Armenia obsidian sources (black circles) included in this study, archaeological sites referenced in the text (black squares), and source complexes (dashed lines) as conceptualized in the tests. Localities with different names but identical compositions are represented by a single dot (e.g., Djraber, Gyumush, Fontan, etc. localities of Gutansar). No attempt is made here to precisely represent the full primary and secondary distribution of the obsidians. For this study, Syunik complex specimens did not include sources only recently surveyed by our team, such as Mijnek Satanakar, Pokr Satanakar, and Merkasar. Although not conceptualized as such for this study, Gutansar and Hatis can be considered the Hrazdan-Kotayk group, while Khorapor is part of the Vardenis group. Apnagyugh-8 is unof fi cially known as Kmlo-2. Digital elevation data from SRTM3 (Shuttle Radar Topography Mission dataset version 3), and base map shared and modi fi ed under Creative Commons terms from Wikimedia Commons. 

Contexts in source publication

Context 1

... smaller than Belgium or the state of Maryland, Armenia has more than a dozen obsidian-bearing volcanic centers, resulting in one of the most obsidian-rich natural and cultural landscapes in the world. At numerous Palaeolithic sites, obsidian comprises the majority, if not the entirety, of the lithic assemblage. This is the case at two sites recently excavated by the Hrazdan Gorge Palaeolithic Project (Adler et al., 2012): Nor Geghi 1, an open- air Lower Palaeolithic site, and Lusakert 1, a Middle Palaeolithic cave site ( Fig. 1). Their lithic assemblages are more than 99% obsidian. At Lusakert 1, in particular, after four excavation seasons (2008 e 2011), 13,970 obsidian artifacts have been recovered (excluding 5970 fragments smaller than 25 mm) from 11.9 m 3 of sediment. That is, 1174 obsidian artifacts were recovered per cubic meter. On average, 470 obsidian artifacts were excavated daily with spatial data recorded by two Leica total stations. In 2011, the project began a new program of obsidian studies, including artifact sourcing as well as source surveys and characterization. During the 2012 season, we analyzed over 1700 artifacts in our fi eld house using portable X-ray fl uorescence (pXRF). We established that the Nor Geghi 1 and Lusakert 1 assemblages were both approximately 93% obsidian from Gutansar (sometimes spelled as Gutanasar), the nearest volcanic center with abundant obsidian resources. The remainder came from numerous sources throughout Armenia, including Hatis, Pokr and Mets Arteni, Pokr Sevkar, Geghasar, and the Tsakhkunyats sources. Publications on these fi ndings, included detailed source and spatial data, are currently in preparation. The focus of this paper is a methodological development that arose out of our work, speci fi cally a desire to have obsidian source information available on-site. The methods we document here will be deployed in future seasons, becoming a routine component of our excavation toolkit and constituting a key strategy in our approach to site surveys and assessments. As hundreds of artifacts were sourced in our fi eld house during the 2012 season and the results compared to the spatial data, it became clear to us (for reasons discussed in Section 2) that an artifact ’ s geological origin would be useful information to have on- site during an excavation or survey. Our tests of visual classi fi cation yielded little success. Gutansar obsidian is highly variable in appearance, and as Table 1 shows, any “ exotic ” artifacts from other sources were overlooked and grouped into types with Gutansar artifacts. Thus, we began our efforts to develop methods for shifting pXRF from our fi eld house into the fi eld itself. As discussed in Section 2.4, pXRF-based obsidian sourcing has hitherto been largely conducted in laboratories, museums, and fi eld houses. Although our colleagues are using dust- and waterproof instruments with 10- h batteries, obsidian sourcing remains embedded in post- excavation studies and is rarely, if ever, done in the fi eld. A critical factor in the uptake of obsidian sourcing in the fi eld is the time needed for each measurement. In recent pXRF-based obsidian studies (Table 2), analyses took 2 e 6 min. The most common duration is 5 min, corresponding to 12 artifacts per hour. Our excavations at Lusakert 1 yielded, on average, 70 e 80 artifacts per hour. Although this suggests a need for 45-s measurements, we aimed for 10 s so that these analyses could become syncopated with the excavation activities. Additionally, we deemed it was insuf fi cient simply to conduct a measurement in that time. After 10 s, we wanted the instrument ’ s built-in LCD to display an artifact ’ s source so excavators and surveyors could instantly know the result. Reducing measurements from a mode of 300 s to just 10 s and having the instrument ’ s onboard software automatically do the data analysis is bound to raise challenging assumptions about the validity and reliability of our approach. We show that the technological capability to source obsidian artifacts rapidly on-site exists, but the methods to do so effectively were previously undeveloped. Our tests with Armenian geological specimens and Palaeolithic artifacts demonstrate the high ef fi cacy of the two methods we report here. Obsidian artifact sourcing conducted rapidly on-site can trans- form the ways in which our discipline approaches subjects involving raw-material procurement, transport, and use as well as the organization of space and the identi fi cation of activity areas. In this paper, “ pXRF ” refers to the ruggedized instruments about the size and shape of a cordless drill. In contrast, some researchers (e.g., Craig et al., 2007; Liritzis and Zacharias, 2011; Speakman and Shackley, 2013) consider “ pXRF ” to include benchtop instruments that could be transported from an analytical laboratory to a similar context in a museum or fi eld house. Such instruments have been used to good effect for sourcing obsidian artifacts (e.g., Cecil et al., 2007; Speakman et al., 2007; Liritzis, 2008). These systems, though, require a computer and electrical outlet, and they would not fare well outdoors. We focus on handheld analyzers designed for use in the fi eld (e.g., geological exploration) and industrial settings (e.g., scrapyards, manufacturing). Obsidian sourcing at an archaeological site may yield insights that inform the excavation strategies. Consider, for instance, that most of the obsidian at our two sites originated ...

Context 2

... software features: (1) “ pass/fail ” mode for testing consumer products and (2) “ pseudo-elements, ” commonly used to automatically calculate percent oxide or element ratios. This method follows Hughes ’ approach to classi fi cation, and it remedies the key weakness of our fi rst method: a lack of quantitative elemental data. pXRF is often used to test consumer products for compliance with regulatory limits. For example, the software can be used with pass/fail criteria for the European Union ’ s Restriction of Hazardous Substances Directive (RoHS). RoHS prohibits, for example, Pb above 1000 ppm. A Pb concentration over 1300 ppm results in the measurement highlighted in red to mark a “ fail. ” If Pb is below 700 ppm, the result is a green “ pass. ” A yellow “ inconclusive ” results if the level falls in the intermediate range. Thresholds may be set for any element; however, the algorithms are rather simple. The software assumes “ pass ” is below a low threshold and “ fail ” is above a high one. There is no way to de fi ne an intermediate range as “ pass. ” Another issue is illustrated by Fig. 5. Suppose one wishes to distinguish two obsidians based on Elements A and B. Neither source can be de fi ned by ranges of Elements A and B due to overlaps. There is, though, an axis on which the two populations are perfectly differentiated: line c - d. This axis combines Elements A and B in a discriminant function (line e-f ). With more elements, discriminant analysis (DA) derives the best discriminating axes in higher dimensions. These functions can also classify new observations (i.e., analyzed