A Calibrated, Anti-Hallucination Computational Benchmark for the Indus Valley (Harappan) Script
Live Interactive Explorer (EN / TR): https://indus.emare.dev/?lang=en · Zenodo DOI:
10.5281/zenodo.24383348· PDF Preprint:INDUS_DEDUCT_PREPRINT.pdf· Claim Ledger:CLAIM_LEDGER.md
INDUS-DEDUCT is an end-to-end, reproducible 13-layer computational epigraphy benchmark across 5,679 Indus inscriptions (5,805 parsed lines; 713 raw signs). Every pipeline stage obeys a strict anti-hallucination contract:
- Zero speculative upstream leakage: Upstream
sanskrit,translation, andnotescolumns are stripped prior to modeling and guarded bytest_01_zero_speculative_contamination. - Calibrated against real scripts & negative controls: Benchmarked at matched size and length (
N = 2,047) against Ur III Sumerian (CDLI), Linear B, Linear A, and a structured non-linguistic stamp negative control. - Zero fabricated phonetic readings: We explicitly separate verified structural/stratigraphic proofs (Tier A) from comparative/typological leads (Tier B) and retired/falsified claims (Tier C). See
CLAIM_LEDGER.md.
| Deliverable | Path | Description |
|---|---|---|
| Full Academic Preprint (PDF) | paper/INDUS_DEDUCT_PREPRINT.pdf |
11-page peer-review-ready preprint with 3 embedded high-res figures (Layers 0–12) |
| Executive Claim Ledger | CLAIM_LEDGER.md |
1-page Tier A (Proven) / Tier B (Supported) / Tier C (Retired/Null) referee map |
| Interactive Web Portal | portal/index.html |
6-tab live FaunaTech Doğal Atlas v2.0 browser (294 live TTF glyphs, 690 restorations, calibration & stratigraphy) |
| Open Damaged Restoration Catalog | reports/damaged_inscriptions_restoration_catalog.csv |
Calibrated Top-5 restorations for all 690 single-gap damaged inscriptions across 25 sites |
| Automated Verification Suite | tests/test_claim_ledger.py |
11/11 machine-verified invariants + SHA-256 integrity (DATA_MANIFEST.json) |
- Five-Way Calibration (Layer 8): Retires within-sequence conditional entropy
z_H1(Non-Linguistic Stamps = -245.34σvs.Indus = -108.85σ) and establishes three validated discriminators separating Indus and attested natural scripts from structured non-linguistic stamps:- Unseen-trigram generalization over position baseline:
Indus +6.38 pp (+0.432 bits),Ur III Sumerian +10.71 pp,Linear B +0.81 ppvs.Non-Linguistic -2.95 pp (-0.960 bits). - Kober bridge–ending
MI z:Indus +53.67σ,Ur III Sumerian +20.40σ,Linear B +18.47σvs.Non-Linguistic +0.07σ. - Blind 2-affix stacking (
Suffix:Prefix):Indus 207:1(4.96×right-edge bottleneck) vs.Ur III Sumerian 4:39,Linear B 22:11,Non-Linguistic 0:0.
- Unseen-trigram generalization over position baseline:
- Logo-Syllabic Core & Sub-Graphemic Grammar (Layers 1B, 5B, 7):
N_core = 240–271graphemes (95.7%tokens);: X :four-dash frame converts medial stems (2.9%) to terminals (62.2%, 8/8 pairs,p = 4.38e-45) and transfers zero-shot across unseen base glyphs (LOBO Δlog-loss = -31.3%). - 700-Year Paleographic Drift & Cross-City Sync (Layer 12): De-confounding Mohenjo-daro excavation datums (
n = 1,253within-area;DKG(S)n = 595) and replicating against HarappaPeriod 2–5(n = 637,Spearman rho = +0.289, p = 0.0376) uncovers two BH-FDR paleographic drift chains:#156 → #154(q = 0.0083) and the 3-stage sprouting chain#838 → #803 → #806(q = 0.0339,#806#1 late sign in Harappaq = 0.0106), while coreSEAL:Sgrammar remains frozen across 700 years (p > 0.09). - Document Registries & Maritime vs. Riverine Logistics (Layers 9–12):
- Mohenjo-daro Copper Tablets (
TAB:C, 99.6% MD-exclusive):Text ↔ Icon MI z = +19.66σ(96.2%purity) + 23 single-sign reverse equations (#341quadruped↔ 503 615 740), yet 4 tests refute bare animal nouns (H1) in favor of compound institutional/ritual formulas (H2). - Harappa Double-Sided Tokens (
TAB:I/B, 772 objects): FixedEntity ↔ Quantitylock (+9.55σ) and: X : + 400clitic stacking (95.2%,p = 1.16e-8). - Riverine Tablets vs. Maritime Cargo:
98.9%of tablets occur inland (enriched in freshwaterGavigharial62/66andFish47/47), whereas65.3%of cargo sealings (TAG*) and~66%of circular Gulf seals (SEAL:C) are coastal/maritime — anchored by Twin-Humans#091(0%final,11.89×trade/cargo liftp = 1.48e-6,42.35×in foreign corridorq = 2.86e-4).
