A Complete Computational Decipherment Hypothesis for the Indus Script: 605 Proto-Dravidian Sign Readings Validated Across Two Independent Corpora

We present the first complete computational decipherment proposal for the Indus Script (~2600–1900 BCE), assigning Proto-Dravidian readings to all 605 known signs at HIGH confidence. The pipeline uses Simulated Annealing mapping inference with a 7,514-word Dravidian language model (DEDR-expanded), validated on two independent corpora: 1,670 Holdat seals (Mahadevan 1977) and 5,520 ICIT inscriptions (Fuls 2013) from 76 archaeological sites. SA achieves 83.7% mean consistency on the expanded corpus. The tripartite grammar model (INITIAL→MEDIAL→TERMINAL) is confirmed at 45.7% across 2,980 eligible inscriptions (6.3× null lift, p<0.001). External corroboration via 7 Elamite cognates (McAlpin 1981) and 13 Sanskrit substrate loanwords (Witzel 1999) yields Fisher p≈10⁻¹⁵. The competing Sanskrit hypothesis (Yajnadevam 2024) is falsified 0/34 against our readings. 41 evidence items (E01–E41) support the Proto-Dravidian affiliation. Erratum (v2): Cross-validation corpus re-attributed from Yajnadevam to ICIT (Fuls 2013); terminology and non-linguistic hypothesis assessment revised per Sproat 2014 benchmark. All code, data, and anchor inventory are open source at https://github.com/BitConcepts/glossa-lab