Research objective
Build the most transparent, archaeologically grounded and reproducible Indus-sign corpus possible, then identify structural predictions that every interpretation must explain.
The laboratory does not begin by assuming Tamil, Proto-Dravidian, Indo-Aryan, Munda or any other language. Each proposal receives the same data, disclosure rules and held-out tests.
Why start with infrastructure?
Indus inscriptions are normally very short, sign inventories differ, the underlying language is unknown and no secure bilingual inscription has been found. A proposed reading can therefore look persuasive through flexible sign matching, selective examples or unrecorded editorial choices. The laboratory makes those choices visible and reversible.
Corpus register
One record per physical object, with images, excavation context, object type, chronology, reading direction and uncertainty.
Sign crosswalk
Mappings among Mahadevan, Parpola, CISI/ICIT and future sign lists without silently merging variants.
Hypothesis registry
Pre-declared sign values, language assumptions, expected patterns, failure conditions and revision history.
Test bench
Grouped holdouts, baselines, uncertainty intervals, ablations and blind predictions on unseen inscriptions.
Minimum corpus record
| Field | Purpose | Rule |
|---|---|---|
| object_id | Stable identity of the physical artefact | Never assign one ID to multiple impressions. |
| source_catalogue_id | Links to CISI, Mahadevan or excavation publication | Preserve every original identifier. |
| site / context | Site, mound, trench, locus and layer | Use “unknown” rather than infer provenance. |
| chronology | Archaeological phase and date range | Keep contextual and stylistic dates separate. |
| object_type / material | Seal, sealing, tablet, pottery, copper, ivory, etc. | Record controlled terms and original description. |
| image / rights | Reference image and reuse status | Do not republish an image without documented permission. |
| reading_direction | Observed or inferred sequence direction | Store confidence and justification. |
| sign_sequence | Transcription under a named sign list | Mark damage, gaps, ligatures and uncertain segmentation. |
| motif / associated marks | Animal, narrative motif, numerals and non-text marks | Do not silently treat imagery as text. |
| bibliography / notes | Audit trail | Cite page, figure and retrieval date. |
Release 0.2: open seed corpus
This release converts the openly licensed, work-in-progress digitisation by Michael Carlson into two audit-friendly tables. It contains 179 inscribed sides from Mohenjo-daro, 1,003 encoded sign tokens, 182 primary sign IDs in the seed, and a 396-row sign crosswalk. The source snapshot is fixed at commit ad2f1e218a34b8c33c57de0d6cb8d99272765bbb.
Use with caution
This is a secondary digital transcription and has not yet been checked object by object against every CISI plate. No seal images are redistributed. Crosswalk entries reproduce the source author’s mappings; they are candidates for review, not equivalence judgements by IndusValley.org. Verify a record against the published corpus before using it to argue for a reading.
Digital dataset: Michael Carlson, Corpus of Indus Seals and Inscriptions digitization, MIT licence. The underlying CISI publications and images retain their separate rights.
Release 0.3: CISI plate audit, Batch 1
The first object-level coverage audit checks M-1 through M-25 against the published Mohenjo-daro plates in Corpus of Indus Seals and Inscriptions I. All 25 catalogue identifiers were located on printed plate pages 1–15. The frozen open repository contains records for 23 of them; M-2 and M-11 are absent from that snapshot.
What this audit establishes
It establishes catalogue presence, plate location and digital-record coverage. It does not yet certify every P-code sign identification. Every M-1–M-25 row therefore keeps sign-by-sign validation marked pending independent verification. Missing JSON records are reported as gaps in this repository snapshot, not as missing artefacts.
Plate source: Corpus of Indus Seals and Inscriptions I: Collections in India, Mohenjo-daro plates 1–15. Digital comparison: Michael Carlson’s open digitisation, frozen at commit ad2f1e218a34b8c33c57de0d6cb8d99272765bbb.