- Mohenjo-daro Copper Tablets (
- Exact Phonetic Decipherment Boundary (Layers 4B, 5B, 8, 11A): Known-answer 23-seal Max-CSP locks
23/23(100%) of closed Linear B (+22.75σ) and Ur III Sumerian (+7.00σ) atk = 0, fails atk = 0under realistic 74% open Sumerian (z = -0.11σ), and recovers21/23(+4.97σ,68.2%shared signs) atk = 2bilingual anchors. Expanding foreign Indus seals24 → 42 → 55strictly increasesDOF(29 → 43 → 45), provingk ≥ 2bilingual anchors are required. - Leak-Free Visual-Neural LOCO Damaged Restoration (Layer 11B): Across
285real damaged inscriptions held out over12cities (STRICT_LOCO), our visual-glyph + neural MLP hybrid improvesedge_allTop-1 from53.5% → 57.5%(Top-5 82.9%, baseline10.2%) andinterior_allTop-1 from30.0% → 40.0%(Top-3 60.0%).
git clone https://github.com/emraresoftware/indusemare.git
cd indusemare
python3 -m unittest tests/test_claim_ledger.py -vThis immediately verifies all analytical invariants across Layers 0–12 (reports/*.json) and SHA-256 checksums for bundled data/external/* tables. To respect primary epigraphic concordance rights (ICIT / CISI), raw upstream corpus dumps are not redistributed inside the repository (DATA_MANIFEST.json records their exact SHA-256 hashes for full local verification).
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
# Place upstream raw corpora under data/raw/ and data/calibration/ (verified via DATA_MANIFEST.json)
.venv/bin/python src/layer0_ingest.py
# Strict 11/11 verification including all 17 DATA_MANIFEST.json SHA-256 hashes + SQLite checks
INDUS_STRICT_CORPUS=1 python3 -m unittest tests/test_claim_ledger.py -v
# Launch the bilingual (EN / TR) interactive FaunaTech Doğal Atlas portal locally
python3 -m http.server 4188 --bind 127.0.0.1 --directory portalIf you use INDUS-DEDUCT code, benchmark tables, or the 690-inscription damaged restoration catalog, please cite the permanent Zenodo archive:
@software{emre_2026_indus_deduct,
author = {Emre and {Emare INDUS-DEDUCT Research Collective}},
title = {{INDUS-DEDUCT: A Calibrated, Anti-Hallucination Computational Benchmark for the Indus Valley Script}},
month = nov,
year = 2026,
publisher = {Zenodo},
version = {v1.0.1},
doi = {10.5281/zenodo.24383348},
url = {https://doi.org/10.5281/zenodo.24383348}
}- Indus Corpus: Derived from the Interactive Corpus of Indus Texts (ICIT; Bryan K. Wells & Andreas Fuls) and Corpus of Indus Seals and Inscriptions (CISI; Asko Parpola et al.), stripped of all speculative columns in Layer 0.
- Calibration Corpora: Ur III Sumerian gold corpus (CDLI / MTAAC), Linear B transliterated words/tablets, and Linear A open corpus (
LinearAWords.json).