artifacts). The general form of a discriminant function is: where D 1⁄4 discriminant function; c 1⁄4 discriminant coef fi cients; X 1⁄4 original variables (e.g., elements); i 1⁄4 number of variables; A 1⁄4 constant. Such an equation is compatible with the software ’ s “ pseudo- elements. ” For example, a geologist may be more interested in the ratio of CuO to NiO in a rock outcrop than the absolute Cu and Ni contents. The software can report a pseudo-element de fi ned as (Cu  1.252)/(Ni  1.273), and the calculations are done automatically and displayed onscreen. If that geologist is interested in identifying rocks above a particular ratio, pass/fail thresholds could be de fi ned (and quantitative data for Cu and Ni would still be acquired to calculate the pseudo-element). Consequently, we derived discriminant functions, entered them as pseudo-elements, and calculated thresholds for six Armenian obsidian source complexes. It must be emphasized that, to calculate the pseudo-elements, fully quantitative elemental data (i.e., data corrected for matrix effects, calibrated, and converted into element concentrations) are acquired simultaneously. Five elements were selected for DA: Zr, Sr, Rb, Nb, and Fe. For two reasons, these are among the elements most frequently used for obsidian sourcing (e.g., Seelenfreund et al., 2002; Silliman, 2005; Negash and Shackley, 2006; Carter and Shackley, 2007; De Francesco et al., 2008; Phillips and Speakman, 2009; Shackley, 2009). First, their concentrations vary little within obsidian fl ows but can vary as much as three orders of magnitude among vol- canoes (Rapp and Hill, 2006:225). Second, due to X-ray physics, elements in “ mid-Z ” portion of the periodic table are measured especially well with XRF (e.g., Giaque et al., 1993). Shackley (2008) notes these elements are “ some of the most sensitive ” for sourcing (203). Therefore, this method follows a Hughesian approach: using the best-measured, most-discriminating elements. These elements are also measured using the “ main ” X-ray fi lter and, therefore, can be simultaneously quanti fi ed. Other elements, including Ba and Ti, have proven useful in prior obsidian sourcing studies. However, Ba must be measured using Niton ’ s “ high ” fi lter, which would increase the measurement duration. Similarly, Ti can be detected using the main fi lter; however, precise quanti fi cation at its concentrations in obsidian requires the “ low ” fi lter. As noted in Section 4.1, measurement uncertainty is inversely proportional to the square root of the measurement time. Fig. 6 shows the results of our tests regarding this relationship for Zr, Sr, Rb, Nb, and Fe. For all fi ve elements, uncertainties are below 10% (at the 2 s -level) after 10 s. In fact, for Sr and Rb, the uncertainty is 6% after 10 s, and for Zr and Fe, it is 4%. Consider that NAA has uncertainties of 2 e 5% for elements such as Rb and 5 e 10% for Sr and Zr (Glascock et al., 2007: 346). After 10 s, these elements have uncertainties comparable to NAA data, and there is little point in longer measurements. With a limit of 15 pseudo-elements, de fi ning populations using two to three discriminant functions restricted us to fi ve to seven groups. Using the XLSTAT Pro statistical package, we derived three discriminant functions based on more than 1800 analyses from six obsidian source complexes (Fig. 1): Arteni ( n 1⁄4 356), Geghasar ( n 1⁄4 58), Gutansar ( n 1⁄4 886), Hatis ( n 1⁄4 330), Syunik ( n 1⁄4 81), and Tsakhkunyats ( n 1⁄4 78). Therefore, as a trade-off for fully quantitative elemental data, the distinction between, for example, Pokr and Mets Arteni obsidians would not be evident in the fi eld with this con fi guration, but subsequent data analyses could discern them. If distinguishing Pokr and Mets Arteni was important, newly derived functions could do so. We derived three discriminant functions using Nb, Zr, Sr, Rb, and Fe, and since pseudo-elements are assumed to yield positive numbers, the functions ’ constants were increased so that all results are positive. The fi nal equations ...

