The Indus Script
How AI Helped to Understand a 4,000-Year-Old Information System
Four thousand years ago, someone in Mohenjo-Daro carved a tiny stone seal. On it: a unicorn with muscles so finely rendered you can count the tendons, and above it, a string of mysterious signs that no one has been able to read since the city fell silent. That seal survived earthquakes, floods, and the full weight of four millennia. It looks like it was made yesterday morning.
For a century, scholars tried to crack the Indus script the way you crack a code — by finding the language hidden inside. They all failed. Every last one. Not because they weren’t brilliant, but because they were asking the wrong question. The Indus seals were never letters in a language. They were something far more interesting: the world’s first information system. A structured code that identified people, goods, and transactions across an entire civilization — without a single word of grammar.
This is the story of how that discovery was made. It involves seven groups of researchers spread across a hundred years and five continents. A Finnish professor who spent half a century chasing fish-shaped rebuses. A software engineer from Kolkata who saw tax receipts where everyone else saw prayers. Three scholars who declared the whole thing wasn’t writing at all — and got death threats for their trouble. A Tamil Nadu government that offered a million dollars to anyone who could solve the riddle. An information theorist who studies cat behavior and compression algorithms. A systems theorist in Toronto. And an artificial intelligence that, in one three-hour conversation, ran the statistics that changed everything.
The result: 98.3% of all seal inscriptions are unique. The signs follow strict positional rules. Stroke marks function as numerals, clustered in the penultimate position. The system had the capacity to generate over a million distinct identifiers. This was not a civilization that wrote poems. This was a civilization that invented administration — four thousand years before the barcode.
We may finally have the answer. And it’s not what anyone expected.
Keywords: Indus Valley Civilization, Indus script, archaeological decipherment, information systems, structural analysis, Bronze Age, Mohenjo-Daro
Contents
Chapter One: Eighty Handshakes. 15
Chapter Two: The Riddle Nobody Could Solve. 22
Chapter Three: The Concordance. 28
Chapter Four: What Mahadevan Built 35
Chapter Five: What Survives Four Thousand Years. 39
Chapter Six: The Fish and the Star 44
Chapter Seven: The Entropy of Signs. 49
Chapter Eight: The Language That Never Was. 54
Chapter Nine: The Bombshell 59
Chapter Ten: The Format Is the Finding. 65
Chapter Eleven: Search for M-1. 70
Chapter Twelve: The Entropy Chasers. 75
Chapter Thirteen: Counting Strokes. 81
Chapter Fourteen: What Would Remain of Us. 86
Chapter Fifteen: The Tax Collector 91
Chapter Sixteen: Near-Duplicates and Prefix Trees. 97
Chapter Seventeen: The Silence of Things. 100
Chapter Eighteen: The Graffiti Connection. 105
Chapter Nineteen: Administrative Continuity. 110
Chapter Twenty: The Scent of the Seal 115
Chapter Twenty-One: The Compression. 119
Chapter Twenty-Two: The Machine. 124
Chapter Twenty-Three: The Paper 130
Chapter Twenty-Four: The First Information Civilisation. 135
Chapter Twenty-Five: What We Do Not Know.. 140
Chapter Twenty-Six: The Prize. 145
Chapter Twenty-Seven: Eighty Handshakes Back. 151
Preface
Here is a fact that should bother you more than it does: four thousand years is not a long time.
I know it sounds long. It sounds like the kind of number that belongs in a museum, next to a dusty pot and a small card explaining that this pot was used for “ritual purposes,” which is what archaeologists write when they have no idea what something was used for. Four thousand years sounds ancient. It sounds like the deep past, the kind of past that has nothing to do with you, your mortgage, or your phone battery.
But consider this. If you could somehow line up every person who ever lived between now and the time of the Indus Valley Civilization, and if each of them shook the hand of the next, you would need about eighty people. Eighty. That’s a mid-sized wedding. That’s a Tuesday morning lecture hall at a second-tier university. Between you and the person who carved a tiny stone seal in Mohenjo-Daro, there are fewer handshakes than there are people in the average grocery store at five o’clock on a Friday.
That person — the one who held the chisel — was not a caveman. He was not a grunt. He was someone who woke up in a city with a drainage system more advanced than most of what Europe would manage for the next three thousand years, who walked past public baths and standardized brick houses, who lived in a society that traded with Mesopotamia and managed its economy with a system of tiny carved stones that we, in the twenty-first century, still cannot fully read.
This book is about those tiny carved stones. They are called seals, and they are magnificent.
Each one is about the size of a postage stamp. On one side, an animal — usually a unicorn, sometimes a bull, an elephant, a rhinoceros, a tiger. The carving is so precise, so anatomically detailed, that when you see a high-resolution photograph for the first time, your brain refuses to accept the date. It looks modern. It looks like something a jeweler in Milan might have made last week, not something dug from the rubble of a city that fell silent around 1900 BCE.
Above the animal, a line of signs. Between three and twenty of them, usually around five. Small, crisp, distinct. Some look like fish. Some look like jars. Some are just lines — vertical strokes, one or two or three. Nobody knows what they say. Nobody has known for a hundred years. Not for lack of trying.
The Indus script is the most famous undeciphered writing system in the world, if it is a writing system at all. That “if” has started more academic knife fights than any other two letters in the history of archaeology. People have devoted entire careers to it. Friendships have ended over it. In 2004, three scholars published a paper arguing that the signs weren’t writing at all, and they received death threats. Actual death threats. Over squiggles on rocks. Humanity is a remarkable species.
For a hundred years, the operating assumption was straightforward: the signs represent a language. Find the language, crack the code, read the seals. This was the approach that had worked for Egyptian hieroglyphs (with the help of the Rosetta Stone) and for Linear B (with the help of a brilliant architect named Michael Ventris and a lot of wartime codebreaking experience). It seemed reasonable. It seemed like the only approach.
It wasn’t.
Every attempt to find the language behind the Indus signs has failed. Dravidian, Sanskrit, Munda, Proto-Elamite — all proposed, none confirmed, all fiercely defended, and all quietly set aside when the evidence refused to cooperate. A hundred years. Hundreds of scholars. Thousands of papers. Zero decipherment.
And then something happened that was not supposed to be possible.
In March 2026, in a city called Toronto — which is about as far from Mohenjo-Daro as you can get without leaving the planet — a systems theorist sat down with an artificial intelligence and asked it to download the only fully open-access corpus of Indus seal inscriptions from a GitHub repository. The AI did so. The theorist asked it to count the signs, calculate their entropy at each position, test whether the sequences were unique, and check whether the stroke marks behaved like numbers. The AI did all of this. It took about three hours.
The results were not ambiguous.
Ninety-eight point three percent of all seal inscriptions in the corpus were unique. Not approximately unique. Not “pretty different.” Statistically, significantly, overwhelmingly unique, to a degree that could not be explained by chance. The signs at position one were drawn from a narrow set. The signs in the middle were wildly diverse. The signs at the end were moderately constrained. Stroke marks — simple vertical lines, one through five — clustered in the penultimate position, and the number two accounted for sixty-seven percent of all numerals. The system had the combinatorial capacity to generate over a million distinct sequences.
This was not a language. This was not a prayer. This was not a name.
This was a code. A structured, positionally constrained, statistically regular identification system. Something like a license plate. Something like a passport number. Something like a barcode — four thousand years before the barcode was invented.
Which means the Indus Valley Civilization was not, as has been assumed, a literate society that happened to write very short things. It was an information society — possibly the first in human history — that managed complex economic and administrative functions through structured identifiers rather than written language. They did not need an alphabet. They had something better: a system.
Now, I should be honest with you, because this book is going to ask you to think carefully, and you deserve to know what you’re getting into.
I am not an archaeologist. I am not a linguist. I am not an Indologist. I am a systems theorist with a background in philosophy and an unhealthy interest in patterns. My co-author on the research paper at the heart of this book, Treasure A. Hunt, is an information theorist who has developed a framework for analyzing constraint-governed systems — a framework that turned out to be extraordinarily useful when pointed at four-thousand-year-old seals. The statistical analysis was performed by an artificial intelligence, Claude Opus 4.6, built by Anthropic, which did not complain about the hours and did not need coffee.
None of us set out to solve the Indus script. We set out to see whether a structural approach — one that asks “how is this system organized?” rather than “what language does it encode?” — could extract meaningful information from the seals without reading a single sign. It could. It did. And the answer it produced was both surprising and, in retrospect, obvious: the format is the finding.
But we did not do this alone. No one does anything alone, despite what the Nobel committee would like you to believe. This book tells the story of seven groups of researchers whose work, over the course of a century, made our analysis possible. Each of them found a piece of the puzzle. Each of them was, in some specific and important way, right. And each of them was, in some equally specific and important way, incomplete.
Iravatham Mahadevan, a Tamil civil servant who gave up a prestigious career to spend decades building the only open concordance of Indus signs. Without his work, there would be no corpus to analyze. He died in 2018, seven years before this book was written, and every number in our research traces back to the data he compiled by hand.
Asko Parpola, a Finnish professor who spent fifty years pursuing a Dravidian reading of the signs through an ingenious rebus principle. His reading may yet prove correct — for the semantic layer beneath the administrative one.
Steve Farmer, Richard Sproat, and Michael Witzel, who in 2004 had the courage to argue that the Indus signs are not writing. They were vilified. They were partially right. The signs are not writing in the linguistic sense. But they are not nothing, either.
Rajesh Rao and Nisha Yadav, computer scientists who brought entropy analysis and n-gram statistics to the field, proving that the signs carry structured information even if no one knows what kind.
Bahata Ansumali Mukhopadhyay, a software engineer from Kolkata who published seven papers arguing that the seals record taxes, trade licenses, and commodity controls. She was the first person to see what the system actually does. Our statistics confirm her.
K. Rajan and R. Sivanantham, whose government-commissioned study found that sixty percent of graffiti marks on ancient pottery in Tamil Nadu share morphological parallels with Indus signs — raising the astonishing possibility that the administrative tradition survived for over a thousand years after the cities fell.
And Treasure A. Hunt, whose information-theoretic framework gave us the analytical lens to see what was there all along.
Each of these people gets their own chapters in this book. Not because I want to be generous — though I do — but because the story of the Indus script is the story of how science actually works. Not as a single eureka moment, but as a slow, fractious, occasionally vicious accumulation of partial truths, each one built on the ruins of the last, until one day someone puts them together and the picture finally resolves.
There is also a chapter about what would happen if our civilization disappeared tomorrow. What would survive four thousand years? Not your phone. Not your house. Not your city. The concrete would crumble in five centuries. The steel would rust in a millennium. The data would die in decades. What would remain are things made of stone, things made of ceramic, and — if we were very lucky — things the size of a postage stamp, carved by someone who understood that beauty is not decoration but durability.
The Indus people understood this. Their seals survived because they were made to survive. And the information encoded in them survived because it was encoded in structure, not in meaning. You do not need to know what the signs say to know how the system works. The format is the finding. That sentence will appear several times in this book, because it is the most important sentence in it.
