Constructional Grammar and Semantic Constraints in the Indus Script

Structural recovery can constrain an undeciphered inscriptional system before its sound values are identified. This study develops a construction-aware account of Indus sign sequences through family-level corpus control, prediction across sites, matched sequence comparisons, and artifact-register conditioning. An earlier predictive analysis used 2,034 exact-sequence families; a subsequent quarantine yielded a reference of 2,022 families and 9,302 sign occurrences. Replaying recovered model code on the current reference decreases mean negative log-likelihood from 4.33935 to 4.32551 across 6,778 held-out token predictions, a gain of 0.01384 with family-bootstrap 95% interval [0.00979, 0.01789]. The contexts were selected during the earlier research program; the evaluation does not independently establish nomination-free generalization. A fresh recount of the reference reproduces a distinction between W520 and W740: excluding W033-conditioned W520 occurrences, W090 never immediately follows W520 in 136 cases, whereas it follows W740 in 69 of 994 cases. The contrast remains after excluding every W220-bearing sequence family and in a sensitivity analysis using disjoint family groups. Further results identify construction-sensitive numeral order, a directed quantitative hierarchy, and noun-class selection associated with resolution. The functional ledger assigns roles to 7,846 of 9,302 tokens, with complete assignments in 1,004 families; these measures describe model coverage rather than translation accuracy. Graphic and distributional evidence support a limited semantic layer, while specific grammatical labels remain interpretations. The resulting model supplies testable constraints for lexical and phonetic testing without identifying a spoken language or claiming a complete decipherment. The accompanying public supplement contains analysis code and reports; corpus-level inputs are withheld pending verification of redistribution permissions.