Skip to content

About

Reproducible statistics on the undeciphered Indus script: structure, same-size comparison with languages and a code, sign groups. Not a decipherment.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

11 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Indus sign statistics

A small, reproducible look at the structure of the Indus script (c. 2600–1900 BCE): how its signs are distributed, where they sit in a text, how it compares with real languages and with a non-linguistic code cut to the same size, and which signs keep the same company.

This is not a decipherment. No sign is given a sound or a meaning here. Every number below is computed from the data, and every reading of any sign stays open.

Commissioned and directed by Capt. Anil Sharma; implementation with an AI assistant (Claude, Anthropic).

Context-neighbour pairs

Findings

All figures come from results/ and are rebuilt by ./run.sh.

  1. Corpus. 2,543 inscribed objects, which reduce to 1,902 distinct sign sequences once repeated texts (multiple impressions, tablet series) are merged. 591 sign types; mean length 4.65 signs, longest 17. 36% of sign types occur only once, and the 10 commonest signs make up a third of all occurrences.

  2. Signs keep places. One sign (code 740, the "jar") makes up about 11% of all occurrences and stands at the same end of a text in 70% of its uses. It almost never stands at the other end. Several other signs are just as tied to one end or to the middle.

  3. Between a language and a code. Against same-size samples of Vedic Sanskrit (syllables), Old Tamil (aksharas), English (words) and chess openings (moves):

    • Indus signs are more tied to their position than any of the three languages, about as much as chess moves (position information 0.36, against 0.10–0.22 for the languages and 0.33 for chess).
    • Once position is held fixed, the order of signs still carries structure at the level of Old Tamil (22% below a slot-preserving shuffle, against 24% for Tamil, 18% for Sanskrit, 12% for English and 27% for chess).
    • Signs rarely repeat within a text (4.4% of texts), close to English words (4.6%) and below the syllable scripts (8–13%). This fits signs that stand for words more than for single sounds.
  4. One very predictable end. At one end of a text five signs fill 59% of the slot, led by the jar. Old Tamil phrases show a similarly predictable end (their recurring suffixes). The Indus pattern fits a closing marker such as a suffix, but a title or a classifier would produce the same shape, and so does the opening of a chess game.

  5. Signs that look alike behave alike. Grouping signs by the neighbours they keep, 27 of 94 common signs have as their closest match a sign with a catalogue code within 10 of their own, against 2.2 expected by chance (p ≈ 0.002, 500 shuffles). The catalogue numbers group signs by shape, and the rendered pairs show the same shape families: stroke counts 3/4/5 (number-like), the fish with and without added marks, cone, branch, jar and figure variants. Modifiers change a sign without moving it out of its slot.

  6. The text barely depends on the picture. Pairing each distinct text with the animal carved on its object (the dataset's own iconography labels; six motifs with at least 25 texts), the sign-motif mutual information is 0.188 bits against 0.175 ± 0.007 for shuffled motifs, only 1.9 sd above chance. The label "Bull1" alone carries 913 of the 1,134 texts (most likely the one-horned bull, the commonest seal animal; the dump does not define its labels). Picture and text carry largely separate information, as expected if the animal is an emblem and the text a name or title.

  7. Later Hindu sacred symbols do not resemble the signs more than other traditions' symbols do. A pre-registered blind test (blind/PREREGISTRATION.md) drew 10 Hindu sacred symbols and 10 each from Christian, Chinese and Egyptian tradition in one line style, and let a computer find each one's closest Indus sign by skeleton shape. Mean distance: Hindu 2.30 px, Christian 2.27, Chinese 2.28, Egyptian 1.98; one-sided permutation p = 0.73. Under the rule fixed in advance this is no excess resemblance found. Random signs from real pictographic scripts match about as well. Simple shapes (a circle with a cross, three lines) find look-alikes in every tradition (blind/matches.png). This tests the script's signs only: the seal pictures, where the scholarly case for continuity rests, are not in the data.

  8. The pictures follow fixed rules, and some seals carry no writing at all. 202 seals were coded by eye from the photographic plates of the Corpus of Indus Seals and Inscriptions, vol. 1 (Joshi & Parpola 1987), 75 from Mohenjo-daro, 87 from Harappa and 40 from Lothal (results/seal-coding-202.csv). The cult stand appears only before the one-horned bull (50 of 50 where readable); the trough only before bison, buffalo and rhinoceros (14 of 14). Swastika seals (25 of 25) and geometric seals (14 of 14) never carry text; animal seals and picture-less seals almost always do. Several Lothal seals have text along the top and an empty field below, as if the text were cut first. The pages were chosen for variety, so this is not a random sample of all seals, and the coding is one person's reading of photographs.