One more thing. In January 2025, the Government of Tamil Nadu announced a prize of one million US dollars for the decipherment of the Indus script. The prize is named after Iravatham Mahadevan. As of this writing, there is no official portal to submit a claim, which suggests that the government did not entirely expect anyone to do so. We intend to prove them wrong. But this book is not about the prize. This book is about the fact that eighty handshakes separate you from the person who carved that seal, and that person was not primitive, was not simple, and was not silent. They spoke, clearly and precisely, in a language made of structure. And after four thousand years, we are finally beginning to hear them.
At the back of this book, you will find the full text of the research article that started everything, along with the complete Python code that produced every number cited in these pages. The article is technical. The code is not pretty. But both are there for a reason: so that anyone with a laptop and an internet connection can verify every claim in this book within thirty seconds. In an age when trust in institutions is collapsing faster than Bronze Age cities, reproducibility is the only currency that still holds value. The data is open. The corpus is on GitHub. The code uses nothing but Python’s standard library. If you think we’re wrong, you can prove it before your coffee gets cold.
A note on how this book is organized. Each of the seven research groups gets three chapters: one about who they are and what they found, one about how their discovery fits into our analysis, and one that uses their insight to illuminate some aspect of how the Indus civilization might have actually worked. These third chapters are not speculation. They are inference — the kind of careful, evidence-based reconstruction that asks not “what did they believe?” but “what would a system like this require?” If the seals are identification codes, then someone must have issued them. If the strokes are numbers, then someone must have counted something. If the format is consistent across hundreds of seals, then someone must have enforced the format. These are not leaps of imagination. They are the necessary consequences of the data.
Boris Kriger
Toronto, 2026
Chapter One: Eighty Handshakes
There is a party trick that historians never perform at parties, mostly because historians are rarely invited to parties, and when they are, they tend to talk about crop yields in Mesopotamia until the room clears. But the trick goes like this: pick any moment in recorded history, and count how many human lifetimes it takes to reach it. Not generations — lifetimes. Fifty years each, end to end, no overlap.
To reach the time of Jesus, you need forty people. Forty. You could fit them in a school bus. To reach the building of the Great Pyramid at Giza, you need about ninety. That’s a midsized wedding reception, if the bride’s family is large and the groom’s side doesn’t travel well. To reach the founding of the Indus Valley Civilization — Mohenjo-Daro, Harappa, Dholavira, the great cities of the Bronze Age — you need roughly eighty people standing in a line, each one passing the memory of everything they knew to the next.
Eighty handshakes. That is the distance between you and the person who carved seal M-1 in Mohenjo-Daro. Between your morning coffee and a civilization that built cities with grid-pattern streets, indoor plumbing, standardized weights, and a script that nobody can read. Eighty people. Not a chasm. A corridor.
We tend to think of the deep past as impossibly remote, separated from us by an unbridgeable gulf of time and strangeness. This is a failure of imagination, not of mathematics. The Indus Valley Civilization flourished from roughly 2600 to 1900 BCE. It covered an area larger than ancient Egypt and Mesopotamia combined. At its peak, it may have supported five million people. Its cities were among the most carefully planned in the ancient world — not thrown together in a rush of imperial ambition, but laid out with an almost obsessive attention to proportion, drainage, and public space.
If you imagine the entire three-hundred-thousand-year history of Homo sapiens compressed into a single twenty-four-hour day, everything we call “recorded history” — every pharaoh, every empire, every revolution, every symphony — fits into the last twelve minutes. The Indus Valley Civilization falls somewhere around 11:50 PM. We are, all of us, living in the last few seconds before midnight. The past that we call ancient is, in the timescale of our species, absurdly recent. We just don’t feel it, because human memory is short and human arrogance is long.
And yet, around 1900 BCE, this vast and apparently prosperous civilization began to fade. Not with a bang — there is no evidence of conquest, no burn layers, no mass graves. The rivers shifted. The monsoons changed. The cities depopulated slowly, like a party where the guests leave one by one and the host doesn’t notice until the music stops. Within a few centuries, the great urban centers were abandoned, their bricks scavenged for later construction, their streets buried under silt and sand and the quiet indifference of time.
What survived was small. Pottery. Beads. Figurines. And, above all, the seals.
There are about four thousand known Indus inscriptions, found on seal stones, pottery, copper tablets, and other objects recovered from more than sixty archaeological sites across what is now Pakistan and northwestern India. The seals themselves are typically made of steatite — a soft stone that hardens when heated, which is a rather elegant piece of materials engineering for a civilization that supposedly had not yet invented the idea of an idea. They are small, rarely larger than a few centimeters on a side, and they were clearly made to be pressed into something soft — clay, wax, perhaps a surface we have not identified.
On one face: an animal, usually a so-called unicorn (which is almost certainly a bull shown in profile, with both horns overlapping into one, but “unicorn” has stuck because archaeology loves a good name more than a good explanation). The animal is exquisite. If you have not seen one, stop reading right now and search for “Mohenjo-daro seal M-1.” I’ll wait. I have nowhere else to be. Look at the musculature. Look at the line work. Look at the strange object beneath the animal’s head, which has been called a “standard,” a “manger,” a “ritual object,” and a “filter” — which tells you exactly how much we know about it, which is nothing.
Above the animal: the signs. A short sequence, usually three to seven symbols, arranged in a line. They are read right to left, which we know from the spacing — the signs are sometimes cramped on the left side of the seal, as though the carver ran out of room, which is a charming human touch in an otherwise inscrutable system. The signs are diverse: roughly four hundred distinct types have been identified, though the most common sixty-seven account for about eighty percent of all occurrences. This is a Zipfian distribution, for those who care about such things, and you should, because it means the signs are not random noise. Random noise doesn’t follow Zipf’s law. Language does. And, as it turns out, so do structured codes.
That is more or less the sum total of what was known, with certainty, before our research began. Four thousand inscriptions. Four hundred sign types. Average length of five. Read right to left. Associated with animals. Meaning: unknown.
Unknown for a hundred years.
Let that sink in. A hundred years of professional scholarship, hundreds of published papers, dozens of proposed decipherments, at least three major international conferences, one million-dollar prize, and the answer to the most basic question — what do these signs mean? — remains: we don’t know.
But here is the thing about “we don’t know.” It is not a wall. It is a door. It is the most productive sentence in science, because it is the only one that leads somewhere. Every other sentence — “we know,” “it is obvious,” “the evidence clearly shows” — is a destination. “We don’t know” is a departure. And in March 2026, that is exactly where we departed from.
Not to find out what the signs say. We may never know that. But to find out how the system works — which turns out to be a question that the signs themselves are eager to answer, if you ask it correctly.
The Indus Valley Civilization existed for roughly seven hundred years. In that time, it produced some of the most sophisticated urban planning in the ancient world, maintained trade networks stretching to the Persian Gulf, and developed a standardized system of weights and measures so precise that the smallest weight — about 0.86 grams — could measure out individual beads of carnelian. These were not people who did things approximately. These were people who counted. Everything. With an intensity that borders on the devotional.
And the seals, as we will see, were part of how they counted. Not in the mystical sense. Not in the sense of “sacred numerology.” In the brutally practical sense of a civilization that needed to track who owned what, who shipped what, who was authorized to do what, across dozens of cities and thousands of kilometers. The seals were not holy objects, though they may have been beautiful ones. They were functional. They were administrative. They were, in the most precise meaning of the word, information technology.
Consider what that means. Four thousand years before the first database, before the first spreadsheet, before the first punch card, before Charles Babbage had the idea of having ideas about computation — a civilization in the floodplain of the Indus River developed a system for encoding identity and transaction information in a format so robust that it survived the collapse of the cities, the erosion of the rivers, and forty centuries of absolute silence. The system outlasted everything: the language, the religion, the government, the people themselves. Only the format endured.
Eighty handshakes back, someone understood this. They held a seal, pressed it into clay, and produced an impression that said, in effect: this is mine, this is what I do, and this is my number. They did not know that their system would outlast their language, their religion, their memory, and their name. They did not know that four thousand years later, a machine would read the structure of their code and recognize it for what it was.
But they made it beautiful anyway. Which tells you something about the kind of people they were.
Chapter Two: The Riddle Nobody Could Solve
In 1875, a man named Alexander Cunningham published a drawing of a small stone seal he had found in the ruins of Harappa, in what was then British India. The seal bore an animal figure and a line of strange signs that Cunningham, who was otherwise an extremely competent archaeologist, could make neither head nor tail of. He noted it, catalogued it, and moved on. He had a subcontinent to survey, and one odd seal was not going to slow him down.
He could not have known that this single object was the first visible thread of a mystery that would consume scholars for the next century and a half. The Indus script had entered the modern world, and it had absolutely no intention of cooperating.
The full excavation of Mohenjo-Daro and Harappa began in the 1920s under the direction of Sir John Marshall, and it became clear almost immediately that the seals were not anomalies. They were everywhere. Thousands of them, scattered through the ruins like business cards at a conference that ended badly. Each one bore the same basic format: animal below, signs above, pressed into stone with a precision that suggested both standardization and care. The civilization that made them clearly attached great importance to them. The question was: what did they say?
The early attempts at decipherment were enthusiastic and, in retrospect, hopelessly premature. Scholars compared the Indus signs to every known script in the neighborhood: Sumerian, Proto-Elamite, Brahmi, even Easter Island’s rongorongo, which is about as geographically plausible as comparing Toronto street signs to Martian rock formations. Each comparison produced tantalizing similarities and zero results. The signs stubbornly refused to map onto any known writing system.
The problem, as it gradually became clear, was structural. Most successful decipherments in history have relied on one of three things: a bilingual text (like the Rosetta Stone), a known language (like the Greek that helped unlock Linear B), or a large corpus of varied texts (like the thousands of cuneiform tablets that eventually yielded Sumerian). The Indus script had none of these advantages.
There was no bilingual inscription. Despite decades of hoping, no Rosetta Stone of the Indus has ever been found. There was no confirmed underlying language. The leading candidates — Proto-Dravidian and early Indo-Aryan — were both contested, and neither could be proven from the signs alone. And the texts were maddeningly short. The average Indus inscription contains about five signs. The longest known inscription has perhaps seventeen. Five signs is not a sentence. It is barely a phrase. Trying to decode a language from five-sign samples is like trying to reconstruct a novel from a handful of license plates.
And yet, for a hundred years, that is precisely what people tried to do. The history of Indus script decipherment reads like a novel in which every character is convinced they are the protagonist and none of them are right. In the 1920s and 1930s, scholars proposed connections to Brahmi. In the 1960s and 1970s, a wave of Soviet and Finnish scholars pursued the Dravidian hypothesis with mathematical rigor. Asko Parpola developed an ingenious rebus-based reading that proposed Dravidian words for the pictographic signs: a fish for “min,” meaning both “fish” and “star” in Dravidian languages. It was elegant. It was plausible. It was never confirmed.
In India, the decipherment became entangled with politics in ways that would be comic if the stakes were not so high. The question of whether the Indus people spoke a Dravidian or Indo-Aryan language was not merely academic — it was a proxy war for competing visions of Indian identity. Tamil politicians championed the Dravidian hypothesis because it supported the claim that southern Indian culture predated the north. Hindu nationalists favored the Indo-Aryan hypothesis because it placed Vedic civilization at the headwaters of the Indus. The script, mute and impartial, became a screen onto which everyone projected their preferred history.
Meanwhile, the signs sat on their seals and said nothing.
By the late twentieth century, the field had settled into a pattern of diminishing returns. New decipherment claims appeared regularly, each one greeted with brief excitement and long skepticism. A cottage industry of amateur decipherers emerged, many of whom had read exactly one book about the Indus Valley and were now convinced they had cracked the code. Professional scholars grew weary. The Indus script became what academics call a “career graveyard” — a problem so resistant to solution that working on it was widely considered professional suicide, which is the academic equivalent of actual suicide but with worse catering.
And then, in 2004, the bomb went off.
Steve Farmer, a comparative historian; Richard Sproat, a computational linguist; and Michael Witzel, a Harvard Sanskritist, published a paper titled “The Collapse of the Indus-Script Thesis.” Their argument was simple and devastating: the Indus signs are not writing. They are too short, too few, and too formulaic to encode language. They are, at best, a collection of nonlinguistic symbols — marks of identity, religious motifs, clan emblems — but not a script in any meaningful sense.
The reaction was volcanic. In India, the paper was treated as an insult to national heritage. Farmer reportedly received death threats. Sproat was attacked in print and online. Witzel, already controversial for his work on Vedic chronology, became a lightning rod for nationalist anger. The academic response was only slightly more measured: critics accused the authors of Western bias, insufficient data, and a failure to appreciate the complexity of short-text systems.
But the paper also did something that none of the critics fully acknowledged: it forced the field to confront its own assumptions. For a century, everyone had assumed that the Indus signs encode language. Farmer, Sproat, and Witzel asked: what if they don’t? And if they don’t, what are they?
That question — not what the signs say, but what kind of system they belong to — turned out to be the most productive question anyone had asked in a hundred years. It just took another twenty years, and a machine, to answer it.
The answer, when it came, was not what anyone expected. The signs are not language. Farmer, Sproat, and Witzel were right about that. But neither are they random, decorative, or trivial. They are a structured information system — rule-governed, positionally constrained, statistically regular, and combinatorially vast. They are, in short, the opposite of nothing. They are everything you need to run a civilization without an alphabet.
The riddle was never “what language do the signs encode?” The riddle was “why did we assume they had to encode a language at all?” And the answer to that question says more about us than it does about the Indus Valley. It says that we cannot imagine intelligence without speech, structure without grammar, civilization without writing. It says that our definition of “advanced” is embarrassingly narrow. It says that for a hundred years, we looked at the most sophisticated non-linguistic information system in the ancient world and called it a mystery, because we didn’t have a category for what it actually was.
We do now.
Chapter Three: The Concordance
Iravatham Mahadevan was not supposed to become the most important Indus script scholar of the twentieth century. He was supposed to become a bureaucrat, and for most of his career, he was an excellent one.
Born in 1930 in Thanjavur, Tamil Nadu, Mahadevan joined the Indian Administrative Service in 1954 and spent the next twenty-seven years doing the things that IAS officers do: running departments, writing policy, managing the machinery of a young and chaotic democracy. He was by all accounts diligent, effective, and quietly brilliant — the kind of civil servant who makes countries function while politicians take the credit and journalists take the photographs.
But Mahadevan had a problem. He was interested in things that had nothing to do with his job. Specifically, he was interested in the scratchy little marks that appeared on rocks in caves and on pottery shards in Tamil Nadu — marks that most people assumed were decorative, or random, or left by bored goatherds with sharp sticks and too much time on their hands. Mahadevan thought they were writing. He was right.
In the early 1960s, a chance encounter with the historian K. A. Nilakanta Sastri set him on a path that would define his second life. Sastri mentioned, almost casually, that the cave inscriptions of Tamil Nadu were an unsolved problem. Could Mahadevan have a look? He could. He did. And what he found was Tamil-Brahmi — the earliest known form of Tamil writing, dating to the third century BCE, which established beyond doubt that Tamil literacy stretched back far earlier than most scholars had assumed.
This was important work, and it would have been enough for most people. But Mahadevan was not most people. He had caught the bug — the particular, incurable bug that afflicts anyone who spends time with ancient inscriptions. The symptoms include a tendency to squint at rocks, an inability to walk past a museum without entering it, and a growing conviction that the most important things in the world are very, very small marks on very, very old surfaces. There is no known cure, and the prognosis is a lifetime of quietly muttering in dead languages.
Mahadevan wanted the Indus script.
In 1970, he received the Jawaharlal Nehru Fellowship to study the Indus signs. He was forty years old, still a serving IAS officer, and about to embark on the project that would consume the rest of his life. Over the next seven years, working with a grant from the Indian Council of Historical Research and published by the Archaeological Survey of India, he produced a book that is still, nearly half a century later, the single most important reference work in the field.
It is called The Indus Script: Texts, Concordance and Tables, and it was published in 1977. If you have never heard of it, that is because it is exactly the kind of book that changes everything without anyone noticing — the scholarly equivalent of the person who lays the foundation of a building and is never mentioned again when the ribbon is cut. It is not a decipherment. It does not propose a reading of the signs. It does not argue for Dravidian or Indo-Aryan or any other language. What it does is something far more fundamental and far less glamorous: it catalogs every known Indus inscription, sign by sign, position by position, with cross-references, frequency tables, and concordances that allow any researcher to see, at a glance, which signs appear where, how often, and in what combinations.
This sounds dry. It is dry. It is the driest book you will ever be grateful for. Because without it, nothing that followed — not Parpola’s rebuses, not Rao’s entropy analysis, not Mukhopadhyay’s administrative interpretation, and certainly not our own statistical work — would have been possible. The concordance is the foundation. Everything else is built on Mahadevan’s patient, meticulous, unglamorous counting.
Gregory Possehl, a distinguished American archaeologist, called Mahadevan “a careful, methodical worker, taking care to spell out his assumptions and methods.” He noted that Mahadevan preferred “tentative conclusions” and “working hypotheses” over “set ideas and fait accompli.” This is the highest praise you can give a scholar, and it is the kind of praise that never makes headlines. Nobody writes a news article about a man who carefully counted things. But that careful counting is the bedrock of all science. Without it, you are just guessing loudly.
Mahadevan retired voluntarily from the IAS in 1981, at the age of fifty-one. He had given the government twenty-seven years. He would give the Indus script thirty-seven more. He spent the last three decades of his life refining his concordance, publishing papers on the Dravidian connection, establishing the Indus Research Centre at the Roja Muthiah Research Library in Chennai, and corresponding with scholars around the world. He received the Padma Shri, one of India’s highest civilian honors, in 2009. He was, by then, eighty years old and still working.
There is a detail about Mahadevan that captures the man better than any biography. During a visit to a temple, he was observed running his index finger along the carved inscriptions, feeling the grooves in the stone. When asked about it, he held up his hand and said: “Can you tell me why this finger is shorter than this one?” pointing to his index and middle fingers. “That’s how much this finger has been run on rough rock-faces to sense the letters.” The man had literally worn down his own body reading stones. This is not scholarly dedication. This is love.
He died on November 26, 2018, at the age of eighty-eight. His ashes, at his request, were immersed in the Bay of Bengal. He had asked that no religious rituals be performed. He was, in the end, a man of evidence — a man who believed that data, carefully gathered and honestly presented, was the only thing that lasted.
He was right about that, too.
The concordance Mahadevan built is now the only fully open-access corpus of the Indus script. Other databases exist — notably the Interactive Corpus of Indus Texts, maintained by scholars in Finland and India, which contains roughly five thousand five hundred entries — but they require institutional access and are not freely downloadable. Mahadevan’s data, digitized and structured into machine-readable format by independent researchers and hosted on GitHub, is the only corpus that anyone with a laptop and an internet connection can use without asking permission.
This is not a trivial point. When we began our analysis in March 2026, the choice of corpus was not incidental. We could have requested access to the larger databases. We chose not to. The reason is simple and, we think, important: if your central claim is statistical, then your credibility depends entirely on reproducibility. Anyone who reads our paper must be able to download the same data, run the same code, and get the same numbers. If the data is locked behind institutional gates, the work cannot be verified by the majority of the scholarly community. We prioritize transparency over sample size. That is a philosophical decision as much as a methodological one, and we owe it, ultimately, to Mahadevan’s example. He built a concordance, not a fortress. He wanted people to use the data, not to admire it from behind glass.
The specific corpus we used contains one hundred and seventy-nine Mohenjo-Daro unicorn seals. One hundred and seventy-nine is not a large number. It is, in fact, a small number, and several reviewers have pointed this out with the gentle persistence of people who are paid to find problems. But it is a clean number. Every seal in the corpus comes from one site, depicts one animal type, and is formatted in structured JSON with grapheme-level metadata. It is, in other words, a controlled sample — exactly the kind of dataset you want when your goal is to detect positional patterns rather than to catalog the entire sign inventory.
From those one hundred and seventy-nine seals, we extracted one thousand and three sign tokens belonging to one hundred and eighty-two unique types. The mean inscription length was five point six signs, with a median of five. The shortest inscription was one sign. The longest was thirteen. And from this modest dataset — this tiny, lovingly preserved fragment of a vanished world — every statistical result in this book was derived.
Every number in our paper traces back to Mahadevan. Every pattern we found was hiding in the data he compiled. He built the concordance in 1977. We ran the statistics in 2026. Between those two dates: forty-nine years, half a dozen competing theories, three major academic scandals, and one million-dollar prize. And still the foundation held. Because Mahadevan did not build a theory. He built a tool. And tools, unlike theories, do not go out of fashion. They do not get attacked in journals or defended at conferences. They simply work, quietly, until someone picks them up and uses them for something their maker never imagined.
In January 2025, the Government of Tamil Nadu named its Indus script prize after him. The Iravatham Mahadevan Prize: one million dollars for the decipherment of the script. It is a fitting tribute — not because Mahadevan himself cracked the code, but because he made it possible for anyone to try. The man who counted the signs never got to read them. But he made sure that the next person could.
That is what a concordance does. It does not answer questions. It makes questions answerable. And that, in the long run, is worth more than any answer.
Chapter Four: What Mahadevan Built
Now that you have met the man, let us meet his data.
The corpus we used for our analysis is not the concordance itself — Mahadevan’s 1977 book is a physical volume, and while it remains the authoritative reference, it is not something you can feed to a computer. What we used is a digitized version: a GitHub repository maintained by independent researchers who took Mahadevan’s meticulous cataloguing and converted it into structured, machine-readable JSON files. Each file represents one seal. Each seal contains a list of graphemes — the individual signs — with their identifiers, positions, and metadata.
