Eugen S. Teodor
National History Museum of Romania, Bucharest

In the accustomed curators' language, we learn about "gold", "silver", "bronze" (a.s.o.), as raw materials for metallic artifacts. As soon as the archaeologist enters the world of physics, the prior "taxonomy" becomes obsolete. Dealing with intricate metallic compositions, in large series of analyzed archaeological objects, as happened in Archaeomet Project, it became obvious the necessity of specialized tools dedicated to the elemental compositions.

In the fall of the last year, facing the elemental results for almost 200 medieval "silver" coins, it also became obvious that the graphic presentation of those results will be unreadable. Two options where therefore left: first - to study the most competitive methods for multivariate analyzes, to decide which is the best, to buy the product, to learn the methodology, to implement the model...; second - to figure out a method to deal with the results, a new method, oriented directly towards the needs. Time and money evaluation decided in favor of the latter.

The chosen method is not quite "original", a variant being previously tested for pottery shapes (which will be presented, only one week later, in the International Congress for Medieval Studies, Kalamazoo, USA). The only one essential difference is that pottery shapes are classified against a regular number of characters (8), while the elemental analysis works with an undetermined number of variables (5 or as well 25).

Briefly, the method implies the next steps:
• 1. to establish a definite lot of artifacts of the same type (with the same predominant element, for instance silver), measured with the same technique;
• 2. to establish how many metallic elements were identified inside the lot (for instance 7, 13, or 16);
• 3. to establish the range of figures for each element;
• 4. to transform each range of values (310-991; 4-22; 0.2-3.56) to a range from 0 to 100;
• 5. to pick a target (first object in a table) and to compare its ranges, for all elements, with all the other objects (the comparison set);
• 6. the differences between the "target" and the "comparison item" is accounted as a sum of all differences divided to the number of elements (that would be the average gap); these results are transferred into a new table;
• 7. the average gap figure is classified within the next thresholds:
  • 0 = X < 0,1 = identity
  • 1 = X [0,1...0,5] = very close resemblance
  • 2 = X [0,5...1,5] = close resemblance
  • 3 = X [1,5...3] = loose resemblance
  • 4 = X [3...5] = very loose resemblance
  • 5 = X >5 = no analogy
• 8. to pick the next target (the second record in the table) and to compare it with all the rest of the objects;
• 9. the process of relationship classification is closed when all objects are classified in relation with all the other objects;
• 10. the results are further processed in a serial table, looking for "analogical groups" and giving the final "group number classification".

The comparison result "0" (zero) is theoretical; the results "1" encounter only for very accurate measurements, or by chance; the results "2" are pretty rare and theoretically point out a close resemblance, for artifacts allegedly produced in the same place, with the same tools, raw materials and technology; the results "3" might point out to related artifacts, like the "raw material source" (e.g. one coinage set) to re-melted metal objects (e.g. adornments), but the possible interpretations vary pretty much; results like "4" point out to a distant relationship, like a "raw material source" and its derivates resulted by mixtures with other materials; "4" is a relationship occurring between well defined and homogenous groups of artifacts.

The fact that we can't use any results over 5% in the analogical judgment has to warn us about the accuracy of the record on the physicist laboratory. The tolerance of ±3% is catastrophic for this kind of analogy seeking, because a "very close resemblance" (for example 0.45% "average gap") can be reported as "very loose resemblance" (3.45%). Or, how many technologies or devices available today provide less then 3% error? ...

For now there is only one way to get out from the "mice trap": to repeat the reading several times and to record each reading. The average of ten readings might reduce the risks environ 10 times; a 5% technological error might be reduced to around 0.5%, which is almost all right...

Even so, the results from the test-lot from Moldavian coinage (supplied with some contemporary foreign coins, as comparison terms) show some credibility. Some groups resulted by average gap analysis are quite homogenous; for instance, from 8 items in Group 1, 7 came from Teius hoard; from 8 items in Group 5, 6 came from Iasi hoard; from 15 items within Group 8, 10 came from Alexandru I, and the rest from Petru I. This homogenity couldn't ever occur if the laboratory results would be unreliable, or the computing wrong.