3 food webs (e.g. Bonsall et al. 1997), particularly in temperate regions (i.e. with C 3 vegetation in the watershed) and water systems where the main source of dissolved inorganic carbon is atmospheric CO 2 (Finlay and Kendall 2007).
Co-analysis of both δ 13 C and δ 15 N allows additional dietary discrimination. Nitrogen in plants may be assimilated with little fractionation from atmospheric N 2 as well as from soil (Nadelhoffer and Fry 1994). Plant δ 15 N, therefore, varies according to the environmental source (von Wirén et al. 1997). Metabolic fractionation of nitrogen stable isotopes occurs at each trophic level of the food chain, resulting in 'stepped' 15 N enrichment. Within a single biome, plants have lower δ 15 N values than herbivores, which in turn have lower values than carnivores (DeNiro and Epstein 1981;Katzenberg 2000). The offset between humans and diet has been proposed to be up to c. 6‰ (O'Connell et al. 2012). Measurement of bone collagen δ 15 N can thus indicate the relative importance of plant versus animal food sources in diet, and by extension the trophic level of the consumer. Additionally, δ 15 N values enable discrimination between terrestrial and freshwater food sources. Again, the two food sources may be distinguished by δ 15 N values, which are typically higher in freshwater foods. Elevated δ 15 N in aquatic foods is a consequence of the larger number of trophic levels in both marine and freshwater food webs and may also result from enrichment through bacterial activity (Schoeninger and DeNiro 1984;Schoeninger 2010).

Dietary models
Stable isotope data for Lateglacial and Holocene huntergatherers in Britain, while limited, can shed light on aspects of past behaviour, including resource availability, subsistence strategies and mobility. Conventionally, linear models have been employed to quantify the proportions of different foods in diet from carbon and nitrogen stable isotope measurements (e.g. Richards et al. 2000;Bocherens and Drucker 2006;Stevens et al. 2010). However, such models are problematic.
Linear Mixing Models (LMMs) have significant limitations: & Mathematical constraints limit the number of food sources that can be robustly modelled. Generally, dietary proportions have been quantified from one (δ 13 C) or two proxies (δ 13 C and δ 15 N). However, multiple isotopically distinct food sources are attested in the archaeological record, often greatly exceeding the number of dietary proxies. Where the number of food sources exceeds the number of proxies by more than one, multiple dietary 'solutions' may be generated by LMMs (Phillips et al. 2005; e.g. Bocherens and Drucker 2006)-and these are of limited utility. & The greater precision offered by combining variables (Phillips et al. 2005) has led to the 'lumping' of food sources and the creation of binary (i.e. terrestrial vs. marine) dietary models of hunter-gatherer diet (e.g. Richards et al. 2005

Bayesian modelling of palaeodiet
In principle, Bayesian mixing models (BMMs) offer more realistic reconstructions of dietary intake than conventional LMMs (Fernandes et al. 2014;Parnell et al. 2014). Uncertainties in trophic level offsets and food-source isotope values, as well as variation in dietary routing and food group elemental composition, can be incorporated into BMMs. The Bayesian mixing model FRUITS-Food Reconstruction Using Isotopic Transferred Signals-is used here to evaluate the relative caloric contribution of multiple food sources to an individual's whole diet (Fernandes et al. 2014(Fernandes et al. , 2015. Four models of British hunter-gatherer diets were generated using FRUITS: MODELS 1 and 2 are protein routed concentrationdependent models, which assume dietary protein was directly routed to bone collagen. Model 1. Late Upper Palaeolithic (LUP) diets reconstructed using LUP food source stable isotope values.
Model 2. Mesolithic diets reconstructed using Mesolithic food source stable isotope values.
Model 3. LUP diets reconstructed using LUP food source stable isotope values.
Model 4. Mesolithic diets reconstructed using Mesolithic food source stable isotope values.