If you have never looked at a JSON file, it is essentially a way of organizing information that computers find congenial and humans find tolerable. It looks like a series of nested brackets and labels, and it reads like a tax form written by someone who actually enjoys tax forms. But inside those brackets lies the raw material of discovery: every sign, on every seal, in every position, tagged and ready for analysis.
The repository contains data from Mohenjo-Daro, organized into two folders: M-1 through M-99 and M-100 through M-199. We restricted our analysis to the unicorn seals — those bearing the iconic single-horned animal — because they form a coherent, internally comparable subset. Mixing unicorn seals with bull seals, elephant seals, and rhinoceros seals would be like analyzing license plates from five different countries at once: possible, but noisy. The unicorn seals gave us a clean signal.
One hundred and seventy-nine seals. One thousand and three sign tokens. One hundred and eighty-two unique sign types. A mean inscription length of 5.6 signs. A median of five. These are small numbers, and we have never pretended otherwise. A critic could reasonably ask whether 179 seals are enough to draw conclusions about a civilization that produced thousands. The answer is: it depends on what you are looking for.
If you are looking for a complete dictionary of the Indus sign system, 179 seals are woefully insufficient. If you are looking for a phonetic decipherment, they are useless. But if you are looking for structural patterns — for the rules that govern how signs are arranged, which positions they prefer, and how they relate to one another — then a controlled, homogeneous subcorpus is not a limitation. It is an advantage. You cannot hear a melody in a crowd. You need a quiet room.
The first thing we did was the simplest: we listed every inscription as a sequence of sign identifiers and checked how many were unique. This is the kind of analysis that sounds trivial until you see the result. Of 179 inscriptions, 176 were unique. Three were repeated — and even those repetitions involved only two distinct sequences, one appearing three times and one appearing twice. The uniqueness rate was 98.3 percent.
Ninety-eight point three percent. Let that number settle in your mind. If these seals were prayers, you would expect repetition — the same prayer on many seals, just as the same verse appears in many hymnals. If they were names of gods or clans, you would expect clustering — a few dominant sequences appearing over and over. If they were random, you would expect a certain level of accidental repetition based on the frequency of the signs involved. We tested this with a permutation model: we shuffled the signs randomly across the same inscription lengths ten thousand times, preserving the overall frequency of each sign, and asked how often the shuffled corpora achieved the same uniqueness rate as the real data.
They never did. In ten thousand random shuffles, the null model produced a mean uniqueness of around 91 percent. The observed 98.3 percent was off the chart. The probability of achieving this level of uniqueness by chance was less than one in a thousand. The p-value was below 0.001, which in statistical language means “this is not an accident.”
What does 98.3 percent uniqueness mean in plain language? It means that almost every seal in the corpus carries a distinct inscription. It means the system was designed to produce difference, not repetition. It means that whatever these signs encode, it is something that needed to be different for each seal — like a name, a registration number, a transaction code. Something individual. Something identifying.
This single result, by itself, rules out several categories of interpretation. The seals are not standardized religious inscriptions. They are not clan symbols. They are not fixed-text amulets. They are not the Bronze Age equivalent of “Live, Laugh, Love” printed on a kitchen wall. Each one says something different, and the system that produced them was explicitly designed to ensure this.
We did not know, at this point, what exactly they said. We still don’t. But we knew something almost as valuable: we knew the kind of thing they were. And the kind of thing they were is an identification system. A registration code. A structured identifier unique to its bearer or its transaction.
The world’s oldest known barcode. Carved in stone. By hand. Four thousand years ago.
If that does not make you want to know more, you may close this book. But I suspect you won’t.
Chapter Five: What Survives Four Thousand Years
Let us conduct a thought experiment. It is the year 6026. Four thousand years have passed since you read this sentence. Toronto, the city where this book was written, no longer exists. Not in the romantic, archaeological sense of “ruins” — there are no picturesque columns or moss-covered walls. Toronto is simply gone. The concrete has powdered. The steel has rusted into unrecognizable lumps of iron oxide. The glass has shattered and been ground to sand by four millennia of freeze-thaw cycles. The CN Tower, that absurd and beloved needle, collapsed sometime in the twenty-third century and was scavenged for scrap before the scrap itself corroded into nothing.
The books are gone. Paper lasts a few centuries under ideal conditions. Under non-ideal conditions — floods, fires, the slow gnawing of insects and mold — it lasts decades. Every novel, every newspaper, every legal document printed before the digital age has disintegrated. The digital records lasted even less. Hard drives fail in years. Servers require power, cooling, and maintenance. When the infrastructure collapsed — and it did collapse, because infrastructure always does when nobody is left to maintain it — the cloud evaporated. Every email, every database, every social media post, every medical record and financial transaction and family photograph vanished in a cascade of silent magnetic decay.
The plastics have degraded into microparticles, dispersed through soil and ocean sediment. A careful chemist in the year 6026 might detect anomalous polymer concentrations in certain geological layers and wonder what produced them. She would be right to wonder. But she would not find a plastic bottle, or a keyboard, or a credit card. Those things do not last four thousand years. Nothing we make lasts four thousand years.
Almost nothing.
Stone lasts. Ceramic lasts. Gold lasts, because gold is too lazy to react with anything. Certain alloys last, under certain conditions. And structures carved into stone — if the stone itself survives erosion, burial, and tectonic activity — can persist for extraordinary periods.
The Indus seals are made of steatite. Steatite is a form of soapstone — soft enough to carve with precision, but capable of being hardened by heating. Once fired, it becomes remarkably durable: resistant to water, to acid, to the kind of everyday chemical assault that destroys most materials. A steatite seal, properly made, can survive more or less intact for thousands of years under burial conditions. And many of them did. The seals excavated at Mohenjo-Daro are, in many cases, in superb condition. The lines are sharp. The details are clear. The signs are legible. Four thousand years of silence, and the stamps look like they were pressed this morning.
This is not an accident. Whoever designed these seals chose their material with the same care they chose their format. Steatite was not the easiest material to work with — it required skill, tools, and access to quarries. But it was the right material for an object that needed to endure. The choice of stone was, itself, an information decision: it said, “This mark is permanent. This identity is not temporary.”
Now consider what this implies about us. If our civilization ended tomorrow — not through a dramatic apocalypse, but through the kind of slow decline that has ended every other civilization in history — what would a researcher find in four thousand years? What would survive to represent us?
Not our literature. Not our science. Not our music, our films, our philosophies, our laws. Not even our buildings, most of which are designed to last decades, not centuries, let alone millennia. What would survive would be precisely the things we value least: landfills, nuclear waste deposits, certain types of industrial ceramics, and — if we were very lucky — a few objects made of materials we chose for durability rather than convenience.
We have no such objects. We have nothing equivalent to the Indus seals. Our identification systems are printed on plastic cards with magnetic stripes that degrade in years. Our passports are booklets of laminated paper. Our most important data is stored on devices that require electricity to read and that no one will know how to power in five hundred years, let alone four thousand.
The Indus people, by contrast, built their identification system on stone. They encoded their information in physical form — carved, not printed; pressed, not displayed. They chose a format that would outlast the language it might have been associated with, the religion that might have sanctified it, the government that might have issued it. The format survived because the format was designed to survive.
There is a lesson here, and it is not a comfortable one. We think of ourselves as the most advanced civilization in history. We have computers, satellites, nuclear power, genetic engineering. But we have not solved the problem that the Indus people solved four thousand years ago: the problem of permanence. Our information is abundant and fragile. Theirs was sparse and indestructible. Their seals will outlast our entire digital civilization by millennia. And the information encoded on them — their positional structure, their uniqueness, their numeral patterns — is still readable today, by anyone with a laptop and a copy of Python.
We cannot say the same about a single thing we have built.
The archaeologist of the year 6026, sifting through the geological stratum that was once North America, will find anomalous isotope ratios, microplastic residues, and perhaps a few ceramic fragments. She will know that something was here. She will not know what we thought, what we loved, what we argued about. She will not know our names. She will have no concordance.
The archaeologist of 2026, sifting through the excavation layers of Mohenjo-Daro, has the seals. Has the signs. Has the structure. Has the data. And from that data, she can reconstruct something astonishing: not what the Indus people said, but how their system worked. Not their words, but their logic. Not their language, but their information architecture.
That is what survives. Not meaning, but format. Not content, but structure. And structure, as we will see, turns out to be enough.
Chapter Six: The Fish and the Star
Asko Parpola has been trying to read the Indus script since 1964. He was thirty years old when he began, a young Finnish scholar with a background in Sanskrit and a taste for puzzles. He is now ninety-three. He has not stopped.
Over the course of six decades, Parpola has produced the most sustained, rigorous, and imaginative attempt at phonetic decipherment that the field has ever seen. His approach is built on a deceptively simple idea: the rebus principle. A rebus is a way of using a picture to represent a sound rather than an object. If you draw a picture of an eye followed by a picture of a can, you can read it as “I can.” The picture of the eye does not mean “eye” — it means the sound “I.” The meaning has been stripped away; only the sound remains.
Parpola proposed that the Indus signs work the same way, and that the language hidden beneath them is Proto-Dravidian — the ancestor of the Dravidian language family, which today includes Tamil, Telugu, Kannada, and Malayalam, spoken by over two hundred million people in southern India. His most famous reading involves the fish sign, one of the most common symbols in the Indus script. In Dravidian languages, the word for fish is “min.” The word for star is also “min.” Parpola argued that the fish sign, when appearing in certain combinations, does not represent a fish at all. It represents a star — or, more precisely, a deity associated with a star.
This is elegant. It is linguistically sophisticated. It draws on deep knowledge of Dravidian etymology, Vedic mythology, and comparative symbolism. Parpola’s 1994 book, Deciphering the Indus Script, remains the most comprehensive treatment of the phonetic approach, and it is a monument to scholarly dedication. The man has spent more time inside the Indus sign system than perhaps any living person.
And yet, after sixty years, the decipherment remains unconfirmed.
This is not because Parpola is wrong. It may be that he is right — partially, at least. The fish-star connection is plausible. The Dravidian substrate in the Indus region is well attested. There is no inherent reason why some Indus signs could not encode Dravidian phonetic values. The problem is that there is no way to prove it. Without a bilingual text, without a long inscription, without independent confirmation from a second source, any phonetic reading of the Indus script remains — in the strictest scientific sense — a hypothesis. A very good hypothesis, supported by a lifetime of careful work. But a hypothesis nonetheless.
Parpola knows this. He has never claimed certainty. In his writing and his lectures, he is careful to distinguish between what he considers established and what remains speculative. He uses phrases like “I suggest” and “it seems likely” with a precision that puts most scholars to shame. He is not a showman. He is a craftsman, and he knows the difference between a finished product and a work in progress.
In October 2024, Parpola wrote to our co-author, Treasure A. Hunt, in response to her position paper on non-phonetic approaches to the Indus script. His words were characteristically generous and precise. He called her paper “a welcome exception to the papers now poured on the decipherment of the Indus script,” while maintaining his commitment to phonetic decipherment. This is the mark of a true scholar: the ability to recognize value in an approach that contradicts one’s own, while standing firm in one’s convictions. It is also, incidentally, exactly the attitude that the Indus script field needs more of and has historically had almost none of.