  9. Seal texts are individual, and the animal does not follow the text. Of 1,469 seals with text, 88% carry a text found on no other object, while 65% of tablets share their text with another tablet, as moulded series would. Of 40 texts found on two or more seals, 82% keep the same animal on every copy, but shuffling the animal labels gives 84% ± 2% (the one-horned bull is 85% of seals), so a text does not stay with its animal more than chance; 7 texts appear with different animals, for example on a one-horned-bull seal and an elephant seal. No part of the text (opening sign, closing sign, middle) is clearly tied to the animal. This fits a text that identifies an individual and an emblem that does not belong to that individual, such as a sign of office or kind of business, better than a personal clan sign. It is a reading, not a proof: the one-horned bull's dominance limits the test, and the animal labels are the dataset's own (checked against the corpus plate headings for the joinable cases). results/picture-test.txt

The same grouping method was first run on Old Tamil, where the answer is known. It recovers letters that share a consonant strongly (11.8 sd above chance) and letters that share a vowel weakly (2.4 sd). It works, but only partly.

Toward the language (steps 1–7)

What can be said about the language behind the signs, short of reading them?

  1. Numbers come before what they count, weakly. The short-stroke number signs (codes 3–7, 13–19, 27, 55) end a text in 3.0% of their uses, against 15.9% ± 9.3% for other signs of the same frequency (only 1.2% of 500 control sets are as low). A second measure, which signs take several different numbers on each side, is even (21 against 21). Numbers before nouns fits Dravidian, Indo-Aryan and Munda alike, so this argues only against a Sumerian-type order, where numbers follow. The result depends on the reading direction (right to left, taken from the jar sign's text-final place). results/numerals.txt

  2. Ending patterns do not separate the candidate families. Inside words, Vedic Sanskrit (written syllables), Old Tamil (aksharas) and Santali (a Munda language, from the Santali Wikipedia) all make the last unit the most predictable (0.68–0.76 of their own unit entropy) and none makes the second-last predictable. In short phrases, the Indus end is more fixed than any of them (0.68 against 0.78–0.88) and its start less (1.02 against 0.81–0.89): a formula with a fixed close. results/endings.md

  3. The outside record of the language is two names. Sumerian texts describe Meluhha (the usual identification of the Indus region) only in Sumerian words: "black land", highland mes timber, carnelian, francolins, a bird called ha-ia₃, gold, tin-bronze, boats. A seal names "a translator of the Meluhhan language". The two personal names Nanaza and Samar are, by the specialists' own account (read here only secondhand), the only possible words of the language. Two names cannot choose a family. results/meluhha-evidence.md

  4. Numbers are used two ways: counting, and fixed numbered names. Taking the signs that follow a number far more often than chance (order shuffled inside each text, 500 times), one group takes many different values, as counted things would: 390 (a forked branch) after 3 to 8, z +9.7; 220 after 3 to 9; 405 and 407 (branches with twigs) after 3 to 8; 900 (an arc) after 3 to 5; 156, mostly on tablets. Another group always takes the same value, like a fixed name: 585 follows 7 in 15 of its 25 uses, 575 always follows 7, 423 always follows 3. The jar sign is preceded by a number less often than chance. Most of these pairs are on seals rather than tablets, which fits counts and numbered names inside identifiers better than bookkeeping. Shapes are described here, not read. results/number-anchor.md, results/number-anchor-signs.png

  5. A slot before the closing jar holds one unit, sometimes a numbered name. Only three fixed number compounds exist: 7 + 585 (15 uses), 7 + 575 (12) and 3 + 423 (5); the 7 appears both as seven strokes in a row and as 4 + 3 in two rows. In about half their uses exactly one sign follows them, and the sign after them is the jar 72% of the time. Counted pairs (a number + 390, 220 and so on) behave differently: they end the text themselves in 62% of uses and are followed by the jar only 15% of the time. Where 7 + 585 stands in the frame "33 _ jar", single signs (923, 760, 176, 904) also stand. This suggests a frame of the form [...] [one unit] [jar], where the unit is a single sign or a numbered name acting as one word, with counting in a different kind of text. The numbers are small (32 uses; each frame seen 2 to 5 times) and the reading direction is inferred, so this is a pattern, not a proven grammar. results/numbered-names.txt