Context 3

... during an eight-hour shift. These algorithms can also be used with other materials and custom libraries. In the Niton software, this is called Spectral Fingerprint mode. First one uses the Teach Fingerprint routine to record spectra from reference materials and create a custom library. Each measurement takes 60 s. Then one measures a specimen using Match Fingerprint , and the built-in screen shows the closest matching reference in realtime. To identify a matching reference material, the algorithms use a Pearson ’ s chi-squared ( c 2 ) test as a means to determine goodness of fi t between the spectrum from a specimen and those in the library (Hejzlar and Pesce, 2009). This statistic is one of the most common metrics for “ best fi ts ” in spectrometry. Speci fi cally, the Niton software uses measured X-ray intensities from 24 pre-set elements to calculate the mean squared distance (MSD) between measurements from a specimen and each reference material within the library. If the spectra from a specimen and a reference material perfectly match, the MSD is zero. As differences increase, the MSD increases as well. As soon as one starts measuring a specimen, the LCD displays the name of the one or two best matches from the library and the MSD for each. Thus, calculated and displayed in realtime, is a statistical metric of the matches between a specimen and the reference population. One weakness of this method is the pre-set list of 24 elements: Sb, Sn, Pd, Ag, Al, Mo, Nb, Zr, Bi, Re, Pb, Se, W, Ta, Zn, Hf, Cu, Ni, Co, Fe, Mn, Cr, V, and Ti. These elements allow identi fi cation of over 400 metal alloys, but many are not particularly useful for obsidian. Nb and Zr are very useful for obsidian sourcing. Ta and Hf, for example, can be useful for obsidian but occur at concentrations that require the sensitivity of NAA. Sr and Rb, however, are missing, and elements such as W and Ni are unlikely to yield much discriminating power. At present, elements cannot be added or removed to the Spectral Fingerprint mode (but this may change in the future). Nevertheless, we felt that this method was worth exploring due to the multitude of elements used to derive the fi ngerprints, following Harbottle ’ s approach. Another drawback is that X-ray intensities are not converted into element concentrations. Measurements are reported in cps/ m A (i.e., X-ray counts per second per microAmp of X-ray tube current) rather than parts per million (ppm) or weight percent. This simpli fi es the algorithms for realtime calculations against potentially hundreds of reference materials. Because the Spectral Fingerprint mode is comparative, only measured X-ray intensities are recorded, and no matrix or overlap corrections are made to the measurements. Nevertheless, this method is quantitative because all X-ray intensities for the references and specimens are recorded, statistically matched, and available as a spreadsheet for docu- mentation or further statistical testing. Measuring and reporting X- ray intensities was once standard (e.g., Ambrose et al., 1981; Brown, 1983; Godfrey-Smith and Haywood, 1984; Shackley, 1988), and most of the foundational obsidian studies are based on intensities (e.g., Jack and Heizer, 1968; Jack and Carmichael, 1969). Today, however, reporting X-ray intensities is commonly considered insuf fi cient (e.g., Shackley, 2005), but while uncommon, it is not unheard of (e.g., Astruc et al., 2007). We concur that, if possible, geochemical measurements of artifacts and specimens should be fully quantitative and generally compatible with data collected by archaeometric laboratories. As noted in Section 4.2, all artifacts and geological specimens were previously analyzed using fully calibrated pXRF. For our specialized applications, however, we are willing to consider a method based on X-ray intensities rather than elemental data, and as noted in Section 3.2, there already are distinct technique-based communities of practice in obsidian sourcing. Thus, while we are interested in the long-term utility of our raw data, our primary goal is pro- ducing correct source identi fi cations useful to our team during excavations and surveys. Lastly, metal identi fi cation commonly involves fl at surfaces and, for alloys like stainless steel, minimal surface alteration. It was unknown at the outset of this study how the curvature and weathering of artifacts would affect our results. Using ceramics, pXRF tests by Forster et al. (2011) demonstrate that elements are not equally affected by surface effects. Light elements, like Si, Ca, and Ti, are strongly affected by convex, concave, and grooved surfaces, whereas heavy elements like Rb, Sr, and Zr are largely im- mune. The same trend was reported for surface coatings. Therefore, if suitable elements dominate the fi ngerprint in Spectral Fingerprint mode, the results could be largely unaffected by surface irregularities. The instrument was “ taught ” twelve obsidian spectra for this test (Fig. 1): Pokr Arteni, Mets Arteni, Gutansar, Hatis, Geghasar, Syunik-1 (Pokr and Mets Sevkar), Syunik-2 (Mets Satanakar and Bazenk), Tsakhkunyats-1 (Arkayasar/Kamakar), Tsakhkunyats-2 (Ttvakar), and Tsakhkunyats-3 (Damlik). Each type was represented by one polished specimen characterized by NAA, XRF, and EMPA. Although the software has the capability to store hundreds of fi ngerprints, we used one specimen per obsidian type for two reasons. The fi rst was speed. Obsidian sources are, with few ex- ceptions, geochemically homogeneous (e.g., Glascock et al., 1998: 18, Hughes, 1998: 107, Shackley, 2008: 198), and with one specimen each, we taught the software all twelve obsidians in less than 20 min. We could pick up a new instrument and use this method in minutes. Second, the software reports only the two best matches, so using two or more specimens would mean not knowing the second-best match. When the fi ngerprints are plotted (Fig. 3), our data suggest that most elements provide a degree of discrimination. This might be due partly to the uncorrected nature of these data (i.e., they are unadjusted for interferences from other elements). This is sometimes due to an element already in the fi ngerprint. For example, Co appears to vary considerably, but the measurements likely re fl ect interferences from Fe X-rays. In contrast, the Bi data likely include interferences from Ba, an element useful in obsidian sourcing. Thus, fi ngerprints include subtle contributions from elements beyond the pre-set list, and such interferences become fortuitous (e.g., Ba is not included in the pre-set list of elements, but it nevertheless contributes to the overall fi ngerprint due to its uncorrected contributions to the Bi X-ray intensity measurements). First, 485 geological specimens were analyzed once for 10 s. Of these specimens, 459 corresponded to the twelve Armenian obsidians for which the software had fi ngerprints in its reference library, and 26 came from Georgia and Turkey. All specimens were analyzed on unprepared surfaces, which varied from smooth fl ake scars to weathered exteriors. This test used a cardboard stand, designed to be lightweight and duplicated anywhere in the world. Second, the 154 artifacts were measured for 10 s, and the instrument was used as it would be in the fi eld. Operation was entirely handheld without a laptop, electricity, or test stand. Analyses were started by the trigger and ended automatically after 10 s. Positioning was aided by the instrument ’ s internal camera. The artifact surface selected for analysis was simply the one considered least likely to puncture a fi lm over the measurement window. Analyses were conducted outdoors in conditions that varied from direct sunlight to light rain. Fig. 4 demonstrates what the pXRF displays. For Gutansar and Hatis, the match was suf fi ciently de fi nitive that only the best fi t and its c 2 value were displayed. For the Pokr Arteni specimen, the correct source is marginally better than the B-ranked Mets Arteni. For a Syunik-Sevkar specimen, Syunik-Satanakar is a close second- best fi t. The Tsakhkunyats-2 specimen is correctly identi fi ed, and Tsakhkunyats-3 is ranked as the next-best match. Table 3 summarizes the A- and B-rank matches and their c 2 values for the 459 Armenian geological specimens, and Supplementary Table S1 gives specimen-by-specimen results. Only seven were misidenti fi ed. Five specimens from Mets Arteni were matched to Pokr Arteni, just 3 km away (with Mets Arteni as close B-rank matches). A Syunik-Sevkar specimen was matched to Syunik-Satanakar, and vice versa. For both, the c 2 values were virtually identical. To establish how the software would handle obsidians not in the library, we analyzed 26 specimens from Georgia and Turkey. Table 4 summarizes the A- and B-rank matches and their c 2 values for these specimens, and Supplementary Table S2 gives specimen-by- specimen results. The Nemrut Da ǧ specimen yielded c 2 values so high that the screen displayed “ No Match, ” and the Bingöl B specimens had c 2 values above 1, indicating poor matches. The other nine Turkey specimens had c 2 values between 0.43 and 0.92, and their A- and B-rank matches are anomalous compared to Table 3. For example, the Sar ı kam ı s ̧ specimen had an A-rank match of Syunik-Sevkar, but the B-rank was Pokr Arteni, not Syunik- Satanakar. Hence, A- and B-rank matches that make little geochemical sense can suggest a mismatch. Georgian specimens were matched to Tsakhkunyats-3 and Hatis, not Chikiani, and their c 2 values may be too small to be recognized as erroneous. Therefore, we propose including a Chikiani fi ngerprint in future work. Table 5 summarizes the A- and B-rank matches and their c 2 values for the 154 artifacts, and Supplementary Table S3 lists artifact-by-artifact results. Generally the artifacts have higher c 2 values than the geological specimens. This is likely due to a com- bination of artifacts ’ uneven, weathered surfaces ...