So where does Parpola fit in our story? He fits in the same place the fish fits in the sign: not where you expect.
Our analysis does not contradict Parpola. It reframes him. If the Indus seals are structured identification codes — registration numbers, transaction markers, administrative identifiers — then the signs that compose them may have multiple layers of meaning. The positional structure is one layer: a fixed format that governs where certain types of signs can appear. The numeral values are another layer: stroke marks that encode quantities. But beneath these administrative layers, there may well be a semantic layer — a level at which individual signs carry meaning derived from the objects they depict or the sounds those objects represent.
Parpola’s fish-star reading, in this model, is not wrong. It is deep. It may capture something real about the origin and semantic content of individual signs, even if the system as a whole functions as a positionally constrained code rather than as written language. The fish sign may indeed have been chosen because of its Dravidian phonetic value. But on the seal, it may function not as a word but as an element in a code — a position-specific identifier whose meaning is structural rather than linguistic.
Think of it this way. The letters in a license plate are drawn from the Roman alphabet. Each letter has a name, a sound, a history stretching back through Latin to Phoenician to Egyptian. The letter A descends from the Phoenician aleph, which depicted an ox. But when you see the letter A on a license plate, you do not think of oxen. You think of a registration code. The semantic history of the symbol is real, but its function in the system is structural. Both layers coexist. One does not cancel the other.
This is what we believe is happening in the Indus script. Parpola may have found the ox. We found the license plate. And both discoveries are necessary to understand the system as a whole.
In our paper, Section 6 — the reconciliation section — is dedicated to this idea. We argue that the major approaches to the Indus script are not contradictory but complementary. Each one captures a different layer of a multi-layered system. Parpola captures the semantic layer. Farmer, Sproat, and Witzel correctly identify the non-linguistic administrative function. Mukhopadhyay identifies the economic domain. Rao and Yadav confirm the statistical regularity. And our analysis maps the positional architecture that holds it all together.
Nobody was entirely wrong. Nobody was entirely right. The Indus script was too complex for any single approach to capture, and too simple for any single approach to decode. It sat, for a century, in the space between language and non-language, between writing and symbol, between meaning and structure — waiting for someone to realize that the space between categories is itself a category.
Parpola spent sixty years staring at fish and seeing stars. That is not failure. That is vision. And the stars he saw may yet turn out to be real — just not in the way anyone, including Parpola himself, originally expected.
Chapter Seven: The Entropy of Signs
Entropy is one of those words that sounds like it should be complicated, and is, but not in the way you expect. In everyday speech, entropy means disorder — the slow unraveling of things, the tendency of your desk to accumulate papers and your life to accumulate complications. In thermodynamics, it measures the number of ways a system can be arranged without changing its outward appearance. In information theory, which is where we need it, entropy measures something simpler and more profound: surprise.
Imagine you are watching someone type a message, one letter at a time. If the next letter could be anything — any of the twenty-six letters in the English alphabet, with equal probability — then each letter carries maximum surprise. You have no idea what is coming. The entropy is high. But if you already know the message is in English, your uncertainty drops. After the letter Q, the next letter is almost certainly U. After T-H, the next letter is probably E. Context reduces surprise. The entropy falls.
Shannon entropy, named after Claude Shannon, the father of information theory (and no relation to our AI colleague, Claude, who was named after Claude Shannon, which makes the whole thing pleasantly recursive), quantifies this surprise in bits. A perfectly random system has maximum entropy. A perfectly predictable system has zero entropy. Everything interesting happens in between.
When we applied Shannon entropy to the Indus seal inscriptions, we were asking a specific question: does the amount of surprise vary by position? In a natural language, you would expect some positional variation — the first word of a sentence is less predictable than the last, because grammar constrains endings more than beginnings. But in a structured code, positional entropy should show a very different pattern. The first position, if it marks a category, should be constrained — low entropy. The middle positions, if they carry unique identifiers, should be diverse — high entropy. The last position, if it marks a suffix or class, should be moderately constrained.
We looked at the thirty-five inscriptions in our corpus that were exactly five signs long. Five is the median length, and using a fixed length eliminates the noise introduced by comparing inscriptions of different sizes. For each of the five positions, we calculated the Shannon entropy of the signs that appeared there.
The results were not subtle.
Position one: 3.13 bits. This is low. Only fifteen unique signs appeared in this position, and the three most common accounted for sixty percent of all occurrences. The first position is a bottleneck — a narrow gate through which only certain signs pass. It behaves like a category marker: a prefix that classifies the seal into a group.
Position two: 4.44 bits. This is high — dramatically higher than position one. Twenty-five unique signs appeared here, with no dominant trio. The second position is wide open, drawing from a large and relatively uniform pool of signs. It behaves like an identifier — a variable that distinguishes one seal from another within the category set by position one.
Position three: 4.45 bits. Nearly identical to position two. Another high-diversity, high-entropy position. The middle of the inscription is where the system generates uniqueness. This is the engine room, the place where combinatorial power resides.
Position four: 3.77 bits. Lower than the middle positions but higher than the first. Twenty unique signs, with the top three accounting for nearly half of all occurrences. And here is where it gets interesting: this is where the stroke numerals cluster. Position four, in five-sign inscriptions, is the penultimate position — the second-to-last slot — and it is disproportionately occupied by simple vertical strokes representing the numbers one through five. We will return to this in Chapter Thirteen. For now, note the pattern: position four is numerically constrained.
Position five: 3.98 bits. Moderate entropy. Twenty-three unique signs, but with a noticeable concentration in a few common types. The final position behaves like a suffix — a closing marker that may indicate a terminal category, a status, or a check.
What you see, if you plot these numbers, is a profile. It rises from low to high, peaks in the middle, and descends toward the end, with a dip at the penultimate position where numerals cluster. This is not the profile of a language. In language, entropy tends to be more uniform across positions, because grammar distributes predictability more evenly. This is the profile of a structured code — a system where different positions serve different functions, just as different fields in a passport number or a license plate carry different types of information.
To make sure this pattern was not an artifact of our particular sample, we ran two additional tests. First, a Friedman test — a nonparametric statistical test that asks whether the entropy values across positions differ significantly from what you would expect if position did not matter. The result: chi-squared of 14.7, with a p-value of 0.005. Position matters. The differences are real.
Second, we calculated Kendall’s W, a measure of concordance that quantifies how consistently the positional pattern holds across individual inscriptions. The result: W = 0.105. This is a small but significant effect — meaning the pattern is consistent enough to be detected statistically, but not so rigid as to be mechanical. The system has structure, but it also has flex. It is rule-governed, not rule-imprisoned. Like a well-designed form: everyone fills in the same fields, but everyone writes something different.
We also computed bootstrap confidence intervals for the entropy at each position, resampling ten thousand times. The intervals confirmed that the differences between positions are genuine and not artifacts of small sample size. Position one is reliably low. Positions two and three are reliably high. Position four dips. Position five partially recovers. The profile is stable.
This is the moment when the seals stopped being mysterious and started being legible — not in the sense of readable, but in the sense of structurally transparent. We could not tell you what any individual sign meant. But we could tell you what each position in the inscription was doing. And that, it turns out, is enough to reconstruct the architecture of the system.
The format is the finding. I told you this sentence would come back.
Chapter Eight: The Language That Never Was
There is a peculiar assumption buried in the history of Indus script studies, so deeply embedded that most scholars never bother to examine it. The assumption is this: if a civilization is advanced, it must have writing. And if it has writing, the writing must encode a language. These two propositions seem so self-evident that questioning them feels perverse, like asking whether water is really wet. But they are not self-evident. They are cultural prejudices dressed up as axioms. And the Indus Valley Civilization is the proof.
We know, with certainty, that the Indus people were advanced. Their cities had grid-pattern streets, standardized brick dimensions, and drainage systems that would not be matched in Europe for three thousand years. Their weights followed a binary-decimal system of astonishing precision. They traded with Mesopotamia, the Persian Gulf, and Central Asia. They built the Great Bath of Mohenjo-Daro — a structure so carefully waterproofed that it would satisfy a modern building inspector, which is saying something, because modern building inspectors are satisfied by almost nothing.
And yet, despite all this sophistication, they may not have had writing in the way we normally understand it. They had a sign system — clearly. They had symbols that recurred in consistent patterns — obviously. But whether those symbols encoded a spoken language is a question that, after a hundred years of effort, remains unanswered. And perhaps the reason it remains unanswered is that the answer is no.
This is not a deficiency. This is a revelation.
Consider what writing does. Writing encodes speech. It takes the sounds that come out of a person’s mouth and converts them into visible marks that can be decoded back into sounds by another person at another time. This is an extraordinary invention. It allows language to survive the death of its speaker, to travel beyond the range of a human voice, to accumulate across generations. Writing is, in a real sense, the technology that made cumulative civilization possible.
But writing is not the only way to encode information. It is not even the oldest. Long before the first cuneiform tablet was pressed in Mesopotamia, people were using notches on sticks, knots in cords, tokens in clay envelopes, and marks on pottery to keep track of things — quantities, transactions, ownership, obligations. These systems encoded information, but they did not encode language. They were administrative, not linguistic. They recorded facts, not sentences.
The Inca quipu is perhaps the most famous example. A system of knotted strings, used across the Andes to record everything from census data to tribute payments to historical narratives, the quipu encoded information in a format that modern computer scientists recognize as structurally similar to a database. The position of a knot, the color of a string, the number of twists — each carried meaning. But it was not writing. It did not encode Quechua or any other spoken language. It was its own system, operating by its own rules, and it worked brilliantly for an empire that stretched four thousand kilometers without a single written word.
The Indus script may belong to the same family. Not a failed writing system. Not a primitive precursor to “real” writing. But a different kind of information technology altogether — one that encoded administrative data in a structured, positionally constrained format without passing through the intermediate step of language.
This is a genuinely radical idea, and it makes many people uncomfortable. In the West, we have spent two and a half millennia associating civilization with literacy. The Greeks defined barbarism as the absence of Greek writing. The Romans spread their empire through Latin documents. The Enlightenment enshrined the printed word as the vehicle of reason. To suggest that a civilization could be informationally sophisticated without being literate feels, to many scholars, like a contradiction in terms.
But it is not a contradiction. It is a category error — the error of assuming that information must take the form of language. The Indus seals prove otherwise. Their structure is not linguistic. Their statistics are not linguistic. Their function, as our analysis demonstrates, is administrative. They encode identity, not speech. They register transactions, not sentences. And they do so with a precision and consistency that rivals any contemporary administrative system.
Meanwhile, the spoken language of the Indus people — whatever it was — survived through a different channel entirely. It survived the way most human knowledge survived before writing: through oral transmission. Through stories told around fires, songs taught to children, rituals performed each generation with the exact same words, the exact same gestures, the exact same intonation. The Vedic tradition of India preserves hymns that were composed over three thousand years ago and transmitted orally with such precision that the pronunciation of individual syllables has been maintained across millennia. The word “Sanskrit” itself means “perfected” or “well-formed” — a language so carefully constructed and so meticulously preserved that it became a monument to the power of the human voice.
