If you’re reading this by email subscription please note, yesterday’s post has significant documentation updates on the process.
So earlier in the week I spent a great deal of time on RRC 22/1 and produced an interpretive die map after having used Co-Pilot to try to come up with the die clusters. Today I’ve used Claude to re do this work to demonstrate how much more robust it is as a tool and to see how it changes my thinking of my own work. One of the things I like best about it is how it tells me what it is doing. The artifacts it produces are far more useful, intelligible, checkable. I find this useful for running thought experiments about possible manufacturing practices.
Let’s start at the end.
Prompt:
How long did it take to reach this point?
[of course this was 2.21 pm EST, but the robot doesn’t live in my time zone.]
Prompt:
estimate the amount of electricity and water this co-work session used
The other cost of course is my allowance of usage. This is my second session. I used pretty much my whole limit yesterday testing functionality not just on coin stuff but also for coding and creating apps.
Claude has session limits calculated on 5 hour increments.
And these all go to my weekly allowance:
52 weeks in a year at 200 dollars a year means $3.85 per week. So yesterday and today together cost me 50 cents.
I began my session by pasting in the briefing from yesterday’s session (you can read that on yesterday’s updated blog post). I need to remember in future to start by telling it that I’m giving information that I don’t want it to use yet. It tried to start working straight away. I selected Opus 5 and told it to use High but not Max effort.
I used co-work rather than chat and started with a much more complex prompt:
The above information is for context and method. The next challenge is to fetch the die links for http://numismatics.org/crro/id/rrc-22.1. Produce aBipartite die-link graph that crossing minimisation, and then create and artifact that show the minimum number of anvils necessary to have no overlapping lines. Then produce an artifact that assumes the reverse dies were produced in the ancient greek alphabetical order of their labels, with the double letter series coming after the the single letter series. Take the time you need to do this task most efficiently with the least usage cost.
As you can tell from the time log above I made a mistake by not telling it where to put AB in the sequence. I also included more human language than is necessary. The strike outs mark what I would exclude in future.
Prompt writing is a skill I will continue to develop. And I enjoy considering questions of data visualization. It is high usage to have Claude fetch its own data or search the web. I may try designing my data set and then uploading it with a careful prompt. I won’t spend much time asking it questions about my electricity and water usage and that in and of itself was a really complex ask. That said I value knowing my time today was about one dishwasher cycles worth of resources. My family of four runs a cycle every day, sometimes two or even three if we are entertaining (as we don’t use disposal materials unless compostable).
Intellectually these images have caused me to lean away from a two anvil model. Instead I think it most likely there was a single anvil and a die box filled with pre cut dies set into punches ready to be used and after a work session reverse dies were returned to a box while the obverse die was changed and then striking resumed.
Why do I think this? If we ignore alphabetical order there are very few necessary crossings in the die clusters:
There are many more crossings if we force alphabetical order, but I think that we can still imagine a one anvil set up with a die box.
I’m still struck how different RRC 25 is from 22, 26, or 27 in its die clusters. I think it was manufactured using a different logic of work flow which I want to think more about it.
I also want to think about how this process seems to prioritize NOT using dies to breakage and fairly regular changing dies.
I wonder if we would find similar patterns of manufacturing in issues struck by Campanian mints?
I did it. I paid for a subscription for Claude. I will never use co-pilot again. Claude is far superior.
I tested it on RRC 27/1. I’m not showing you all it gave me but rather portions of the three artifacts it produced based on minimalist prompts. It took only four very simple prompts (included below) and less than 30 minutes. I am not showing the first artifact as it was not useful beyond under standing the number of clusters at play. And the images here are only parts of the whole artifact produced. The parts I selected as most illustrative and meaningful.
dies within each column are reordered — using a barycenter heuristic, iterated to convergence — to pull linked dies toward the same row and push unrelated ones apart, cutting down how often lines cross
This was drastically faster than my reasoning along side copilot’s clustering for RRC 22/1.
THIS IS NOT NECESSARILY THE ONLY SOLUTION. BUT it does help us see how production MIGHT have been done. And that to allow for a die box of shared reverses we need a minimum of three anvils before the sharing if feasible, assuming of course the die ids are correct.
Claude says:
That a fully non-overlapping arrangement exists with three anvils, where none existed with two, is a real structural difference in the evidence; it makes three anvils a live possibility here, not a preferred conclusion — a single anvil reusing dies over time fits the same pairings equally well.
