01 / OBSERVED PATTERN

Even unread forms
have familiar places.

Some signs prefer the beginning; others favor the end. Position tells us something about use, without telling us pronunciation.

02 / STATISTICALLY REPRODUCED

Some neighbors
have an order.

Numeral-like signs occur before a conservative set of measure-associated signs unusually often. The association is observed; the numerical and metrological interpretations remain external hypotheses.

The direction test

A separate, train-only bigram model prefers normalized reading order in 79.00% of eligible unseen strings, averaged across folds.

03 / HELD-OUT PREDICTION

A missing mark.
A useful context.

Cover one sign in a previously unseen string. Guessing from nearby signs works better than always choosing the most common sign.

The model recovers constraints in a sequence. It has no verified dictionary, and this task does not test whether it understands the signs.

04 / THE LIMIT OF COVERAGE

Every token annotated.
No language recovered.

Giving every field a tentative label makes a useful index. It does not make the form readable. Rare signs often inherit their annotation from context and the five-role grammar’s assumptions.

The long tail carries little evidence

Exact duplicate-line share is 35.32%. 284 sequences recur across artifacts; 77 across sites. Repetition is evidence of recurrence, not independent confirmation of meaning.