If the Indus people had a similarly rich oral tradition — and there is no reason to think they did not — then their language, their literature, their philosophy, and their mythology could have survived for centuries without a single written word. The seals did not need to carry language. Language was carried by people. The seals carried something else: structure. Format. Code. The administrative skeleton of a civilization that ran on information long before information learned to spell.
This is not a demotion of the Indus civilization. It is a promotion. It means they solved a problem that most civilizations solved by brute force — the problem of tracking identity and transactions across a vast territory — with an elegant, compact, non-linguistic system that was, in its own way, more advanced than writing. Writing is verbose. Writing requires grammar, syntax, morphology, and a reader who shares the writer’s language. The Indus code requires none of these things. It requires only structure. And structure, unlike language, does not need to be translated. It speaks for itself.
Four thousand years later, it still does.
Chapter Nine: The Bombshell
The year is 2004. The place is an academic journal called the Electronic Journal of Vedic Studies, which sounds like exactly the kind of publication that no one outside a university library would ever read, and you would be right. Its readership consists almost entirely of Indologists, Sanskritists, and a handful of computational linguists who have wandered in by accident and stayed for the arguments. It is, in other words, the last place you would expect a bomb to go off.
But a bomb went off.
Steve Farmer, Richard Sproat, and Michael Witzel published a paper titled “The Collapse of the Indus-Script Thesis: The Myth of a Literate Harappan Civilization.” The title alone was an act of war. In academic publishing, titles are usually diplomatic to the point of meaninglessness: “Toward a Revised Understanding of Sign Distribution in Mature Harappan Contexts” or something equally soporific. Farmer, Sproat, and Witzel chose a title that said: everything you believe is wrong, and here is why.
Their argument was built on a few key observations, each one delivered with the precision of a scalpel and the subtlety of a sledgehammer.
First: the inscriptions are too short. The average length is about five signs. The longest known inscription has about seventeen. No known writing system produces texts this consistently brief. Sumerian, Egyptian, Chinese, Maya — all produced long texts alongside short ones. The Indus signs appear almost exclusively in short sequences, and this brevity, the authors argued, is not an artifact of preservation. It is a feature of the system. Short texts do not encode language. They encode something else.
Second: the sign inventory is too large for a simple symbol set but too small for a full writing system. With roughly four hundred distinct signs, the Indus system falls into an uncomfortable middle ground. Alphabets use twenty to forty signs. Syllabaries use fifty to a hundred. Logographic systems like Chinese use thousands. Four hundred is too many for sound-based writing and too few for word-based writing. It fits, however, with a nonlinguistic symbol system — a set of icons, marks, and identifiers that communicate through convention rather than grammar.
Third: the signs appear almost exclusively on small, portable objects — seals, tablets, pottery. There are no long texts on walls, no monumental inscriptions, no literary works, no administrative archives of the kind found in Mesopotamia and Egypt. If the Indus people had a full writing system, where are the books? Where are the legal codes? Where are the accounts, the letters, the treaties? Their absence, the authors argued, is not a gap in the archaeological record. It is the archaeological record. The Indus people did not write long texts because they did not write.
The response was immediate, furious, and deeply personal.
In India, the paper was interpreted not as an academic argument but as a political attack. If the Indus script was not writing, then the Indus Civilization was not literate. And if it was not literate, then India’s claim to be the cradle of one of the world’s oldest literate traditions was weakened. For Tamil politicians who saw the Indus script as evidence of Dravidian primacy, and for Hindu nationalists who saw it as proof of Vedic antiquity, the paper was an insult to the nation. Farmer reportedly received death threats. The political response was so intense that the scholarly content of the paper was largely drowned out by outrage.
Among professional scholars, the reaction was more measured but still divided. Some linguists and archaeologists found the argument compelling. Bryan Wells and Andreas Fuls responded with statistical counterarguments. Rajesh Rao and his team published the 2009 Science paper on conditional entropy, arguing that the Indus signs showed linguistic-like statistical properties. Iravatham Mahadevan, in one of his last major interventions, defended the script thesis on structural grounds. The field, always fractious, split further.
But here is what nobody said at the time, and what this book argues now: Farmer, Sproat, and Witzel were right about the most important thing. The Indus signs are not linguistic writing. The statistical evidence — our evidence, from the positional entropy profile, the uniqueness rate, the bigram sparsity, the numeral clustering — all points to a non-linguistic system. The inscriptions are too unique, too positionally structured, and too formulaic to be encoding sentences in a spoken language. On this point, the 2004 paper was correct.
Farmer, Sproat, and Witzel proposed that the Indus signs constitute a nonlinguistic symbol system. This was correct — and, as Sproat has argued at length in his 2023 book Symbols (Springer), nonlinguistic systems can possess rich, complex structure. The error was not theirs but the field’s: the broader scholarly community heard “not writing” and interpreted it as “not interesting.” That misreading shut down the most productive line of inquiry for twenty years. If the signs are not language, then what kind of structured system are they? That question — the natural next step after Farmer, Sproat, and Witzel cleared the ground — was not asked with sufficient force until now. Our analysis attempts to answer it: a positionally constrained, combinatorially powerful registration code. This is entirely consistent with what Farmer, Sproat, and Witzel proposed. It simply adds the structural detail that their 2004 paper, focused on demolishing the linguistic thesis, did not pursue.
Because the Indus sign system is not simple. It is not primitive. It is not uninteresting. It is, as our analysis shows, a highly structured, positionally constrained, combinatorially powerful information system with the capacity to generate over a million distinct identifiers. It is rule-governed. It is consistent. It encodes multiple layers of information in a compact format. It is, in every functional sense, an engineered system. And engineering is not a lesser achievement than writing. It is a different achievement — and, for the specific problems the Indus Civilization needed to solve, it may have been a better one.
The 2004 paper cleared the ground. It demolished the assumption that the Indus signs must encode language, and in doing so, it opened the space for the question that nobody had thought to ask: if not language, then what? For twenty years, that space remained empty. Not because the question was unanswerable, but because the field was too angry, too polarized, and too invested in the old categories to walk through the door that Farmer, Sproat, and Witzel had kicked open.
We walked through it. And what we found on the other side was not a dead end but a landscape — a structured, measurable, analyzable system that had been waiting, in plain sight, for someone to look at it with the right tools.
So here is what we owe to the three most hated men in Indus studies. We owe them the question. We owe them the cleared ground. We owe them the courage to say that the emperor’s clothes, however beautiful, were not clothes at all — they were something else entirely. And that something else, it turns out, is far more interesting than anyone imagined.
They got death threats for this. History should give them something better.
Chapter Ten: The Format Is the Finding
There is a moment in every investigation when the evidence stops being evidence and starts being a picture. You stare at scattered data points for days or weeks or, in the case of the Indus script, a hundred years, and then something shifts. Not the data. The data was always there. What shifts is the frame. You were looking at trees. Now you see a forest. And the forest was always a forest. You just didn’t have the word for it.
For us, that moment came when we laid the results side by side. The uniqueness rate: 98.3 percent. The positional entropy profile: low at the start, high in the middle, dipping at the penultimate position, moderately constrained at the end. The bigram fill rate: 1.66 percent of all possible two-sign combinations actually appeared. The numeral clustering: stroke marks concentrated in position four of five-sign inscriptions. The near-duplicate pairs: eight pairs differing by exactly one sign, and the variation occurring exclusively in the middle positions, never at the edges.
Each of these results, individually, was interesting. Together, they were unmistakable. They described a system with the following properties: fixed format, variable content, positional constraints, embedded numerals, combinatorial power, and near-zero repetition. This is not the profile of a language. It is not the profile of a prayer book. It is not the profile of a collection of names, or clan markers, or religious amulets. It is the profile of a registration system.
Think about your driver’s license. The format is fixed: state code, followed by a sequence of letters and numbers, in a specific order, with specific lengths. The content is variable: your license number is different from everyone else’s. Certain positions carry specific types of information: the state prefix, the sequential identifier, perhaps a check digit. The system generates uniqueness by design — no two licenses are the same. And if you collected a thousand licenses and analyzed them statistically, without knowing what any of the characters meant, you would find exactly what we found in the Indus seals: positional entropy variation, near-total uniqueness, embedded numerals, and a compact, rule-governed structure.
Or think about a passport number. Or a tax identification number. Or an ISBN on a book. Or the barcode on a jar of peanut butter. All of these systems share the same deep architecture: they encode identity, category, and sequence in a positionally structured format that produces unique identifiers from a finite set of elements. The specific characters differ. The underlying logic is identical.
The Indus seals are the oldest known system with this architecture. Four thousand years before the International Standard Book Number was assigned to its first paperback, someone in Mohenjo-Daro was pressing a steatite seal into clay and producing an impression that functioned, structurally, in exactly the same way. Different technology. Same logic. Same purpose. Same result.
This is what we mean by “the format is the finding.” We did not decode the signs. We did not discover what language the Indus people spoke. We did not read a single word. What we discovered was the format itself — the positional architecture, the statistical signature, the structural DNA of the system. And that format tells us, with statistical certainty, what kind of system we are looking at. Not a writing system. Not a religious notation. An information system. A structured registration code. The world’s first.
Now, a reasonable person might ask: isn’t this a disappointment? After a hundred years, don’t we want to know what the signs say? The answer is yes, of course we do, and we may yet find out. But consider what we have gained in the meantime. We have gained the answer to a question that is, in some ways, more fundamental than “what do the signs mean?” We have gained the answer to “what kind of thing is this?”
For a century, the Indus script sat in a category called “undeciphered writing system.” That category shaped every attempt to understand it. Every scholar who approached the seals assumed they were looking at writing — abbreviated writing, perhaps, or highly formulaic writing, but writing nonetheless. And every attempt to read it as writing failed, because it is not writing. The category was wrong. The question was wrong. And a wrong question, no matter how brilliantly pursued, produces only frustration.
Farmer, Sproat, and Witzel destroyed the old category. They said: this is not writing. Good. That was necessary. But they did not build a new category. They said: nonlinguistic symbols. This is like saying: not a bird. Technically correct, but not very helpful if you are trying to identify what you are looking at.
Our analysis builds the new category. It says: this is a structured registration system. It is positionally constrained, combinatorially powerful, and designed to produce unique identifiers. It functions like a modern identification code. It is sophisticated, internally consistent, and engineered. And it tells us something remarkable about the civilization that created it: these people understood the principles of information management four thousand years before anyone else.
The format is the finding. Not because the format is all there is, but because the format is what survived. The language is gone. The religion is gone. The government is gone. The people are gone. But the structure endures — in the positions of the signs, in the patterns of their distribution, in the mathematics of their arrangement. And structure, unlike meaning, does not require translation. It speaks in numbers. And numbers, unlike words, do not lie.