I as a numismatist cannot imagine a situation where dies were taken out of the anvil, saved and then returned to service later to be used with other reverses. If it were ever done at all I would only see that as useful in a strange slow minting operation where there were dies in the storage over a long period of time. It is a poor model for this actual data set.
Chain B only needs two anvils to untangle the data.
We can imagine either a slow ramp up of production or we could imagine an intense early phase of striking that then slows as the target quota becomes closer.
Again other solutions are possible this just lets us see one possible reconstruction.
What I can say is that RRC 25/1 looks very different in its patterns of manufacture and certainly did not require two anvils let alone three. This complexity and the long chains is closer to what we saw with RRC 22/1.
When I do similar work I will use Claude Co-Work as it runs in the background. These problems are time intensive to do in the simple interface. If you are playing around with similar models I’d LOVE to hear what you’ve tried.
Andrew Riggsby raised good points on FB.
I wonder how much of the Claude session is appropriate to share. Is there ever too much? It is possible of course to download the whole thing. Or are just my prompts what would be most useful.
Here is a PDF of the whole session briefing produced by Claude that would let anyone reproduce and continue this work if they uploaded to Claude themselves.
The prompts that I wrote and thus are my original contribution:
Prompt 1:
This website has a list of die links. Create a Create a Clustered Bipartite Die-Link Graph and show me a preview image. http://numismatics.org/crro/id/rrc-27.1
Prompt 2:
produce a near identical graph but optimizing the order of the dies to show as few over lapping lines as possible
Prompt 3:
Theorize that there were two anvil work stations with obverse dies operating at the same time with a shared pool of reverse (punch) dies used to strike the coins. Put reverse dies in the middle and create lines without over laps to obverse dies to the left and right.
Prompt 4:
Same problem, stay focused on chain A and create a visual representation reflecting a hypothesis that three anvils were in use.
These are the prompts that got the data visualizations at the top of the post. If you open the briefing you’ll see I continued with RRC 26/1 (two anvil hypothesis was sufficient to remove overlaps). My intent is to next run RRC 25/1 and 22/1 as I did the first manually and the second with copilot clusters.
None of this is publication quality yet as I would want to triple check everything first.
This morning I did a fun and fast visualization of the die clusters of RRC 25/1. You’ll have to wait until the final publication to see those.
I thought to myself is this similar or different than other striking patterns. I thought RRC 22/1 would be an ‘easy’ point of comparison. It wasn’t easy or fast, but it was interesting and v v different.
At first I was going to ignore the alphabetical sequence of the Greek control marks, but as I looked at the clusters they grouped so nicely around letter sequences that I felt I had to use it.
My first explanation is that there was one anvil working through the letters in alphabetical order with the ‘reverse’ control marked die in the anvil (thus the true obverse!) and that there was a die box next to the anvil with a number of punch dies ready to go.
Next steps.
Check ‘odd’ control marked dies, the ones that seem out of sequence (the greens above).
Check the overlaps to ensure the die ids are correct.
Inspect flans for curvature.
Consider if imitations, ancient or modern, may be effecting the results.
Figure out what the Leake collection is (supposed to have a Pi specimen)
Write Cambridge and ask if they really do have a Pi and if a photo could be sent
Redraw with Greek letters for clarity.
I wonder if Theta (singular) was skipped for its association with death (thanatos). Perhaps H was skipped for fear of confusion with Zeta?
If we do machine best fit. The three largest clusters show significant overlap in the linkages.
But there many single pairs and small chains as well. These I worked into the above clusters in the diagrams above.
If there were not sequential control marks, I likely would have concluded this was at least partially a two anvil one die box system of striking. And that might also be true…
The AB die suggests to me they were engraving dies on an as needed basis at least at the end of the series.
Finally I’ll note that the ratio of obverses to reverses is 1:1.17. This is surprisingly like the ratios we see at the end of the republic. Even as there are die chains and clusters they’re still averaging the same type of usage, near one to one. But still with more reverses, suggesting my idea of the control marked side being in the anvil die less likely.
So let us imagine two anvils and a die box with a selection of sequential control letters selected from the box with some attention to the letter selected but the work continuing in tandem with breaks causing sharing of reverse dies until the die box is exhausted and then refreshed with the next sequence of letters. The three clusters may suggest three points at which the shared die box was refreshed with new reverses. Quite often a reverse would be fully used up on a single anvil die or the two would be swapped at the same time for convenience of work flow, perhaps to rest one set of laborers during the change over.