Shellfish Δ 13 C protein-lipid is calculated to be − 3.5‰ (from data in Ricca et al. 2007). Studies of the isotope values and offsets in marine mammal tissues are limited (for a review, see Newsome et al. 2010). Offset values between consumed tissue protein and lipids to bone collagen were determined from published offsets between diet and keratin, lipid or muscle values for seal. The δ 13 C offset between keratin and bone collagen is c. − 1.5‰ (Bocherens et al. 2014;Crowley et al. 2010). It is assumed there is no significant difference between the δ 15 N values of keratin and collagen (Bocherens et al. 2014). Variation in the carbon and nitrogen stable isotope ratios of different keratinaceous tissues, e.g. hair, claw and whisker, in seals with uniform diets appear to be small (see Hobson et al. 1996).
Errors in mass spectrometric measurements introduce uncertainty in consumer δ 13 C and δ 15 N values. Although measurement error is generally reported to be in the order of ± 0.1‰ for δ 13 C and ± 0.2‰ for δ 15 N for the current dietary reconstruction, uncertainty was cautiously set at 0.5‰ for both δ 13 C and δ 15 N following Fernandes et al. (2014Fernandes et al. ( , 2015. Uncertainty in food source stable isotope values (reflecting differences in preparation methods and seasonal and physiological variations in animal metabolism) were set at 1.0‰ (Fernandes et al. 2014(Fernandes et al. , 2015.

Food group composition
Establishing the proportional contribution of protein/energy in food sources is essential to the accurate reconstruction of diet (Phillips and Koch 2002). Carbon weight composition (wtC%) of plant cereals were drawn from Fernandes et al. (2015). Mean composition values for terrestrial mammals, shellfish, fish and sea mammals were calculated from available food composition data from the United States Department of Agriculture (USDA n.d.) Food Composition Databases (https://fdc.nal.usda.gov/), rounded to a multiple of 5 (cf. Fernandes et al. 2015).
Terrestrial animal protein/energy concentration varies significantly depending on the species and also the portion of the animal consumed. For example, the body fat content of deer is highly variable seasonally and individually: fat content of femur marrow in white-tailed deer fawns can exceed 80%, while total body fat ranged from 2.3% to 48.9% and protein content from 39.2% to 75.5% (Watkins et al. 1991). Wild boar lean muscle, e.g. the tenderloin, has wtC% protein/energy of c. 85:15, while a more fatty portion of the animal, e.g. the belly, has wtC% protein/energy of c. 40:60. It is assumed in our models that all edible parts of the animal were consumed. Intermediate wtC% protein/energy values were used with conservation errors to account for variability in food sources, wtC% protein = 60 ± 5% and wtC% energy = 40 ± 5%.
The protein content of wild plants is highly variable. Plants with relatively high protein content, such as hazelnuts, may have been important in some post-glacial hunter-gatherer diets. Mushrooms with much lower protein content could also have been important. Plant wtC% protein was therefore set at 10 ± 5% and wtC% energy = 90 ± 5%.
Defining the lipid/protein carbon weight composition of fish (marine and freshwater) and sea mammals is non-trivial. The lipid content of fish varies significantly both within and between species (e.g. Berg and Bremset 1998;Pinnegar and Polunin 1999); typically < 1 g/100 g in lean fish such as gadids, and > 10 g/100 g in fatty anadromous species including salmonids and eel (USDA database). Both lean and fatty fish have been identified in the fish assemblages of Lateglacial and Early Holocene sites across Europe (e.g. Pickard and Bonsall 2004;Robson et al. 2016). While freshwater fish are absent from British LUP assemblages and the availability of these resources is debatable (see discussion below), they are relatively common on Lateglacial sites in Continental Europe (e.g. Cleyet-Merle 1990; Bonsall et al. 2016). The mean protein and energy wtC% of marine species typically recovered from British hunter-gatherer sites are c. 75 ± 5% and 25 ± 5% (USDA database), respectively, while those of freshwater fish are c. 65 ± 5% and 35 ± 5% (USDA database; Fernandes et al. 2015). Shellfish have a relatively uniform lipid/protein carbon weight composition with wtC% of c. 90 ± 2.5% and 10 ± 2.5%, respectively (USDA database). Protein and lipid concentrations in sea mammals are also variable. Food composition databases offer a limited range of comparanda. Based on published values, sea mammal (meat and subcutaneous fat) composition values were calculated as wtC% protein = 45 ± 5% and wtC% energy = 55 ± 5%.