Chapter Eleven: Search for M-1
Put this book down for a moment. Pick up your phone, or open your laptop, and type the following into any search engine: “Mohenjo-daro seal M-1.”
I am serious. Do it now. I will still be here when you get back.
What you will see is a small, square stone — roughly the size of a large postage stamp — bearing the image of what archaeologists call a “unicorn”: a bull-like animal shown in profile, with a single horn (probably two horns overlapping, but who among us has not had a bad angle in a photograph). Beneath the animal, a strange object on a pedestal, variously described as a brazier, a standard, a manger, or a sacred filter. Above the animal, a row of signs — compact, precise, incised into the stone with a confidence that makes your own handwriting look like a medical emergency.
Now look more closely. Look at the animal’s body. The ribs are suggested with fine, parallel lines. The musculature of the shoulder and haunch is rendered with the kind of anatomical awareness that Western art would not achieve for another two thousand years. The hooves are distinct. The tail hangs with a naturalistic curve. The eye is alive. This is not a symbol. This is a portrait — a portrait of an animal that someone observed with extraordinary care and reproduced with extraordinary skill, in a space smaller than a matchbox, using nothing but a sharp tool and a piece of heated soapstone.
And this was not a one-off. There are hundreds of these seals, and while the quality varies, the best of them are among the finest examples of miniature carving in the ancient world. The details are so precise that researchers can distinguish individual carvers by their style, the way art historians can identify a Rembrandt from a brushstroke. These were not assembly-line products stamped out by machines. They were crafted by skilled artisans who took pride in their work and who understood that the beauty of the object was not incidental to its function.
This is a point that deserves more attention than it usually receives. In discussions of the Indus script, the conversation almost always focuses on the signs — the mysterious symbols above the animal. The animal itself is treated as decoration, as if the carver added a nice unicorn the way a modern bank adds a watermark: pleasant but secondary. But the animal is not secondary. The animal is the seal’s first layer of information.
Consider: the overwhelming majority of Indus seals depict one of a small number of animals. The unicorn is the most common, accounting for roughly sixty percent of all seals. Then come the humped bull, the elephant, the rhinoceros, the tiger, and a few rarer animals. Each animal type appears with its own characteristic set of signs. The inscriptions on unicorn seals differ statistically from the inscriptions on bull seals or elephant seals. The animal is not random. It is a classifier — a visual marker that assigns the seal to a category before a single sign is read.
This is why we restricted our analysis to unicorn seals. By holding the animal type constant, we could isolate the variation in the signs themselves. But the fact that we needed to do this tells you something important about the system: it has layers. The animal is one layer. The signs are another. The strokes are a third. Each layer encodes a different type of information, and together they produce a multi-dimensional identifier of remarkable compactness.
A passport works the same way. The cover color tells you the country. The printed fields tell you the name, the birthdate, the nationality. The machine-readable zone at the bottom encodes the same information in a format that a scanner can read. Three layers, three formats, one identity. The Indus seal is the same architecture, realized in stone instead of laminated paper, four thousand years earlier.
But there is one dimension of the seal that no passport has ever matched, and that is beauty.
Nobody needs a passport to be beautiful. Its function is bureaucratic, and its aesthetic is accordingly grim — a bad photograph under a plastic sheet, sandwiched between pages of fine print that nobody reads. The passport does not aspire to art. It aspires to functionality, and it achieves it, with the charmless efficiency of a government form.
The Indus seal aspires to both. Its function is administrative — we have established this. But its execution is artistic. The carver who made seal M-1 did not merely produce an identifier. He produced an object of astonishing beauty, in a space so small that the work must have required a magnifying lens or preternatural eyesight. And he did this not because beauty was separate from function, but because beauty was part of function.
A beautiful seal is a secure seal. The more intricate the carving, the harder it is to forge. The fine details of the unicorn’s musculature, the precise curvature of the horn, the subtle rendering of the ritual object beneath — these are not decorations. They are security features. Like the microprinting on a banknote, or the hologram on a credit card, the artistry of the seal is what makes it trustworthy. You cannot fake what you cannot reproduce. And reproducing the work of a master carver, in steatite, at this scale, without modern tools, would have been close to impossible.
So beauty, in the Indus system, is not an indulgence. It is authentication. The seal says: this was made by a craftsman of recognized skill, for a person of recognized standing, within a system of recognized authority. The beauty of the object guarantees the legitimacy of the impression. Aesthetics and bureaucracy, art and administration, are not in conflict. They are collaborators. The unicorn is both a work of art and a security protocol.
Four thousand years later, the beauty still works. You looked at the image, and you felt something — recognition, perhaps, or wonder, or the strange vertigo that comes from seeing something ancient that looks modern. That feeling is the seal doing its job. It was made to command attention, to inspire trust, to declare authority. And it still does, even though the authority that issued it crumbled to dust before Rome was founded.
Not bad, for a piece of soapstone the size of your thumb.
Chapter Twelve: The Entropy Chasers
In 2009, a paper appeared in the journal Science — one of the two most prestigious scientific publications in the world — with the title “Entropic Evidence for Linguistic Structure in the Indus Script.” It was written by Rajesh Rao, a computer scientist at the University of Washington, along with Nisha Yadav and colleagues at the Tata Institute of Fundamental Research in Mumbai. The paper was short, as Science papers tend to be, but its impact was enormous. For the first time, someone had applied rigorous computational methods to the Indus signs and produced a result that was published not in an Indology journal but in the temple of mainstream science.
The argument was elegant. Rao and Yadav calculated the conditional entropy of the Indus sign sequences — the amount of uncertainty about the next sign given knowledge of the current one. They then compared this value to the conditional entropies of known linguistic scripts (Sumerian, Old Tamil, Sanskrit, English) and known nonlinguistic systems (DNA sequences, Fortran code, the Vinča symbols of Neolithic Europe). The result: the Indus signs fell squarely within the range of linguistic scripts and well outside the range of nonlinguistic systems.
The implication seemed clear: the Indus script behaves statistically like a language. It has the kind of sequential structure — the predictability, the patterning, the neither-too-random-nor-too-ordered quality — that characterizes writing systems. Therefore, the authors concluded, it probably is one.
The paper was a sensation. It made headlines around the world. It was covered by the BBC, the New York Times, National Geographic. After decades of stalemate, here at last was hard evidence — published in Science, no less — that the Indus script was real writing. The decipherment community rejoiced. The nonlinguistic camp seethed.
Richard Sproat, the computational linguist who had co-authored the 2004 “Collapse” paper, responded with the fury of a man who has been told that his carefully demolished building has rebuilt itself overnight. He published a detailed rebuttal arguing that the conditional entropy metric was too coarse to distinguish between language and structured nonlinguistic systems. He demonstrated that certain artificial datasets — randomly generated sequences with controlled statistical properties — could produce the same conditional entropy values as the Indus signs. In other words, the test could not tell the difference between a language and a well-structured code. Falling into the “linguistic range” did not prove you were a language. It only proved you had structure.
Sproat was right. And Rao and Yadav were also right. This is the kind of situation that makes scientists want to become plumbers.
What Rao and Yadav had demonstrated, beyond any reasonable doubt, was that the Indus signs carry structured information. Their sequences are not random. They exhibit the kind of sequential dependencies — where the identity of one sign constrains the probability of the next — that characterize complex systems of communication. This was a genuine and important finding, and it has held up under every subsequent analysis, including ours.
What they had not demonstrated was that this structure is specifically linguistic. Sproat’s counterexamples were valid: structured nonlinguistic systems can mimic the conditional entropy of language. A registration code, a DNA sequence, a musical composition — all can fall into the same entropic range as natural language, because all share the property of constrained sequential order. Entropy, as a metric, cannot distinguish between them. It tells you that a system has structure. It does not tell you what kind.
This distinction — between having structure and being a language — is precisely the distinction that our research resolves. Rao and Yadav proved structure. Farmer, Sproat, and Witzel denied language. Both were correct. But neither side had the tools or the inclination to ask the next question: if it has structure but is not a language, then what kind of structure is it?
Our positional entropy analysis answers that question. It shows not just that the signs are structured, but how they are structured — in a positionally constrained format where different positions serve different functions. This is not the structure of language, where grammar distributes constraints more evenly across positions. It is the structure of a code, where each field has a specific role. Rao and Yadav detected the signal. We identified the source.
Nisha Yadav, in a widely quoted statement, summarized the state of knowledge before our work with admirable honesty: “We still don’t know whether the signs are complete words, or part of words, or part of sentences. Our understanding is that the script is structured and there is an underlying logic in the writing.” She was right about the structure. She was right about the underlying logic. And our answer to her question is: they are none of the above. They are not words, or parts of words, or parts of sentences. They are fields in a code. Structured, yes. Logical, yes. But not linguistic.
Rao and Yadav’s contribution to the field cannot be overstated. Before their 2009 paper, the Indus script debate was an argument between philologists and archaeologists — people who dealt in texts, translations, and historical context. After their paper, the debate included computer scientists, statisticians, and information theorists. The tools changed. The questions sharpened. The standards of evidence rose. You could no longer wave your hands and say “it looks like writing to me.” You had to show the numbers.
We showed the numbers. And the numbers showed a code.
Yadav also said something else, something that captures the entire century-long mystery in a single sentence: “The one-million-dollar meaning of the signs and symbols remains elusive.” She was quoting the Tamil Nadu prize, but she was also, perhaps without intending to, stating a deeper truth: the meaning of the individual signs may always be elusive. What is not elusive, what is now firmly established, is the meaning of the system. The system is an information architecture. The signs are its elements. The format is its grammar. And the format, unlike the signs, speaks clearly.
The entropy chasers opened the door to computational analysis of the Indus script. We walked through it with a different question and found a different answer. But we could not have gotten there without them. Science, contrary to popular mythology, does not progress by lone geniuses having sudden revelations. It progresses by stubborn people building on the work of other stubborn people, one careful step at a time, until the accumulation of small advances suddenly resolves into something that looks, in retrospect, obvious.
Rao and Yadav took one of those steps. We took another. The next one belongs to whoever picks up the data and asks the next question. The code is open. The corpus is on GitHub. The tools are free. All that is required is curiosity — and the willingness to let the numbers speak for themselves, even when they say something you did not expect.
Chapter Thirteen: Counting Strokes
Among the four hundred or so distinct signs in the Indus repertoire, some are elaborate: fish, jars, figures with multiple limbs, intricate geometric forms that suggest either deep symbolic meaning or a carver with too much free time. But some signs are disarmingly simple. They are just lines. Vertical strokes, standing side by side like soldiers at attention. One stroke. Two strokes. Three. Four. Five. That is it. No frills, no flourishes, no ambiguity. Just lines.
These are the stroke signs, and they are the closest thing to a smoking gun that the Indus script has ever produced.
In our corpus of 179 unicorn seals from Mohenjo-Daro, stroke signs account for 154 of the 1,003 sign tokens — 15.4 percent of the entire dataset. They come in two varieties: short strokes (half-height) and tall strokes (full-height), each ranging from one to five lines. The short series is catalogued as P121 through P124; the tall series as P144, P145, P147, P150, and P151. Together, they form a coherent set that behaves, statistically, exactly the way you would expect numbers to behave.
