Shinsa team attribution methods

[Justin]

I was wondering if anyone has direct experience with the NBTHK shinsa team when it comes to judging authenticity/attribution for unsigned tsuba. I am interested in their methods.

 

For example, do they just study a piece and use their knowledge and experience to assign an attribution or is there more science to it?

 

They obviously discuss it amongst each other and come up with a collective attribution, however do they refer to references or is it just based on seeing many examples and understanding differences, e.g. size, Mimi, seppa-dai shape, design motifs, file marks, hammer marks, etc.

 

If anyone has first-hand experience watching a shinsa team in action, then I would be most interested in your observations.
 

Edited Thursday at 02:26 AM by Justin

[Tim Evans]

I don’t have the direct experience you specified, but I will put a few thoughts forward. You are asking about kantei procedure for mumei tsuba. There are some differences if the tsuba is kinko or tsubako.

 

Kinko is primarily soft metal and the focus of the art is on the carving. Most of the mumei kinko tsuba that would merit shinsa are from the Edo period where there is more documented about the artists and the schools, because a lot of them are signed, so unsigned work by the same artist or school can be compared. The carving technique details of the famous masters is known by certain experts, so it is possible to confidently attribute a name or workshop to a high quality but unsigned kinko fitting by looking at the unique style, the tool marks, how the tool was moved, etc. This is known as the teguse, or the “hand” of the artist. An example are the specialists who study Goto fittings. You sometimes see kanteisho stating, unsigned, but by Goto so-and-so. A shinsa team may or may not have such a specialist, but I assume in a Japan based shinsa, such a specialist could be called in to evaluate a high quality mumei candidate. This approach can also be applied to a some tsuba artists that worked in iron, such as the well known Higo groups that have distinctive sukashi and softmetal decoration. 

 

However, the majority of unsigned, average and low quality tsuba do not have characteristics that can be identified to a specific person, and so in shinsa they are assigned to broadly defined schools or categories. Particularly in the case of unsigned iron tsubako, most of these categories are of relatively recent invention, made up by tsuba scholars in the 19^th-early 20^th centuries in order to classify similar tsuba into groups. Because of the broadness, there can be multiple opinions about where a tsuba fits. For example, is it Koshoami or Kanayama? Is it Onin or Heianjo Zogan? This can lead to a tsuba having multiple papers with different opinions.

 

This existing unsigned tsuba classification methodology is mostly based on comparative analysis, logical inference, and oral tradition since there are few documented facts supporting the majority of these broad tsuba classifications. We know that the NBTHK utilizes a “kantei points” checklist approach when evaluating sword blades, so there may be a similar approach with tsuba. From a science perspective, there has been a few attempts to analyze the material of sword fittings, but this is expensive and non-destructive methods are limited.

 

A question here, is an NBTHK Hozon paper really worth much on an average quality mumei tsuba? I think in most cases no, because most collectors can classify their mumei tsubako and kinko into these broad categories themselves with some reference books and study. It may be worth submitting if you think the tsuba may make Tokubetsu Hozon, which is useful as an assessment of quality.

[eternal_newbie]

4 hours ago, Tim Evans said:

However, the majority of unsigned, average and low quality tsuba do not have characteristics that can be identified to a specific person, and so in shinsa they are assigned to broadly defined schools or categories.

 

Some additional thoughts on this in particular (it mainly focuses on swords but is applicable to tosogu and all kinds of historical art in general): https://web.archive.org/web/20210123052555/https://blog.yuhindo.com/fungible/

 

"The issue is that there is no additional information inherent in that object that can lead to a more exacting judgment. You know it’s Higo, I know it’s Higo, the judges know it’s Higo, everyone knows it’s Higo because that part is obvious. Past that, on this piece, we run out of clues. The collector is hoping the judges see something that he missed, but in reality, the collector didn’t miss anything. It has the whole story on its face. So a good judge will only judge so far as he can stick his neck out. If the item says no more than Higo, the judge says this without editorializing an opinion out of thin air."

[Infinite_Wisdumb]

I think this is a great use of AI

NBTHK scans all papers into LLM

Agent provides results

Simple, yes?

[Cookie_Monstah]

1 hour ago, Infinite_Wisdumb said:

I think this is a great use of AI

NBTHK scans all papers into LLM

Agent provides results

Simple, yes?

 

They would have to upload their whole archive into the LLM as the basis of reference material. Doing so makes the information no longer proprietary putting them out of business.

[Infinite_Wisdumb]

43 minutes ago, Cookie_Monstah said:

 

They would have to upload their whole archive into the LLM as the basis of reference material. Doing so makes the information no longer proprietary putting them out of business.

If they are doing it on a public instance, correct!

[GRC]

They can’t build an LLM with their own records because there are too many contradictions and errors in their attributions.

