Proposed Manuscript Chronology

[npcompl33t]

Hi @npcompl33t, thanks for publishing the data and the reproduce script. It made it easy to check this properly.

I ran your exact pipeline on the text of my skribe. First I checked I get your numbers on the real text (they matched) then on the skribe's one. Just to summarize what the skriba does: writes with fixed tables per hand × section, a random habit per sheet, and no time, drift or production order.

1. The horseshoe appears without time. The skribe gives aprox the same arch (0.27 vs 0.36 real), the W in MDS3 (0.44 vs 0.34) and the same section order on MDS1: Herbal A, Pharma, Astro, Herbal B, Stars, Text only, Bio. It also passes your triple test, more cleanly than the real text (1.7 vs 4 failing triples out of 35). It still does all this with the sheet habit turned off. So the horseshoe shows that the sections sit on an A→B continuum. It can’t tell us that the order is time.

As I mentioned in the other thread, Guttman horseshoes appear when there is a dominant sequential ordering. 

Its true 'time' may not be the dominant sequential ordering, but I have yet to discover such a signal in Text where that wasn't the case. I'm happy to look at any other examples. 

The only example text statistics exhibited a horseshoe was napiers casebooks, which were written as a log over several years, and the horseshoe verifiably tracked time in that case. 

I'm unable to think of another reason why text bigram statistics would contain a dominant sequential ordering -- I've analyzed over 100 historical reference manuscripts, and to-date the casebooks are the only Horseshoe i've found.

It seems far more likely to me that your model is somehow learning time, because I have yet to observe an actual instance where text could exhibit a horseshoe that does not involve time. Sections, topics, abbreviation conventions, orthography/srcibal differences, were all factors that i have analyzed over the course of these threads, and none produce this signal.