The Romans struck a silver obol with RRC 13 in very small numbers (RRC 13/2), something that Burnett has shown fits well with the fashion for AR obols in Italy more generally. (earlier post). These coins are so small and so easy to lose it is possible more types, even Roman types might appear in future, just as we continue to learn of newer small denomination bronzes in the republican series.
Fractional silver doesn’t start up again until RRC 25 and 26 with drachma which continues into the Quadrigati series and then into the earlier denarii as well. Yes, I’m still worries about RRC 26, but this got me thinking about why specific denominations are struck and their utility. Clearly AR fractions are useful to the Romans or they would not endure as part of the production.
I’ve removed grossly overweight outliers from the CRRO data where it is clearly a cataloging error.
I find it interesting not only is RRC 26 still better regulated than other issues even here but its median and mean averages are closer to the quadrigati than RRC 25.
For all three issues the design of the drachm and didrachm is identical the only change is the size of the flan and die diameter.
This is where RRC 25, 26, 27 look most similar. There is no meaningful difference in the struck bronze weight standard. The AE Litrae has the same design as the silver didrachm. Here there seems a tight fit of the three issues.
If we preserve the logic that the ROMA legend issues must come after all ROMANO issues, I lean right this moment to putting RRC 26 struck issues after RRC 27 and before the quadrigati issues.
I’m still not convinced the cast ae of RRC 26 is really the same issue as the silver or struck bronze. Why must we combine them?! I see no reason. If this is the case I’m happy to push the cast part of RRC 26 earlier.
And then that half litra with the dog… RRC 26/4. It feels free floating, untethered to the other coins in the issue with which it is grouped. And look! I have a note next to RRC 17/1 saying Molinari and Burnett 2015 have associated the Roma/Dog RRC 26/4 with RRC 17/1 based on the Ponte Gini scattered hoard.
If I get hit by a bus someone tell my beloved to scan and share my physical copy of RRC so everyone can have my pencil annotations.
Just as I’ve concluded there is no meaningful difference in the weight standards of RRC 14 and 18 (and 19), I’m comfortable now saying the same between RRC 21, 24, 25, 27 AND 35. RRC 35 was made on a much larger scale, far beyond any of the earlier aes grave issues. I might do the histograms it, but not today.
Haeberlin’s averages are
As = 267.83g from 1168 specimens
Semis = 269.07g from 312 specimens
Quadrans 270.37g from 266 specimens.
The problem is that RRC 26 looks like a different kettle of fish. A higher standard, a smaller issue, no real connection to the silver. I think it is a problem that does not need solving like RRC 19 the scale of the issue makes it of little real meaning for the overall pattern.
Now we see the bronze is also out of sync. Also heavy, and made in smaller quantities and perhaps doesn’t even correspond with the silver at all. It doesn’t have a legend so we could push it between Group 1 (RRC 14, 18) and Group 2 (RRC 21, 22, 25, 27, and 35) and then call it transitional but we may just never know unless a few good hoards appear.
Messy.
Yet more and more I feel RRC 25 and 27 mark something new and interesting at the mint a real emergence of control and regulation. They are substantive and also perhaps more even. Their survival rates across denominations are less radically different that with some other issues. They are complete Uncia to As with no larger units.
What happened? Did the systems that were put in place for RRC 25 and 27 fall out of use because of scaling up of production? Perhaps.
You might have guessed that I’m doing a bit of pre writing here to augment another piece I’m working on. This type of blogging I find very useful as I try to decide what I think of the data.
Why did Crawford put this aes grave with the silver of RRC 26? The types are both revivals and they both have Apollo as the chief denomination but do they really go together?!
This data is so minimal it is hard to draw any conclusions. One new RRC 26 aes grave has appeared on the market and just adding in its weight changes the shape of the data, a nice reminder that this is not a very meaningful body of evidence from which to generalized.
I have not held this piece. I cannot authenticate it. The patina suggests it was recently unearthed likely in Italy. It is great shame we do not have the find data. Aes grave is not legal for export for just this reason.
But the data gets even MORE messy if we put in the specimen the Miles specimen in the ANS which is impossibly heavy. I’ve just about convinced myself that that acorn is really there behind Apollo’s head but I really don’t know what to think about this piece.
