Beat Markov, or publish the caveat

The model plane is allowed structure. It is not allowed ontology. Grouped site holdouts are the honest test. A deeper net that cannot beat a first-order Markov baseline does not get a heroic caption.

Pass · Ontology leak scanBaseline = first-order MarkovGNN weights are not claimed as a production training run

Hold out Mohenjo-daro

test n 1202

Unigram 7.129 bits/token

Markov 7.012 bits/token

Pass · Markov wins

First-order Markov compresses held-out texts better than a unigram. That is structure, not a language ID, and not a trained GNN.

Hold out Harappa

test n 970

Unigram 7.043 bits/token

Markov 7.015 bits/token

Pass · Markov wins

First-order Markov compresses held-out texts better than a unigram. That is structure, not a language ID, and not a trained GNN.

Hold out Lothal

test n 80

Unigram 6.863 bits/token

Markov 6.935 bits/token

Fail · No gain

First-order Markov does not beat the unigram on this holdout. Deep structural patterns are not claimed to generalise.

What a green GNN would still not be

Even if a graph net beat Markov on grouped holdouts, the output would remain a structural prediction — masked-token and link prediction — with FDR-aware caveats. It would not authorise a DVM label, a tax reading, or a press sentence that “AI has deciphered Harappa.”