INDUS

A falsifiable computational workbench for an undeciphered script.

This repository asks what can be learned from sign order, repetition, archaeological metadata, and conservative external anchors without pretending that statistical labels are translations.

Bottom line: the corpus contains strong, generalizable sequential structure. This project maps that structure and produces testable hypotheses. It does not recover the Harappan language, pronunciation, or literal readings.

What is new here

The earlier experiment optimized coverage: it assigned one of five roles to every token. That was useful as an exploratory map, but easy to misread as “100% deciphered.” This release changes the center of gravity from coverage to falsifiability:

• exact duplicates are removed before model fitting;
• held-out strings never appear in training;
• four masked-sign models are compared against one another;
• normalized reading order is tested against exact reversal;
• every role label is observational rather than lexical;
• every headline has a denominator and an explicit forbidden interpretation;
• all 5,679 source rows, 3,219 analytic lines, and 12,910 token-level role assignments remain inspectable.

Results at a glance

Question	Result	What it establishes
Can local context recover an unseen sign?	26.1% top-1, 50.9% top-5	Reusable combinatorial constraints
Frequency-only baseline	11.0% top-1, 24.2% top-5	The context gain is substantial
Does a train-only model prefer normalized order?	79.0% of unseen eligible strings	Directional asymmetry generalizes
Do numeral-like signs precede the conservative measure-sign set?	166 observed vs 67.4 shuffled; empirical p < 0.0005	A strong quantity→measure association
Does sign identity carry positional information?	MI 0.368 nats vs null 0.071; empirical p < 0.001	Signs have differentiated positional behavior
Can every token receive a structural slot hypothesis?	100% annotation coverage	A complete machine-readable role map, not complete decipherment

The masked-sign scores are five-fold means on 2,082 unique complete strings. Test tokens unseen in a fold's training vocabulary are reported as OOV and excluded from accuracy. Read CLAIMS.md before quoting a number.

Corpus anatomy

• 5,679 source records and 3,219 complete, direction-known analytic line records
• 2,082 unique sign strings after exact deduplication
• 606 observed sign IDs and 9,311 tokens in the deduplicated corpus
• median string length 4; maximum 13
• 215 sign types occur once; 395 occur fewer than five times
• the 67 most frequent signs account for 76.1% of deduplicated tokens
• 35.3% of analytic line records duplicate another exact string
• 284 formulas repeat across artifacts and 77 across sites

These numbers are not incidental. They explain both why pattern induction works and why literal decipherment remains underdetermined.

What the role map means

ACGI-5 assigns five deliberately modest structural labels:

OPENING_SLOT → QUANTITY_LIKE → MEASURE_ASSOCIATED → CORE_SLOT → CLOSING_SLOT

The order is soft, not mandatory. The names describe distributional behavior. OPENING_SLOT does not prove “issuer”; CORE_SLOT does not prove “commodity”; CLOSING_SLOT does not prove “approval.” Those remain archaeological hypotheses.

flowchart TD
    A["Pinned secondary corpus"] --> B["Quality audit + direction normalization"]
    B --> C["Deduplicated structural tests"]
    C --> D["Held-out prediction benchmarks"]
    C --> E["Uncertainty-aware slot map"]
    D --> F["Claims with baselines"]
    E --> F

Loading

Reproduce it

Python 3.11+ is recommended.

git clone https://github.com/sriharideveloper/indus.git
cd indus
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python analysis/run_all.py
python analysis/verify_results.py
pytest -q

The full role model performs 24 bootstrap refits and can take several minutes. For the faster independent benchmarks:

python analysis/evaluate.py

To rebuild the reviewer notebook, install requirements-dev.txt and run python analysis/build_notebook.py.

Query a sign or exact sequence after generating results:

python analysis/query.py 740
python analysis/query.py 032-220

Repository map

Path	Purpose
analysis/	corpus parsing, benchmarks, slot induction, contextual inference
data/DATA_CARD.md	provenance, inclusion rules, leakage and licensing
results/evaluation_summary.json	compact machine-readable benchmark summary
results/tables/	fold results, profiles, lexicon and complete role maps
reports/RESEARCH_NOTE.md	results, interpretation, failure analysis and next experiments
docs/CLAIMS.md	claim ledger: allowed and forbidden interpretations
docs/METHODS.md	algorithms, splits, controls and known circularities
docs/RESEARCH_AGENDA.md	tests that could move this toward actual decipherment

Scientific stance

The strongest conclusion is not “we solved the Indus script.” It is:

Sign order is predictable on held-out inscriptions, local context contains enough information to recover many hidden signs, and numeral-like signs have a robust directed association with a conservative measure-sign set. A five-slot model is a useful index over those regularities, but its English labels are hypotheses rather than translations.

This distinction is the project’s most important feature.

Citation

See CITATION.cff. The corpus snapshot is redistributed under its upstream MIT license with provenance and checksum; cite the upstream corpus and the primary studies listed in SOURCES.md.

License

Code and original documentation are MIT licensed. The pinned secondary corpus retains its upstream MIT notice in data/raw/UPSTREAM_LICENSE.