 

Anyone building an LLM for tsuba identification would have to start from scratch to train it with any degree of efficacy.

 

The current system of classification has too many inherent problems (such as not enough categories, and some major voids of information, as well as built in misinformation), so it needs some significant overhauling before building an LLM around the NBTHK’s records.

Garbage in = Garbage out

 

“We” would need to build a group of contributors who are each specialists in specific areas rather than a bunch of generalists making generalizations.

Ideally, a minimum of three specialists for each tsuba “category” (if you can even find that many specialists for each!).

Those specialists would need to reach a consensus, and be more able to potentially parse out some “subcategories”, maybe identify previously unidentified smiths based on workmanship, and maybe even outline a timeline of progression within a tsuba category.

You need people with depth of knowledge and understanding, not generalists who are indoctrinated and obligated to follow “what has already been said in the past”. 
i.e. A fresh perspective with an analytical mindset and a vast data set

vs

Perpetuating and reinforcing a flawed system based on narrow sets of classification criteria.

 

Lofty ideals and goals I know, but doing it with anything less will just build a deficient AI that will only be good at identifying the obvious and horribly inept at identifying anything other than the norm (hmmm…kind of like the NBTHK as it is now)

[Curran]

1 hour ago, GRC said:

They can’t build an LLM with their own records because there are too many contradictions and errors in their attributions.

 

Anyone building an LLM for tsuba identification would have to start from scratch to train it with any degree of efficacy.

 

The current system of classification has too many inherent problems (such as not enough categories, and some major voids of information, as well as built in misinformation), so it needs some significant overhauling before building an LLM around the NBTHK’s records.

Garbage in = Garbage out

 

“We” would need to build a group of contributors who are each specialists in specific areas rather than a bunch of generalists making generalizations.

Ideally, a minimum of three specialists for each tsuba “category” (if you can even find that many specialists for each!).

Those specialists would need to reach a consensus, and be more able to potentially parse out some “subcategories”, maybe identify previously unidentified smiths based on workmanship, and maybe even outline a timeline of progression within a tsuba category.

You need people with depth of knowledge and understanding, not generalists who are indoctrinated and obligated to follow “what has already been said in the past”. 
i.e. A fresh perspective with an analytical mindset and a vast data set

vs

Perpetuating and reinforcing a flawed system based on narrow sets of classification criteria.

 

Lofty ideals and goals I know, but doing it with anything less will just build a deficient AI that will only be good at identifying the obvious and horribly inept at identifying anything other than the norm (hmmm…kind of like the NBTHK as it is now)

 

For me, this is the best answer.

I share similar views with Glen.

 

 I have openly questioned if the NBTHK started using database image matching software for their opinions in the last 10 to 12 years.

I've seen too many errors where I can understand the attribution to [ABC] from a top down digital photo that might help speed up the shinsa process, --- but anyone handling it will realize iron, patina, small tell-tales of manufacture, etc.--> all point to [XYZ] a different school.

 

The topic of AI trying to do tosogu shinsa is a very complex one.

A lot of the data it would use is very conflicting or corrupted.  Ex: I have mumei tsuba with 2 or more NBTHK papers giving very different attributions: sometimes decades apart, but often only a few years apart.

If both sets of papers are in the dataset?

 

Also:  Unless able to deep dive 3-D study a piece of tosogu, I am more of the opinion AI will get lost in its own righteous convictions.

Edited 10 hours ago by Curran

[Hector]

I was talking to an old Japanese tsuba collector and he told me he doesn't bother submitting anything that isn't "extra special" nowadays as he's noted an increasing upturn in either 'Shoami' (apparently a 'catch-all' attribution) or 'Horyu' (don't know) results.

Best,

Hector

 

[Iekatsu]

Glen's suggested approach is the only serious way for this field of study to progress. There needs to be a significant overhaul of the current system and how judgments are made. Ideally this would be driven by the NBTHK, but they are somewhat stuck between a rock and a hard place, because any major changes to the current system would potentially invalidate tens of thousands of issued papers/attributions, and honestly they are overwhelmed as is.

[Infinite_Wisdumb]

You have your answer @Justin

NBTHK can’t be trusted because they do not have a good process nor experts who know what they are talking about.

 

Great discussion gents!

 

[Justin]

Thanks for the feedback.

 

I personally value the work the NBTHK do. I have papers for 90% of my collection and the vast majority of these attributions are accurate (in my opinion). Even when I received a surprise attribution, after doing a bit of research I can see how they made their decision.

 

Hundreds of items are submitted to the NBTHK shinsa every month and I don’t think people would do that if they didn’t think the service was worthwhile.

 

In the end we are paying for someone to give and opinion. It could be a positive outcome or a negative one, but in the end it is just someone’s opinion. For me it is worth the small fee and I learn something from time to time.