A Computational Decipherment Hypothesis for the Indus Script: 185 Proto-Dravidian Readings Validated Across Two Independent Corpora

We present a computational decipherment hypothesis for the Indus Script (~2600–1900 BCE) proposing 185 corpus-attested Proto-Dravidian phonetic readings that cover 92. 8% of the Holdat Indus Valley Seal corpus tokens (6, 501/7, 002). Readings were derived through DEDR-based simulated annealing with anchor amplification across multiple evidence layers including distributional profiling, Elamite cognate matching, and allograph correlation. We validate the hypothesis through six independent tests: (1) an anchored bigram discrimination test showing Dravidian language models fit the readings significantly better than Uniform baselines (57. 8% vs 0. 0%) ; (2) corpus-independent replication on the Mahadevan 1977 concordance (70. 5% Dravidian hit rate) ; (3) 80% agreement with Parpola's (1994) independent iconographic-rebus proposals across 20 tested signs; (4) reading-level conditional entropy of 4. 11 bits, falling within the range of natural languages; (5) 97. 7% inscription uniqueness consistent with a registration-code or guild-identity administrative model; and (6) 76% Proto-Dravidian phonological inventory coverage. The Sanskrit hypothesis is falsified (0/34 agreement with Yajnadevam 2024). Three pipeline bugs were discovered and corrected during a comprehensive audit, and three prior claims from v1/v2 were retracted. The hypothesis requires specialist Dravidianist review before any claim of decipherment can be made. All code, data, anchor inventory, audit trail, and release validation are open source at https: //github. com/BitConcepts/glossa-labErratum (v3): Comprehensive audit corrected two mass-assignment bugs (Phase 239 kur, Phase 312 kol) that inflated v1/v2 reading counts from 605 to the validated 185 corpus-attested readings (400 HIGH confidence total, 185 attested in Holdat). Grammar conformance claim (91. 8%) retracted after permutation null test proved non-discriminative. Token coverage corrected from 100% to 92. 8%. Parpola agreement upgraded from 50% to 80% after fixing Unicode diacritical comparison bug. All v3 numbers are from a single cold re-run on audited data (RELEASEVALIDATION. json).