  6. The slot before the jar is an open class with a bound core. 35% of distinct texts end with the jar. The unit before it is filled 665 times by 161 different units, half of them seen there only once, as a slot for names or identifiers would be. A core of 36 signs stands there far more often than chance: sign 760 in 61 of its 77 uses (z +29.8), and 752, 772 and 773 almost always; several of the strongest are U-shaped, vessel-like signs. In texts that do not end with the jar, these signs end the text less often than other signs (17% against 27%), so they need the jar after them, like a stem with its fixed ending. The counted signs 390 and 405 are the exception: in jar-less texts they end the text 62% and 79% of the time. Two endings, then: an open-class unit plus the jar, or a number plus a counted sign without it. The jar behaves as a closing element bound to what precedes it, which fits a case ending or title marker, and equally a closing formula in a code. results/prejar-slot.md, results/prejar-slot-signs.png

  7. Indus-script seals made abroad follow different rules. Round "Gulf Type" seals with Indus signs were found in Bahrain, Failaka, Mesopotamia and Iran. Their inscriptions are drawn in Laursen (2010, Fig. 11); we transcribed all 29 into sign features, right to left, each sign with a confidence mark (abroad/laursen-fig11-transcription.csv). Three are left out: one is in Linear Elamite, one carries no text, one is pictorial. Against this repository's dataset:

    inscriptions open with the "twins" sign end with a jar-like sign
    home texts (this dataset) 1,902 0.4% 35%
    Gulf Type seals found abroad 21 24% (p ≈ 5 × 10⁻⁹) 14% (p = 0.035)
    Gulf Type seals found in the Indus Valley 5 0 0

    The "twins" sign (two human figures side by side) opening these inscriptions was first noted by Parpola (1994) and set out by Laursen (2010), who calls it "the only rule-like regularity so far attested in this 'western written language'" and concludes that the language of these seals was, "with only a few exceptions ... of non-Harappan origin". Our transcription confirms the opening and adds that these texts seldom close with the jar, the sign that ends a third of home texts: a change in how the text is framed, not only in the names it carries. Whether that means another language, names from another tradition, or a later way of writing (most Gulf Type seals date from about 2050 BC on) is not decided by this. The transcription is one reader's reading of drawings, and the jar-like count includes doubtful U shapes, so the shortfall is if anything understated. results/laursen-abroad.txt

Taken together: order and position statistics can rule some kinds of language out, but they cannot choose between Dravidian, Indo-Aryan and Munda, which look alike on every measure tried. Choosing needs evidence about sounds, which the record does not yet hold.

Relation to prior work

Most of the structural findings above re-derive results that were already published or found in parallel, and should be read as independent replications rather than discoveries:

here prior or parallel work
findings 1–4 (structure, signs keeping places, the jar ending, comparison with languages) Mahadevan 1977; Rao et al. 2009; Yadav et al. 2010; Koshy 2026 (§8–9, with Linear B, Sumerian and Egyptian, size-matched)
finding 5 (look-alike signs behave alike) Koshy 2026 (§4: a weak but real agreement between shape and behaviour)
findings 6, 9 (text independent of the animal) Koshy 2026 (§4), and Marshall (1931), who already reported the same inscription over different animals
finding 9 (seal texts individual) Kriger & Hunt 2026; Koshy 2026 (95% of Mohenjo-daro seal texts unique)
steps 1 and 4 (number signs) Mahadevan; Kriger & Hunt 2026; Koshy 2026 (§6: short strokes go mainly with "forked stem" signs, our 390/405/407)
step 2 (suffixing cannot be told from a phrase-final formula) Koshy 2026 (§14) reaches the same conclusion
steps 5–6 (the slot before the jar) Koshy 2026 (§5, §8: the pre-final sign and the ending behave as one unit; the slot is open)
step 7 (the "twins" opening abroad) Parpola 1994; Laursen 2010; Koshy 2026 (finding 5)

What, as far as we could find, is not already in these sources: the pre-registered blind test of Hindu sacred symbols (finding 7), the shortfall of jar endings on Gulf Type seals measured against home texts (step 7), the particular fixed number compounds 7 + 585, 7 + 575 and 3 + 423 as fillers of the slot before the jar (step 5), and the inclusion of a Munda language (Santali) in the endings test (step 2). The reading direction used throughout (right to left) follows the literature and Mackay's own statement; none of our statistics can confirm a direction, since they would read the same way reversed.

Koshy, N. S. (2026). Indus script: machine-readable seal catalogues from Mackay (1938) and Marshall (1931), with photo-coded text features and a reproducible analysis pipeline. Zenodo, 10.5281/zenodo.22852769. Kriger, C. & Hunt (2026), "Positional constraints, sequence uniqueness, and stroke numerals in Indus seal inscriptions from Mohenjo-Daro" (as cited by Koshy; not read here).

Limits

  • Direction. The sign order is taken as stored. That the stored start is the reading end (the script is mostly read right to left) is an inference from the jar sign's known text-final place, not a field in the data.
  • Sanskrit. The syllable splitter attaches final consonants to the last syllable, which scatters Sanskrit's case endings. The Sanskrit grammar-shape row is not a fair test until it splits by written akshara.
  • Genre. The language samples are verse and prose cut to the Indus lengths, not seal-like texts (names, titles, formulae). A genre-matched comparison, for example Mesopotamian seal inscriptions, has not been done.
  • Entropy. Plug-in estimates on about 9,000 tokens are biased low. Only comparisons between samples of the same size, and against shuffles of the same text, are meaningful.
  • Corpus choices. Identical sequences are merged; the line-break marker is removed; 45 rare signs have no glyph in the font and show as "?". The sign catalogue (the corpus's own numbering) decides what counts as one sign, and other catalogues split and merge signs differently.

Reproduce

./fetch_data.sh   # downloads all inputs from upstream, pinned to commits, sha256-checked
./run.sh          # rebuilds everything in results/

Needs Python 3 with numpy and Pillow. Every script runs offline once the data is fetched.

script output
src/structural.py results/structural.txt: counts, positions, pairs, entropy against shuffles
src/compare.py results/compare.md: the same-size comparison table
src/grammar_groups.py results/grammar-and-groups.md: ends test and context groups
src/shape_family.py results/shape-family.txt: code-distance test
src/motif.py results/motif.txt: text against the carved animal
src/numerals.py, src/numeral_position.py results/numerals.txt: number signs and word order
src/step2_endings.py results/endings.md: ending patterns against Sanskrit, Tamil, Santali
src/number_anchor.py results/number-anchor.md, results/number-anchor-signs.png: signs that follow numbers
src/numbered_names.py results/numbered-names.txt: fixed number compounds and the slot before the jar
src/laursen_abroad.py results/laursen-abroad.txt: Gulf Type seal inscriptions (Laursen 2010) against home texts
src/prejar_slot.py results/prejar-slot.md, results/prejar-slot-signs.png: every filler of the slot before the jar
src/picture_test.py results/picture-test.txt: repeated texts, text slots and object types against the carved animal
(by hand) results/seal-coding-202.csv: picture coding of 202 seals; results/meluhha-evidence.md: sources on Meluhha, quote or null
src/render.py results/*.png: sign sheets
blind/symbols.py, blind/match.py blind/drawings.png, blind/result.txt, blind/matches.png: the blind symbols test

Data and credit

No data is redistributed here; fetch_data.sh downloads each file from its source.

input source licence
Indus sign sequences and sign font yajnadevam/indus-website @ 2434249 GPL-3.0
Rig Veda lines (IAST) Equationist/indusscript-decipherment-stats @ 90c70d8 Unlicense
Tirukkural (original lines only) tk120404/thirukkural @ cf3ff6c Apache-2.0
Chess openings lichess-org/chess-openings @ a6189a3 CC0-1.0
Pride and Prejudice Project Gutenberg #1342 public domain (US)
Santali Wikipedia, dump of 2026-10-01 dumps.wikimedia.org/satwiki CC BY-SA 4.0

Seal pictures were read from the plates of Corpus of Indus Seals and Inscriptions, vol. 1 (Joshi & Parpola 1987, Memoirs of the Archaeological Survey of India 86), as hosted by the Indira Gandhi National Centre for the Arts. No image or text of that volume is reproduced here, only our coding. Sumerian passages are quoted from the Electronic Text Corpus of Sumerian Literature (Oxford).

The Indus sequences follow the Interactive Corpus of Indus Texts (ICIT) of Bryan K. Wells and Andreas Fuls, which rests on the corpora of Iravatham Mahadevan (1977) and Asko Parpola and colleagues (CISI). This study uses only the sign sequences, sites and object types. The dataset's maintainer has proposed a reading of the script; this study neither uses nor evaluates it.

Related work: Rao et al., Science 324 (2009); Farmer, Sproat and Witzel (2004); Parpola, in Qala'at al-Bahrain vol. 1 (1994), pp. 304–315; Laursen, Arabian Archaeology and Epigraphy 21 (2010): 96–134; Nair, arXiv:2604.17828 (2026); Tiwari, NLP4DH 2026.

Licence

Code: GNU Affero General Public License v3.0 (LICENSE). Copyright © 2026 Capt. Anil Sharma. Rendered sign sheets in results/ are drawn with the GPL-3.0 font above. The transcription abroad/laursen-fig11-transcription.csv is released under CC BY 4.0 (credit: Capt. Anil Sharma, indus-sign-statistics). It is our reading of the drawings in Laursen (2010, Fig. 11); the drawings themselves are Laursen's and are not reproduced here.

About

Reproducible statistics on the undeciphered Indus script: structure, same-size comparison with languages and a code, sign groups. Not a decipherment.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages