Sequential Structure Without a Linguistic Diagnosis: A Falsification-Oriented Analysis of the Indus Script
Year: 2026
Type: preprint
Abstract: The Indus Valley script remains undeciphered. Under ordinary cross-validated scoring, real Indus sign sequences show a positive trigram-over-bigram predictive gain (the two preceding signs predict the next sign better than one preceding sign alone), replicating closely across two independently extracted digitizations of the same underlying corpus tradition, including the M77 corpus corresponding to the concordance behind the most cited statistical studies of the script (+0.143 and +0.158 bits), and clearing 200-trial null distributions entirely. However, both corpora contain substantial exact-sequence repetition (23 to 33 percent of sequences). When cross-validation folds are constructed so duplicate sequences cannot appear in both training and their own held-out evaluation, the gain reverses sign for both corpora (-0.088, -0.083); when duplicates are instead collapsed to unique sequence types, it weakens to near zero (-0.023, -0.010). A control experiment, grouping sequences of the same size distribution but without duplicate identity, stays strongly positive, ruling out the alternative explanation that this reversal simply reflects smaller effective training sets rather than duplicate identity specifically. This is not a data-handling error, since the repetition is a genuine corpus property, but it means the ordinary-CV result should not be read as established evidence of sequential structure that generalizes to novel sequences, and we treat resolving this as the paper's central open problem. Independently of that question, our most specific test among the mechanisms we examined shows the ordinary-CV gain is not uniquely diagnostic of the particular linguistic generative mechanism we tested, including after rebuilding that generator and a non-linguistic alternative to match the real corpus's own vocabulary and length statistics; neither reproduces the real magnitude in matched or unmatched form. A sample-size curve on two calibration languages shows this estimator rarely returns a positive result below roughly 3,000 sequences for either, evidence against the ordinary-CV result being generic small-sample behavior. We also find evidence against inscriptions being deliberately constructed as globally unique identifiers, though not against more elaborate identifier schemes. The framework, derived corpora where licensing permits, and all experiments are released openly.
Author: Bose, Joy
Institution: Ericsson (Sweden)
Language: English
Subfield: Language and Linguistics
Field: Arts and Humanities
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7212011102
Cites:
Cited by:
Related to: