INDUS SCRIPT
I led an AI-assisted study that tested where Indus signs recur and whether published readings performed better than chance.
Unreviewed preprint- Role
- Independent researcher
- Duration
- Oct 2026–present
- Team
- Independent project with disclosed AI assistance
- Status
- Unreviewed preprint
First find out which patterns repeat
A sign can turn up in the same position often enough to warrant a closer look. The next question is whether that pattern survives in another collection. I led a study of positional structure in the Indus script and tested published readings against models of what we could expect by chance.
I directed an AI research assistant across nine sessions. We worked with Mahadevan, ICIT-derived and CISI material, checked digit-level OCR, built a SQLite object map and matched signs across catalogues. The public package includes a 1,132-entry sign list, a claim-scoring benchmark and a blind rater kit.
Give structure and meaning separate tests
The study reports 38 signs with at least 85% positional confidence, of which 28 replicated in an independent corpus. Seven of eight pre-registered held-out predictions replicated. The 42 published proposals and two complete keys tested performed no better than the per-sign null models. That result applies to those readings and the tests used here.
I kept pre-registration hashes, repeated seeds and false-discovery-rate controls in the record. The checks were rerun from a fresh clone and corrections were published in response to public feedback. The 54-page manuscript has not undergone formal peer review. Code is MIT-licensed and results use CC BY 4.0. The study makes no decipherment claim.
The record
What the evidence says.
- Preprint pages
- 54
- Updated public manuscript
- Positional signs replicated
- 28
- Independent-corpus study result
Indus Script Statistical Analysis / Workflow
How the study tests signs and proposed readings
- Match the signs
Align signs across the Mahadevan, ICIT-derived and CISI materials.
- Define the baselines
Use null models to estimate the patterns expected by chance.
- Test two questions
Measure positional structure and separately test how published readings perform against per-sign nulls.
- Check replication
Repeat structural tests in an independent corpus.
Worked explanationSeparate position from meaning
- Input
- A sign appears repeatedly in a particular position in one corpus.
- Decision / rule
- Compare the positional pattern with a null model, then test replication in an independent corpus. Proposed readings receive a separate test.
- Outcome
- A replicated positional pattern supports a structural result. It does not, by itself, establish what the sign means.
Scope of this exampleThis explains the unreviewed study’s method, not a new statistical run or a decipherment.
Working through a similar decision?
Talk to me about your project