Release 0.3: sign verification, Batch 2
This pass compares the ordered P-code sequence for every available repository record from M-1 through M-25 with the corresponding CISI impression and the repository’s sign-form descriptions. Results: 20 verified, 3 provisional because of damage or segmentation (M-19, M-22 and M-25), 1 unresolved (M-11), and 1 without a visible inscription in the published fragment (M-2).
Correction to Batch 1
Batch 1 described M-2 and M-11 as repository omissions. The closer plate review refines that statement. M-2’s surviving published fragment shows the animal and standard but no visible inscription, so no transcription should be invented. M-11 preserves only a damaged sign remnant; it remains unresolved until better evidence or an authoritative tracing supports an encoding.
Scope: verification here means that the ordered repository signs agree with the visible plate morphology at the level of the repository’s own P-code descriptions. It does not assign sound or meaning, and it is not yet a second-person independent review.
Analysis sequence
- Freeze a corpus version. Publish exclusions, uncertain readings and duplicate handling.
- Measure structure without meanings. Test directionality, sign position, repetitions, co-occurrence, sequence length and differences by object, site and period.
- Build comparison baselines. Include linguistic scripts, administrative codes, emblem systems and synthetic controls calibrated to similar text lengths.
- Reserve grouped test sets. Hold out entire sites, object families or excavation groups so near-duplicate inscriptions cannot leak into evaluation.
- Register language hypotheses. Declare historical reconstructions, rebus rules, sign values and permissible transformations before examining test results.
- Score predictions. Reward exact, pre-declared predictions and penalise unbounded homophony, sign-value changes and post-hoc exceptions.
- Replicate independently. A second researcher should reproduce the result from the frozen data and instructions.
What counts as progress?
More objects with verified provenance, clearer readings and documented image rights.
Patterns that remain stable across sign lists, sites, object types and resampling.
Pre-registered readings that succeed on genuinely unseen material.
Explicit results showing which attractive proposals do not survive testing.
Foundation hypothesis registry
| Family | Examples of claims to test | Required controls |
|---|---|---|
| Linguistic writing | Sequences encode names, titles, words, morphemes or syllables. | Known short-text writing systems and non-linguistic matched baselines. |
| Administrative notation | Signs encode commodities, quantities, institutions or transactions. | Object-type, context, numeral and motif associations. |
| Emblematic identifiers | Sequences identify persons, groups, offices or lineages. | Repetition across sealings, findspots and ownership contexts. |
| Mixed system | Some signs are linguistic while others are numerical or emblematic. | Component-specific predictions and ablation tests. |
| Proto-Dravidian | Proposed readings follow reconstructed forms and regular morphology. | Comparison across Dravidian branches and chronologically plausible reconstructions. |
| Other language families | Indo-Aryan, Munda or unknown-language proposals. | Identical scoring rules, historical chronology and held-out prediction. |
Evidence standards
- A similar-looking sign is not automatically the same sign.
- A plausible word is not a decipherment unless the mapping predicts other texts.
- Genetic ancestry cannot directly identify language.
- Modern vocabulary must not be projected four millennia backward without historical reconstruction.
- Statistical regularity can establish constraints, but cannot by itself supply sound or meaning.
- Negative results and changed hypotheses remain in the public revision history.
Starting sources
- Roja Muthiah Research Library, Indus Research Centre — Mahadevan-based corpus and concordance access.
- Iravatham Mahadevan, The Indus Script: Texts, Concordance and Tables, ASI Memoir 77 (1977).
- Corpus of Indus Inscriptions — catalogue-oriented documentation and discussion.
- Kenoyer: origin, context and function of the Indus script.
- Rao et al.: Markov modelling and Yadav et al.: n-gram analysis.
- IndusValley.org reproducibility review of the Kee2u project.
Current assessment
Established: the sign sequences exhibit non-random positional and co-occurrence constraints. Open: whether all signs form a linguistic writing system, what language or languages were involved, and how signs mapped to sound or meaning. Not established: any complete decipherment.