Testing Head-Conditioned Terminal Selection in the Indus Script Using Held-Out Formulae

Preprint — not peer reviewed. This study tests whether the strong association between the immediately pre-terminal sign and Indus terminal signs 520/740 can be explained solely by repetition of complete inscriptional formulae. Using a frozen leave-entire-scaffold-out analysis, terminal choice is predicted correctly in 701 of 726 eligible held-out cases (96.56%). None of 10,000 terminal-label permutations reaches the observed accuracy. The effect also transfers between Mohenjo-daro and Harappa, survives seal-only and sign-family exclusions, and independently replicates in a small CISI/Parpola transcription (55/55). The result is structural, not a decipherment. It does not assign phonetic or semantic values to signs 520 or 740. The broad relationship between preceding signs and terminal selection has antecedents in prior Indus-script research and was recently demonstrated independently by Bourdeau (2026). The candidate contribution of this study is the controlled leave-entire-scaffold-out test designed to determine whether the dependency generalizes beyond repeated formulae. Analysis code, exception audit, corpus provenance, and independent-replication materials accompany the preprint.