Context 4

... as they dig. Our methods offer a new way to make fi eld observations because, as demonstrated by our tests of visual classi fi cation, volcanic source is one tangible aspect of the material culture not apparent while excavating. As a routine component of fi eld practice, on-site obsidian sourcing may yield insights when such information would be immediately relevant to excavators and their interpretations. We are also interested in insights from obsidian sourcing during surveys. Surveys are the principal means not only to search for and choose sites for excavation but also to reconstruct the distribution and organization of groups across the landscape. Obsidian sourcing during a survey may offer insights into settlement type, resource management, and other spatiotemporal patterns that inform how we approach surveys and document sites. Consider Kuhn ’ s (1995) provisioning strategies for mobile foragers. Resource planning must account for the frequency and predictability of moves, variety and abundance of foraging opportunities, and availability of replacement raw materials. Kuhn describes two approaches to maintain tool supplies: individual provisioning and place provisioning. Individual provisioning is a strategy in which there is uncertainty about mobility, resource availability, and opportunities for reprovisioning with raw materials, while place provisioning is a strategy with lower mobility and greater predictability in resource availability and distribution. Thus, raw-material transport and management strategies differ based on mobility and landscape knowledge, and strategies on a particular landscape may vary diachronically due to climatic shifts. Therefore, the ability to rapidly identify local and non-local obsidians during surveys has great interpretive potential. At sites just a short distance from an obsidian source, artifacts from only that source may indicate a strategy primarily in fl uenced by place- provisioning. In contrast, sites with artifacts from diverse sources could re fl ect a predominance of individual provisioning. In reality, both provisioning strategies were almost always at work. For example, foragers may have used individual provisioning while traveling from the Syunik region to Gutansar, but while residing near Gutansar, they shifted to primarily place provisioning. It is not a simple either-or proposition, and the palimpsest issue remains. Thus, large assemblages sourced by pXRF and with detailed spatial and stratigraphic data are key to unveiling signi fi cant patterns. Additionally, recognizing the spatial distributions of obsidians at a site could inform test pit locations or reveal different activity areas. Contrasting artifacts from the surface and levels of a test pit or trench add a temporal dimension, and correlating source data with artifact typology may uncover otherwise hidden diachronic changes in mobility between periods. At numerous open-air Palaeolithic sites in Armenia, little more is preserved than, in some instances, millions of obsidian artifacts. With a small team equipped with several pXRF instruments to cover a large area, working at such sites has become practical using the methods we report here, and the resulting datasets would be suf fi ciently nuanced to make meaningful behavioral interpretations. Thus far pXRF-based obsidian studies have been primarily focused on sourcing artifacts at museums, universities, and archive facilities (i.e., collections previously beyond the reach of analytical techniques). Most studies have been conducted in museum labs and similar contexts (e.g., the Smithsonian ’ s Museum Conservation Institute in Phillips and Speakman, 2009; Craig et al., 2010; Field Museum of Natural History ’ s Elemental Analysis Facility in Golitko et al., 2010, 2012; Millhauser et al., 2011). For Craig et al. (2007), “ in situ analysis ” refers to measurements taken “ where the artifacts are stored. ” Despite increasing use of pXRF, obsidian sourcing remains embedded in post-excavation studies. When artifacts are sourced, they have been detached, often distantly, in time and space from their archaeological contexts. For example, Forster and Grave (2012) analyzed 26 Near Eastern artifacts, excavated during the 1930s, at an Australian museum. This division has been the domi- nant paradigm since the 1960s, and obsidian sourcing has been highly productive over the last fi ve decades. Nevertheless, use of pXRF analyzers in laboratories clearly does not take full advantage of their capacities. Our aim is to shift obsidian sourcing from the realm of “ white coats ” in the controlled laboratory environment to “ muddy boots ” in the fi eld. Before reporting the development and testing of our methods for rapid obsidian sourcing, we brie fl y discuss obsidian studies in Armenia and two different approaches to classi fi cation in obsidian sourcing, revealing two technique-based communities of practice. Given the abundance of obsidian sources in Armenia (Fig. 1), it is beyond the scope of this paper to discuss them in detail. Badalyan et al. (2004) synthesize Armenian sourcing studies prior to 2000, including that of Blackman et al. (1998) and research never fully published. Until recently, most source characterization was done with fi ssion-track dating and neutron activation analysis (NAA; e.g., Oddone et al., 2000; Badalian et al., 2001; Chataigner et al., 2003; Kasper et al., 2004; Cherry et al., 2010; Meliksetian et al., 2010, 2013). Rarely XRF has been used (e.g., Keller et al., 1996 used both NAA and XRF). Chataigner and Gratuze (2013a) recently studied Armenian obsidians with laser ablation inductively coupled plasma mass spectrometry (LA-ICP-MS), characterizing most source complexes based on four to eight specimens each. Geologically focused studies include Keller et al. (1996) and Karapetian et al. (2001). Relatively little archaeological data have been published regarding obsidian sourcing in Armenia, especially for the Palaeolithic. Most research has focused on the Neolithic to Bronze Age (e.g., Chataigner et al., 2003; Badalyan et al., 2004; Kasper et al., 2004; Chataigner and Barge, 2007; Cherry et al., 2010; Chataigner and Gratuze, 2013b). We are aware of obsidian sourcing at only one Palaeolithic site, Upper Palaeolithic Kalavan-1. Chataigner and Gratuze (2013b) have analyzed 18 artifacts and assigned them to Hatis (10), Gegham (4), Gutansar (3), and Satanakar in the Syunik complex (1). They also analyzed 20 artifacts from Mesolithic/proto- Neolithic Apnagyugh-8 (referred to by its unof fi cial name Kmlo-2) and attributed them to Gutansar (10), Tsakhkunyats (4), Arteni (3), Hatis (1), Gegham (1), and a source in the Kars province of north- eastern Turkey (1). Thus, hundreds of millennia of obsidian use by modern humans and other hominins has hitherto been largely unstudied in Armenia. There are two schools of thought regarding classi fi cation in obsidian sourcing, and our two methods re fl ect these distinct approaches. Neither approach is fundamentally correct, and in instances where both approaches successfully discern sources, there is no clear reason to choose one or the other (unless there is a bene fi t in time or cost). The fi rst approach is summarized concisely by Harbottle (1982): “ the more elements, the better ” (18). He sees sourcing as a form of taxonomy, applies taxonomic theory in his approach, and argues that classi fi cations based on numerous traits are superior to those based on just a few. Harbottle cites the foundational book Numerical Taxonomy by Sneath and Sokal (1973), whose fi rst taxonomic principle is: “ the greater the content of information in the taxa of a classi fi cation and the more characters on which it is based, the better a given classi fi cation will be ” (5). Thus, Harbottle claims that one tends to “ get the best classi fi cations out of the most information ” (39). Similarly, Glascock (1992) maintains “ it is advisable to use the information on all elements... to use the maximum amount of information ” (17), citing Sneath and Sokal (1973). Consequently, Glascock and colleagues often use 20 or more elements in their source attributions (e.g., Ericson and Glascock, 2004; Glascock et al., 2007). The second approach involves critical selection of elements. Hughes (1984) holds it “ is not necessarily the case... that the in- clusion of larger numbers of variables... results in a ‘ better ’ classi- fi cation ” (3). He contends poorly measured, redundant, or uninstructive elements should be excluded. Citing Hughes (1984), Shackley (1988) holds “ improperly chosen ” elements can “ covertly skew the classi fi cation analysis ” and maintains “ the best course is to choose elements for analysis carefully ” (763). Hence, Shackley (1995) proposes “ a rule of thumb... is to use the fewest variables necessary ” (546), and he frequently relies on four key elements (e.g., Ba, Rb, Sr, Zr in Shackley, 2009; Fe, Mn, Zn, Zr in Negash and Shackley, 2006). These approaches re fl ect different technique-linked communities of practice. Harbottle and Glascock are nuclear chemists who principally use NAA, while Hughes and Shackley have archaeological backgrounds and use XRF. What constitutes “ standard practice ” differs between the NAA and XRF communities. These differences are based on the strengths (and weaknesses) of each technique. NAA typically measures more than 20 trace elements in obsidian; however, early XRF instruments involved manually tuning a spec- trometer to each element, so measuring fewer elements was desirable. By extension, methodological variations in pXRF use should be acceptable if those differences facilitate its strengths (e.g., fi eld use). Our 10-s measurements take advantage of elements measured simultaneously with one X-ray fi lter, the development of faster detectors, and the relationship between measurement uncertainty and time. The tests we report here involved more than 500 geological ...