Food source δ 13 C and δ 15 N values There were three broad categories of isotopically distinct food sources that may have contributed to the diets of post-glacial hunter-gatherers: (1) wild plant foods; (2) terrestrial wild mammals; (3) aquatic resources (freshwater and marine).
& Wild plant foods. Bayesian models of diets should elucidate the role of plant foods in LUP and Mesolithic diets. However, studies of the δ 13 C and δ 15 N values of archaeological wild plant food remains are scant. Modern comparanda may be used as proxies in dietary models (e.g. Meadows et al. 2019). δ 13 C and δ 15 N values of hazelnuts and mushrooms, two protein-rich plant foods from the 'primaeval' forest of Białowieża, Poland, were reported by Selva et al. (2012) to be δ 13 C = − 31.3 ± 0.4 and δ 15 N = − 0.8 ± 0.4 and δ 13 C = − 20.1 ± 0.3 and δ 15 N = − 0.2 ± 0.7, respectively. However, modern proxies may introduce further uncertainties into dietary reconstructions. Although these data were corrected for the Suess effect (recent 13 C-depletion of atmospheric CO 2 resulting from the use of fossil fuels (Keeling 1979)), the samples were not pre-treated to remove lipids. Lipids are generally 13 C-depleted relative to proteins and carbohydrates (DeNiro and Epstein 1978). While δ 15 N values reported in Selva et al. (2012) reflect the plant protein values (as lipids do not contain any nitrogen), the δ 13 C values reflect combined plant protein and energy. To accurately model whole diet, δ 13 C values for both plant protein and energy are required. Therefore, the δ 13 C and δ 15 N values were taken from the mean δ 15 N of hazelnuts and mushrooms of − 0.4 ± 1.0‰ in Selva et al. (2012) and mean energy δ 13 C of − 23.5 ± 1.0‰ and mean protein δ 13 C of − 26.0 ± 1.0‰ published in Bogaard et al. (2013). & Terrestrial mammals. Climatic changes through the Lateglacial and Holocene are mirrored in temporal variation in environmental δ 13 C and δ 15 N (Drucker et al. 2003;Stevens and Hedges 2004). In the following analyses, food source δ 13 C and δ 15 N specific to the time period of each site are utilised. Typical isotope values of Lateglacial and Early-Middle Holocene terrestrial animals were determined from archaeological specimens (see Tables 1 and 2; site location, dates and sample details are provided in Table S1). For our Lateglacial models, stable isotope values of animal remains 14 C dated to the Bølling-Allerød were used. There are multiple analyses of several of the Lateglacial animal bone samples. The values used here for palaeodietary modelling are either the most recently published measurements, or in the case of those measured in separate studies for 14 C determination and dietary studies, the latter values are used. Mean bone collagen δ 13 C and δ 15 N values (with standard error of the mean, cf. Hedges et al. 2007) of Bølling-Allerød terrestrial herbivores (n = 50) were − 20.3 ± 0.1‰ and 1.8 ± 0.2‰, respectively. Incorporating the bone collagen to protein tissue offset and food source uncertainties, Lateglacial terrestrial herbivores have mean protein δ 13 C = − 22.3 ± 1.0‰, mean energy δ 13 C = − 28.3 ± 1.0‰ and mean protein δ 15 N = 3.8 ± 1.0‰. Mean bone collagen δ 13 C and δ 15 N values (with standard error) of Mesolithic terrestrial herbivores and omnivores (n = 36) were − 21.6 ± 0.2‰ and 