We present two machine-readable catalogues of Mohenjo-daro seals, transcribed by an AI model from the excavation reports of Mackay (1938, 709 rows with block/house/room and level) and Marshall (1931, 560 rows with excavation area and depth), together with text features coded from all thirteen of Mackay's seal plates (531 seals, 419 joined to find-spots, inter-pass reliability kappa 0.89), a held-out validated set of 90 distributional constraints on Indus sign sequences (1.08% violation rate on 2,585 held-out applications), and a reproducible pipeline of genre- and size-matched structural comparisons against Sumerian, Egyptian, Linear A/B, proto-cuneiform and Proto-Elamite corpora. Indus sign sequences carry non-first-order dependencies (7-8% excess conditioning to distance 5, where a bigram control collapses to 0.4%), condition as tightly as syllables within a Mycenaean Greek word, and show a text-final slot that is strongly preferred but not closed, replicating independently at Harappa. Genre-matched Egyptian title-and-name labels place their fixed element first and variable element last; Indus does the opposite. Text form does not vary with depth or find-spot across the complete coded corpus, and one earlier positive result failed to replicate on independent data. Ledger, container-label and astral/calendrical readings are not supported. No decipherment is offered or implied: the texts are too short, there is no bilingual, and the language is unknown.
Two seal catalogues AI-transcribed from the 1931/1938 excavation reports (OCR-cross-checked), text features coded from all of Mackay's seal photographs, 90 positional rules mined on half the corpus and validated on the held-out half, and controlled comparisons against six other writing systems. The results are mostly negative and are documented as such: no ledger, no closed name-set, no link between text and find-spot depth, and one of our own earlier claims failed to replicate independently. What stands: sign sequences carry dependencies a bigram model can't produce, the text-final slot is preferred-but-open and replicates at a second city, and a genre-matched Egyptian comparison shows Indus puts its fixed element LAST where Egyptian labels put it first. Not a decipherment — code, data and this note are attached, tear it apart.