This minimal series looks like it might have been aiming at a higher weight standard, c. 290grams. The missing triens and uncia are strange indeed. A bit like RRC 19 which has only two cast denominations. I will only publish results based on Haeberlin numbers to keep my data set relatively consistent in it sourcing of evidence.
I’ve gotten faster now as I come to the end of the data extraction and processing. The As again shows the least variation and the highest average and median. The weight standard looks about the same as the Sickle standard but perhaps more tightly controlled across all denominations.
You can totally see why the semuncia stopped being made. What a radical difference in metal content. Again the As seems a higher weight standard and better controlled, but here the semis is more in line than with the wheel series. I see why Crawford decided to call this a 265 weight standard, but I don’t think it is out of the question it was thought of as on the same weight standards of these group 2 Aes Grave, slightly reduced from group 1 (RRC 14, 18, and 19).
I wonder if it was issued with RRC 22 even if there is no attempt to ensure they read as unified in the designs.
These certainly share a similar Roma design on the As and the didrachm.
The Roma of RRC 19 is markedly different
The RRC 24 Roma of the wheel series is certainly cut from the same cloth as RRC 21 and 22 helping me understand Crawford’s choice in relative sequencing to keep these issues close together.
I’m getting better at writing LLM prompts to extract data. I’ve not yet let it run my analyses as that feels like part of my own thinking at least until I’m confident at my methods and how I will check results. I want no magic black box to spit out answers. I do like less grunt work in transcription, data cleaning, etc… Copilot (university paid subscription) needs shorter sets of data but can handle multiple multiple pages of Haeberlin all at once and produce very usable PDFs with high accuracy (now that I’ve gotten better at prompt writing).
This is a weird picture. The As is much heavier than its fractions. I’m happy (ish) to ignore the tressis (1) and dupondius (14) but the As has 44 specimens and is clearly much more tightly controlled than the fractions as well as heavier.
Why does this series have no uncia?
My notes on Burnett 1989 say that he places this later in the relative chronology because of the standardized reverse on all denominations (cf. prow series) and CHRR 28. But I’m guessing the lack of the unciae is alway a good reason for bringing it down in the sequence of issues.
What should we use as the standard since it looks like the ass is significantly heavier than other denominations. The practical weight standard is 265 grams or so for the fractions, but 271 or so for the As.
I think the As is probably the best unit by which to judge the intended standard of production…
This work is building on my 2023 piece
“The Strangeness of Heavy Bronze” In S. Bernard, L. Mignone, & D. Padilla Peralta (Eds.), Making the Middle Republic: New Approaches to Rome and Italy, c.400-200 BCE (pp. 103-131). Cambridge: Cambridge University Press 2023.
I’m confident in Haeberlin’s weights based on my own reweighing in Berlin. See write up.
I’m corresponding with a very smart graduate student working on Bronze to Iron Age differences in the use of aes rude and thus have been talking about archaeological finds and how we date them.
She wrote me of Lucus Feroniae finds:
“… my dataset comes from the temple sectors excavated in 2000-2010. … In Serafin’s corpus there are no Roman denarii: the nine silver coins are all from southern Italy (Neapolis, Velia and Locri), while Roman bronze coins are often found together with aes rude.In Serafin’s corpus there are no Roman denarii: the nine silver coins are all from southern Italy (Neapolis, Velia and Locri), while Roman bronze coins are often found together with aes rude.”
“For Phase I, several primary foundation contexts contain Roman coins dated to 211-208 BCE. For example, US 79 has an uncia dated to 209-208 BCE (RRC 99,8 [= P series]), and US 93 has another uncia dated to 209-208 BCE together with aes rude. Serafin also records 89 coins from primary foundation deposits, with the latest issues reaching 208 BCE.”
This sent me to re read McCabe (and Russo) in RBN 2024:
RBN articles are restricted from online posting for two years so this one should be available shortly. If you need a copy now either McCabe or I can probably share one with you.
I agree with P being closer to Canusium that Luceria but likely a camp mint. McCabe does a great job mapping military action and regional coin finds.
The reason this P series coin at Lucus Feroniae in Phase 1 excites me is that the site was sacked in 211 BCE:
Here’s Livy 26.11 with the old translation wrongly using brass for bronze the Latin phrase is: aeris acerui = “a heap of bronze”
I’ve got to look at Serafin and consult McCabe! But that isn’t on the docket anytime soon, too many other tasks.
Liv Mariah Yarrow 