4.0 ± 0.4‰, respectively. Incorporating the bone collagen to protein tissue offset and food source uncertainties, Mesolithic terrestrial herbivores and omnivores have mean protein δ 13 C = − 23.6 ± 1.0‰ and δ 15 N = 6.0 ± 1.0‰. & Aquatic food sources can be grouped into two main isotopic categories: freshwater and marine. Marine food sources can be subdivided into three further isotopic groups: (1) shellfish, (2) fish and seabirds, and (3) sea mammals. Establishing the δ 13 C and δ 15 N values of aquatic resources is non-trivial. Carbon and nitrogen stable isotope values of archaeological specimens of marine fish and seal are available for the Mesolithic and later Holocene periods (see Table 3). However, no data for archaeological freshwater resources or marine shellfish are available for Great Britain. Therefore, stable isotope values of archaeological freshwater fish are drawn from continental European sites (see Table 3), while shellfish flesh values are derived from modern comparanda (Bownes 2018). Mean carbon and nitrogen stable isotope values of archaeological marine fish are δ 13 C = − 12.9 ± 0.1‰ and δ 15 N = 14.0 ± 0.1‰ (n = 28). Mean carbon and nitrogen stable isotope values of Lateglacial fish from freshwater systems are δ 13 C = − 21.6 ± 1.0‰ and δ 15 N = 9.0 ± 0.6‰ (n = 7). Mean carbon and nitrogen stable isotope values of Mesolithic fish from freshwater systems are δ 13 C = − 20.6 ± 0.7‰ and δ 15 N = 9.4 ± 0.2‰ (n = 39). The δ 1 3 C and δ 1 5 N values for the Lateglacial and Mesolithic fish are not statistically different (Mann-Whitney U test, δ 13 C p = 0.614; δ 15 N p = 0.392), and so have been grouped for dietary modelling. Mean values for archaeological seal are δ 13 C = − 12.3 ± 0.3‰ and δ 15 N = 18.2 ± 0.3‰ (n = 9). Incorporating the bone collagen to protein tissue offset and food source uncertainties, the marine fish protein values are δ 13 C = − 13.9 ± 1.0‰ and δ 15 N = 16.0 ± 1.0‰; freshwater fish values are δ 13 C = − 21.7 ± 1.0‰ and δ 15 N = 11.4 ± 1.0‰; while seal protein values are δ 13 C = − 14.6 ± 1.0‰ and δ 15 N = 17.6 ± 1.0‰.
δ 13 C and δ 15 N of modern shellfish flesh are used in dietary models (e.g. Bonsall et al. 2009;Montgomery et al. 2013). Although such data should be corrected for the Suess effect, the impact on oceanic δ 13 C has been smaller than on atmospheric δ 13 C. In the North Atlantic, 13 C-depletion of up to 0.8‰ is evident (Eide et al. 2017). Lipid extraction has generally been undertaken before measuring the δ 13 C and δ 15 N signals of shellfish protein. However, lipid extraction has been demonstrated to alter δ 15 N values in a non-predictable manner (e.g. Post et al. 2007;Logan et al. 2008). It is best practice to use the δ 13 C values from a lipid-extracted sample alongside   the δ 15 N of the pre-treated sample (Sotiropoulos et al. 2004).
The shellfish flesh isotope values used for this study were selected from data in Bownes (2018). Only those samples for which (1) whole and lipid extracted δ 13 C and δ 15 N values were available and (2) with C/N ratios in the range 2.9-3.6, characteristic of collagen, following lipid extraction (cf. DeNiro 1985) were included in this analysis. Suess effect corrected mean carbon and nitrogen stable isotope values of shellfish protein are δ 13 C = − 14.6 ± 0.1‰ and δ 15 N = 7.4 ± 0.2‰ (n = 11).

Results and discussion
The diets of 85 individuals from 21 sites (see Table S2) were modelled using FRUITS based on their bone collagen C-and N-isotope signatures (see Tables 4, 5, 6 and 7). (Note-we treated each δ 13 C and δ 15 N paired measurement as a separate individual unless it was specified in publications that duplicate measurements of the same skeletal element/individual had been measured; however, it is acknowledged that this figure represents a maximum number of individuals.) For the most part, we selected targets with direct AMS 14 C dates (Table S2). In the case of the human remains from Thatcham (Fig. 1, site 20), however, dating is based on associated palynological evidence (Churchill 1963). The radiocarbon dates cluster into three distinct periods: 15.0-13.6 ka BP (first half of the Bølling-Allerød Interstadial), 9.5-9.0 ka BP (Early Holocene) and 8.3-6.2 ka BP (early Middle Holocene). Archaeologically, these correspond to phases within the 'Late Upper Palaeolithic' (LUP), 'Early Mesolithic' and 'Late Mesolithic', respectively. Eleven of the 21 sites under consideration are located on or near (within 2 km) the present-day coastline (Fig. 1). While distance to the sea would have varied in response to relative sea-level (RSL) changes and shoreline displacement during the post-glacial period, in most cases, the communities who used the sites during the Lateglacial or Early-Middle Holocene would likely have had direct access to coastal resources. A possible exception is Tilbury near the mouth of the River Thames (Fig. 1, site 21) where during the Late Mesolithic c. 8 ka BP with RSL of c. − 10 to 12 m (cf. Shennan et al. 2006, figure 7), the river habitat was likely freshwater rather than estuarine. Similarly, the Cannington   Park Quarry site, today located c. 1.9 km from the estuary of the River Parrett in southwest England (Fig. 1, site 13), may have been further from the coast during the Early Holocene.

Late Upper Palaeolithic diets (models 1 and 3)
Three LUP sites are included in our study. They comprise two inland sites, Gough's Cave and Sun Hole in southwest England, and one 'coastal' site, Kendrick's Cave in North Wales (Fig. 1, sites 3, 16 and 17). With RSL of c. − 15 to 20 m along the North Wales coast between 12.0 and 11.7 ka BP (Shennan et al. 2006, figure 7), Kendrick's Cave would have been further from the shoreline than it is today.
At Gough's Cave and Sun Hole, the vast majority of the diet was derived from terrestrial resources. Both protein routed (model 1) and nutrient scrambled (model 3) models indicate that at each of these sites the mean proportion of terrestrial (animal + plant) protein and food source calories was at least c. 89% (see SH1 in Tables 4 and 6), and possibly higher. Means and associated standard deviations suggest that the dietary contribution of marine and freshwater resources was likely negligible. Dependence on terrestrial resources is broadly consistent with previous interpretations of Lateglacial diet at Sun Hole and Gough's Cave (e.g. Richards et al. 2000;Stevens et al. 2010).  Richards et al. (2005:392) proposed that the carbon stable isotope values were consistent with 'a diet where approximately 30% of dietary protein was from marine resources' and also argued that the nitrogen stable isotope values indicated 'a mix between protein from terrestrial herbivores, such as Cervus elaphus, and marine mammals, such as seals'. Bocherens and Drucker (2006) highlighted some of the limitations of Richards et al.'s (2005) analysis: the failure to incorporate uncertainty in δ 13 C trophic level shifts, the small sample size of the terrestrial mammals used as a food source baseline and the absence of freshwater resources from the food sources modelled. Bocherens and Drucker (2006) re-evaluated the Kendrick's Cave data using a mixing model for four food sources-terrestrial herbivores, salmon, seal and freshwater fish. They concluded that more than one dietary model fits the data, asserting that 'it is not necessary to incorporate herbivore meat to explain the human isotopic values' (Bocherens and Drucker 2006:441) and diets based entirely on aquatic resources (both freshwater and marine) could account for the stable isotope signatures of the Kendrick's Cave humans.
Here, it is also worth noting that remains of fish and other aquatic foods are generally scarce on Lateglacial sites in Britain, perhaps due to taphonomic and recovery biases (Bocherens and Drucker 2006). Fish remains are generally more friable than mammal bones (Nichol and Wild 1984) and this, combined with the fact that many LUP and Mesolithic sites were excavated before fine recovery techniques were widely practised, limits the chances of recovery. Many coastal sites in the southern half of Britain older than c. 7.0 ka BP were probably submerged during the post-glacial marine transgression, although the rare find of the scales of lemon sole (Microstomus kitt) from Pinhole Cave, Creswell Crags, northern England (Armstrong 1932) is confirmation that marine resources were being exploited at least occasionally during the LUP.

Mesolithic diets (models 2 and 4)
Human stable isotope data from 18 Mesolithic sites were analysed, seven dated to the Late Mesolithic and 11 to the Early Mesolithic (n.b. the specimens attributed to a site at Mewslade Bay are assumed to have originated from Worm's Head; cf. Meiklejohn et al. 2011). FRUITS modelling suggests significant site diversity in dietary intake, ranging from groups whose diets were largely based on terrestrial food sources to those with mixed diets that included terrestrial, marine and freshwater foods.
Coastal hunter-gatherers at Cnoc Coig, Oronsay, had the highest proportion of diet derived from marine foods (with the highest individual mean intake of marine protein in sample CC14 of c. 56% (shellfish 33 ± 12%, fish 11 ± 8% and seal 12 ± 9%) in model 2 and of whole diet calories in sample CC10 of c. 66% (shellfish 25 ± 11%, fish 20 ± 11% and seal 21 ± 13%) in model 4). Previous interpretations of diet at Cnoc Coig have emphasized the importance of marine resources with linear mixing models based on the δ 13 C and δ 15 N values used to suggest that up to 100% of dietary protein was derived from marine foods (cf. Richards and Mellars 1998;Schulting and Richards 2002b;Charlton et al. 2016). Our FRUITS models, using essentially the same data but modelling for food source calories, suggest that terrestrial foods made a significant contribution to overall diet (with individual intake of terrestrial protein up to c. 42% in model 2 (e.g. sample CC1with 20 ± 15% plant and 22 ± 13% herbivore/omnivore protein) and up to c. 36% of total calories in model 4 (e.g. sample CC4 with 19 ± 15% plant and 17 ± 11% herbivore/omnivore calories)). In most samples, the proportion of plant versus terrestrial mammal food in the diet is roughly equal. Both FRUITS models highlight the dietary importance of shellfish as well as fish and sea mammals. The nutrient scrambled model suggests some individuals obtained more than one-third of their calories from shellfish.
The results from Cnoc Coig hint at individual variation in dietary intake (see Fig. 2). For example, the nutrient scrambled model suggests that individual CC12 may have obtained up to 39 ± 13% of calories from shellfish, while individual CC6 consumed a much smaller proportion of shellfish, c. 20 ± 11%. However, individual variation in dietary intake at Cnoc Coig may be inflated by uncertainty over the age-atdeath of the individuals sampled. Although this may also be an issue at other sites, the use of ZooMS to identify small, undiagnostic fragments of bone suggests this may be especially problematic at Cnoc Coig. A nursing effect in infants and young children causes 13 C-and 15 N-enrichment above maternal values of up to c. 1‰ and c. 3‰, respectively (Fuller et al. 2006). Additionally, increased protein requirements during adolescence may result in reduced fractionation of dietary 14 N/ 15 N (e.g. Waters-Rist and Katzenberg 2010). Bone collagen laid down in adolescence may, therefore, have lower δ 15 N than at other stages of life in individuals with monotonous 'lifetime' diets. . NM -no model was generated Similar intra-site variability is evident at Ogof-yr-Ychen in South Wales. However, in this case, terrestrial foods constitute dietary staples; one individual analysed (OY7) had whole diet calorie proportions that were 33 ± 22% plant, 39 ± 18% terrestrial mammal, 12 ± 9% shellfish, 4 ± 3% marine fish, 5 ± 4% seal and 8 ± 7% freshwater (model 4). Schulting and Richards (2002a, Table 1) proposed that the main dietary protein source was marine 63-72% at Ogof-yr-Ychen and that individuals with particularly high δ 15 N had consumed significant quantities of seal meat. Our FRUITS models suggest a mixed diet with a particular emphasis on freshwater fish. For each individual, the mean estimate of total calories obtained from shellfish (12 ± 9% to 25 ± 12% in model 4) is equal to or exceeds that from seals (5 ± 4% to 13 ± 10% in model 4). Richards (2002a:1017) observed that certain individuals from Ogof-yr-Ychen 'show a more balanced use of marine and terrestrial resources that could imply seasonal movements; inland groups may have maintained social links with coastal communities allowing them access at certain times of the year'. Our FRUITS estimates suggest that all but two of the samples analysed from Ogof-yr-Ychen had broadly similar mixed diets. The two samples, C and C*, which may come from one individual, have a slightly higher proportion of terrestrial food calories in diet than the other samples, though both show similar proportions of marine food calorie intake to the other samples from Ogof-yr-Ychen-the main difference is in the contribution to diet of freshwater fish.
At all of the other Mesolithic sites analysed, terrestrial food sources dominated the diet. However, there is variation between and within groups in the proportions of plant and herbivore/omnivore food sources consumed. At Aveline's Hole, for example, the contribution of plant foods to protein was 39 ± 25% to 86 ± 8% (model 2) and to whole diet 39 ± 25% to 80 ± 17% (model 4), while that of terrestrial mammals to protein was 8 ± 8% to 45 ± 25% (model 2) and to whole diet % freshwater whole diet calorie contribution, (d) % marine protein contribution, (e) % terrestrial protein contribution and (f) % freshwater protein contribution 12 ± 16% to 48 ± 25% (model 4). The individual from Bower Farm had a diet with a very high proportion (96 ± 2%) of plant foods unless the unusually low δ 15 N value of 2.8‰ is a reflection of long-term nutritional stress or disease.

Wider issues
We encountered a number of technical and practical issues relating to Bayesian modelling of the British data: & The FRUITS software (v. 3.0 beta) would fail to generate model outputs for certain site datasets or even individual samples, most particularly where we attempted to introduce prior information into our models. Partly for this reason, our dietary models do not include the prior assumption of 5-40% of protein intake recommended by Fernandes et al. (2014). Initially, we attributed these problems to the fact that the number of food sources (six) far exceeded the number of proxies (two-δ 13 C, δ 15 N). To investigate this issue further, FRUITS models were generated for four food sources: plants, terrestrial herbivores/ omnivores, freshwater fish and grouped marine resources.
The results were broadly similar in terms of food proportions in diet, although overall there was an increase in the proportion of plants in diet compared to the six-foodsource models and a corresponding decrease in the proportion of marine food. Roughly the same proportion of samples failed to generate model outputs in both the fourand six-food-source models. It is therefore tentatively suggested here that the problems encountered with model generation and also possibly the large one-sigma standard deviations of certain modelled diets may result from the inadequacies in food source data rather than the number of proxies or the number of food sources. & A third proxy, sulphur stable isotope ratios (δ 34 S), has been incorporated into dietary models. Arguably, δ 34 S is a particularly useful aid to dietary reconstruction in situations where freshwater food sources were available (e.g. Nehlich et al. 2010 Table 6) are particularly surprising given that during the early part of the Bølling-Allerød Interstadial, trees that produce high nutritional value nuts (e.g. acorn, hazelnut) were absent from the British flora (cf. Ince 1996;Hill et al. 2008). & These observations highlight the current lack of wild plant food protein and energy δ 13 C data from Britain. Our FRUITS models were constructed using data from domesticated plant food remains in Bogaard et al. (2013), which may not be appropriate. Further research is vital to determine stable isotope values for local edible wild plants including hazelnuts, which may have been subject to a canopy effect (van der Merwe and Medina 1991). & A related point is that FRUITS assumes the contribution to dietary carbohydrate (CH) from animal tissues is negligible (Fernandes et al. 2014) and (by extension) is obtained almost exclusively from plants. However, CH occurs in some organs (e.g. liver) of terrestrial and marine mammals and some species of shellfish, especially bivalves such as cockles, oysters and mussels (USDA database). Honeytechnically, an animal (insect) product-is also a rich CH source. Moreover, post-glacial hunter-gatherers probably consumed nearly all the edible parts of the animals they exploited for food (meat, organs, bone marrow, etc.), providing higher amounts of CH and fat than meat alone. It follows that our FRUITS models may overestimate the proportion of diet obtained from plants. & Our FRUITS estimates for the contribution of freshwater resources to post-glacial hunter-gatherer diets are also unexpected. Populations living near one of the longest rivers in Britain-at Bower Farm and Staythorpe in the valley of the River Trent-show relatively minor contributions from freshwater fish (1 ± 1% and 4 ± 3%, respectively, in both models 2 and 4) ( Tables 5 and 7). Likewise, freshwater resources were apparently unimportant at the Early Holocene wetland site of Thatcham III in southern England-although the single human bone from this site was found in a secondary context (Churchill 1963) and so the Mesolithic dating is not secure. It is possible, however, that our model estimates are inaccurate; the lack of local baseline data for freshwater fish necessitated using stable isotope data for freshwater fish from a very broad geographic area (for a review of variation in freshwater fish stable isotope values, see Guiry 2019). By contrast, individuals from Ogof-yr-Ychen were modelled to have had a relatively high proportion of freshwater food sources in their diets despite the lack of large bodies of fresh water on or near Caldey Island. This may relate to group movement (see Preston and Kador 2018 for a review of mobility models among British hunter-gatherers) and exploitation of resources across the broader Carmarthen Bay area and beyond. More particularly, it may reflect a further complexity in modelling hunter-gatherer diets-the distinctive isotopic signatures of estuarine resources. The Carmarthen Bay area encompasses extensive estuarine habitats. Estuaries typically support a greater abundance and diversity of potential food resources than other coastal habitats, including fish, shellfish, birds and edible plants, which made them particularly attractive to post-glacial huntergatherers (Bonsall 1981). Fish and shellfish harvested from estuaries may have δ 13 C intermediate between marine and freshwater values (cf. Thornton and McManus 1994), and diets composed largely of estuarine resources may mimic mixed diets of marine and freshwater food sources. This dietary scenario may also hold for the sample from Caisteal nan Gillean II, Oronsay, which had similar δ 13 C and δ 15 N values to some of the individuals from Ogof-yr-Ychen. & Research is also required to establish tissue offsets in marine mammals. In terrestrial mammals, muscle tissue is generally 15 N-enriched relative to bone collagen; however, this was not evident in Hobson et al.'s (1996) study of seal tissues, though their data are limited in number.

Conclusions
FRUITS modelling of the diets of post-glacial hunter-gatherers in Great Britain, based on bulk bone collagen stable Cand N-isotope values, suggests there was significant spatial and temporal variability in subsistence practices. At the three LUP sites included in our analysis (Gough's Cave and Sun Hole in south-west England, and Kendrick's Cave in North Wales), terrestrial food sources appear to have been the dietary staples, although at Kendrick's Cave freshwater resources also made a significant contribution. While marine foods may have contributed to diet at all three sites, they constituted at most a very minor resource.
Previous stable isotope studies of Mesolithic diets have tended to emphasise the role of aquatic (primarily marine) foods and perhaps underestimated the contribution of plant foods. Models of food source calorie contribution to diet indicate that only two coastal sites in this study (Cnoc Coig and Caisteal nan Gillian II on the island of Oronsay, off the west coast of Scotland) show a clear emphasis on aquatic resources, although terrestrial resources also contributed significantly to caloric intake. In contrast, FRUITS models of the Mesolithic populations living along the South Wales coast suggest diets where often the majority of calories were drawn from terrestrial resources, although in at least one case (Ogof-yr-Ychen) aquatic resources appear to have been important. However, these models are treated cautiously in light of the confounding isotope values of estuarine resources.
Predictably, Mesolithic inland sites show a heavy emphasis on terrestrial (animal and plant) resources, though one surprising outcome of our model predictions is the limited evidence for the use of freshwater resources at sites in the vicinity of large rivers or wetlands (Bower Farm, Staythorpe and Thatcham III).
We are acutely aware of the shortcomings of the research presented here. We were unable to generate FRUITS models for certain sites and individuals, possibly related to the small number of proxies relative to food sources and/or deficiencies in the food source data. Moreover, our model predictions are likely biased by inadequate baseline δ 13 C and δ 15 N data for wild terrestrial plant, freshwater and estuarine resources, in particular. Accurately quantifying the proportion of aquatic resources in diet is also crucial in calculating radiocarbon reservoir offsets.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.