The first clue is their distribution. If stroke signs were decorative or symbolic — if they meant “sacred” or “blessed” or “belonging to the temple” — you would expect them to appear more or less uniformly across all positions in the inscription. They do not. In five-sign inscriptions, strokes cluster overwhelmingly in position four — the penultimate slot. Forty-five point seven percent of all numerals in five-sign inscriptions appear in position four. At position one, the figure drops to 5.7 percent. At position five, the same. The strokes are not scattered. They are placed. Deliberately, consistently, in the second-to-last position.
This is exactly where you would expect a quantity field in a structured code. Think about how modern identification numbers work. A credit card number has a bank identifier prefix, a personal account number in the middle, and a check digit at the end. A vehicle identification number has a manufacturer code, a vehicle descriptor section, and a sequential production number. In these systems, the numerical component — the part that encodes a quantity or a sequence — occupies a specific, predictable position within the format. It does not float around. It sits where the format puts it.
The stroke signs in the Indus seals do the same thing. They sit in the penultimate position because the format requires them there. They encode a numerical value — how many of something, or which number in a sequence — and that value is part of a larger identifier that includes non-numerical signs in the other positions. The format governs. The strokes obey.
But here is where it gets genuinely strange.
The digit two accounts for 66.9 percent of all stroke numerals. Two-thirds. Not approximately two-thirds — exactly, crushingly, overwhelmingly two-thirds. One hundred and three out of 154 numeral tokens are the number two, whether in its short form (P122, which appears seventy-six times) or its tall form (P145, which appears twenty-seven times). The digit one accounts for 11.7 percent. Three accounts for 17.5 percent. Four and five together account for less than four percent.
We tested this distribution against a uniform model — the assumption that all digits should appear with equal frequency — and the result was a chi-squared value of 225.3 with four degrees of freedom, which corresponds to a p-value so small that it essentially registers as zero. The digits are not uniformly distributed. The digit two dominates. By a lot.
Why?
This is the most intriguing open question in our entire analysis, and we do not have a definitive answer. But we have hypotheses, and they are revealing.
One possibility is that the strokes encode a quantity of goods. If most transactions involved two units of something — two bales of cotton, two containers of grain, two measures of oil — then the preponderance of twos would reflect the economic reality of the marketplace. Indus trade, as far as we can reconstruct it, was organized around standardized units, and it is plausible that the most common transaction size was two.
Another possibility is that the strokes encode a class or grade rather than an absolute quantity. If the numeral field represents a tier within a hierarchy — first grade, second grade, third grade — then the dominance of two might mean that most seal holders belonged to the second tier of whatever system the seals tracked. This would be consistent with a Zipfian distribution of social ranks: many people in the middle, fewer at the top and bottom.
A third possibility, more speculative but worth noting, is that the strokes encode a weight class. The Indus weight system was binary-decimal: the smallest unit doubled to the next, and the doubling ratio was 1:2:4:8:16. If the numeral on the seal indicated which weight standard applied to the goods being traded, then two would correspond to the second unit in the series — the most commonly used weight for everyday transactions.
We cannot yet distinguish among these hypotheses. What we can say is that the stroke signs are numerals, that they occupy a fixed position in the inscription format, and that their frequency distribution is not random. Something real is being counted, measured, or classified. The system knows what the number two means. We don’t. Yet.
But the strokes whisper. And if you listen carefully, what they whisper is: this is not art. This is not prayer. This is accounting.
Chapter Fourteen: What Would Remain of Us
Let us play a game. It is the year 6026. You are an archaeologist — not of the Indus Valley, but of North America. Your civilization has no memory of ours. Everything we built, everything we wrote, everything we uploaded has been lost. The servers died four thousand years ago. The paper rotted. The concrete crumbled. The steel rusted into anonymous lumps of iron oxide that future geologists will catalog as “anomalous ferric deposits” and publish mildly puzzled papers about.
But you have found something. In a stratum dating to approximately 2025 CE, embedded in a layer of compacted sediment that was once a commercial district, you have recovered 1,003 small rectangular objects. They are made of a durable polymer compound that, while degraded, has retained its basic shape. Each object is roughly 85 millimeters by 54 millimeters — standardized, clearly manufactured, not natural. On one side, a series of raised characters: letters and numbers, arranged in a consistent format. On the other side, a magnetic stripe (now demagnetized) and sometimes a small metallic chip (now corroded beyond function).
You have found a cache of credit cards.
You do not know what a credit card is. You do not know what a bank is, what money is, what a transaction is. You do not speak English, or any language that existed in 2025. The magnetic stripe is dead. The chip is dead. The only information that survives is what is physically embossed on the surface: a name (which you cannot read), a sixteen-digit number (which you can count), and an expiration date (which you can partially parse, because numbers are numbers in any civilization).
What can you learn?
Quite a lot, as it turns out. You can measure the format. You can observe that the sixteen-digit number follows a consistent structure: the first four digits cluster into a small number of recurring groups (these are the bank identification numbers). The middle digits are highly variable (these are the individual account numbers). The last digit is a check digit, computed from the others. The name field is variable in content but fixed in position. The date field uses a two-digit month and a two-digit year.
You can calculate the entropy at each position and discover that it varies: low at the start (few banks, few prefixes), high in the middle (many individual accounts), low at the end (check digit, limited date range). You can compute the uniqueness rate and discover that it is extremely high — almost no two cards are identical. You can identify the numerical fields by their character set and observe that they cluster in specific positions.
You would, in short, discover exactly what we discovered in the Indus seals. Positional entropy variation. Near-total uniqueness. Embedded numerals in fixed positions. A structured, rule-governed identification system. You would not know what a credit card is. But you would know what kind of thing it is. And you would be right.
This thought experiment is not fanciful. It is the precise analogue of our situation with the Indus script. We are the archaeologists of the year 6026. The Indus seals are the credit cards. The signs are the embossed characters. The strokes are the numerical fields. The format is intact. The content is lost. But the format tells us everything we need to know about the system — its purpose, its structure, its logic — even though we cannot read a single character.
The analogy also illuminates something uncomfortable about our own civilization. We think of ourselves as permanent. We think our information is immortal because it is digital, because it is backed up, because it exists in “the cloud” — a term that should already make you nervous, since clouds are famous for disappearing. But digital information is the most fragile kind of information ever created. It requires electricity, hardware, software, and maintenance. Remove any one of these, and it ceases to exist. A hard drive without power is a paperweight. A server without cooling is a furnace. A database without a compatible operating system is a block of inscrutable magnetic dust.
The Indus people, by contrast, encoded their information in stone. Not because they lacked sophistication, but precisely because they had it. They understood, at some level that we have forgotten, that information is only as durable as the medium that carries it. They chose steatite — a material that could be carved with precision, hardened by fire, and pressed into clay to produce impressions that would outlast civilizations. They designed a format that did not depend on the survival of any particular language, religion, or government. The format was self-contained. It carried its own logic. And it worked for seven hundred years, across dozens of cities, without a single software update.
We have not built anything this durable. Our credit cards expire in three years. Our phones become obsolete in two. Our cloud storage depends on companies that may not exist in a decade. If our civilization ended tomorrow, the archaeologist of the future would find less readable information from 2025 than from 2500 BCE. The Indus people, it seems, were better at information management than we are. They just used smaller files.
There is a final twist to this thought experiment. Imagine that the archaeologist of 6026, having analyzed the credit cards and reconstructed the format, writes a paper about it. Imagine that she presents her findings at a conference. And imagine that a colleague raises his hand and says: “But what do the numbers mean? You haven’t decoded the language. This isn’t a real decipherment.”
She would reply: “I don’t need to decode the language. I’ve decoded the system. The numbers are identifiers. The format is positionally constrained. The structure is administrative. I know what kind of thing this is, even if I don’t know what any individual number refers to. The format is the finding.”
That is our reply, too.
Chapter Fifteen: The Tax Collector
Bahata Ansumali Mukhopadhyay is not a professor of archaeology. She is not a linguist, an Indologist, or a philologist. She is a software engineer from Kolkata, and she has done more to advance the practical interpretation of the Indus script than most credentialed scholars combined. This is either an inspiring testament to the power of outsider thinking or a damning indictment of academic gatekeeping, and it is probably both.
Between 2019 and 2023, Mukhopadhyay published a series of papers in Humanities and Social Sciences Communications — a Nature Portfolio journal and, as it happens, the same journal where we intend to submit our own research. Her central argument is deceptively simple: the Indus inscriptions record administrative and economic transactions. The signs on the seals encode information about taxation, trade licensing, commodity control, and access authorization. The seals are not prayers. They are not names. They are not linguistic texts in any conventional sense. They are bureaucratic instruments.
If this sounds familiar, it should. It is, in essence, the same conclusion that our statistical analysis reaches. The difference is methodological: Mukhopadhyay arrived at her interpretation through semantic analysis — studying the signs, their contexts, their associations with specific types of objects and locations — while we arrived at ours through structural statistics. Two entirely different approaches. The same destination. In science, this is called convergent evidence, and it is the strongest kind there is.
Mukhopadhyay’s first major paper, published in 2019, bore the title “Interrogating Indus Inscriptions to Unravel Their Mechanisms of Meaning Conveyance.” It argued that the inscriptions function through a combinatorial mechanism — signs combined in rule-governed ways to produce specific administrative meanings. She identified recurring patterns that suggested the signs encoded categories of goods, types of transactions, and levels of authorization. Her reading was not phonetic. She did not claim to know what language the Indus people spoke. She claimed, instead, that the system operated at a level above language — an administrative layer that could function regardless of what language its users spoke among themselves.
This is a crucial insight, and it deserves more attention than it has received. In a multilingual civilization — and there is every reason to think the Indus Valley Civilization was multilingual, given its geographic extent and its contact with diverse neighboring cultures — an administrative system that depends on a single language is a liability. A system that operates through abstract codes, independent of language, is an asset. It works for everyone, regardless of what language they speak at home. It is, in modern terms, a protocol — a shared standard that enables interoperability across different systems.
Her 2023 paper went further, proposing specific functional categories for the signs. She identified what she believed were markers for taxation, trade licensing, commodity classification, and quality control. Her analysis was meticulous, drawing on the spatial distribution of seal finds (where they were found within the city, and in what archaeological contexts), the associations between specific signs and specific types of objects, and the recurring combinations that suggested formulaic administrative content.
Not everyone was convinced. Some scholars objected that her interpretations were too specific — that the data, while suggestive, did not compel the particular functional readings she proposed. Others noted that semantic analysis, without statistical validation, is vulnerable to confirmation bias: you see what you expect to see, especially when the data is sparse and the signs are open to multiple interpretations. These are fair criticisms, and Mukhopadhyay herself has acknowledged the need for further testing.
This is where our research enters the picture. What we provide is exactly what Mukhopadhyay’s work needed: independent statistical confirmation. We did not set out to test her hypothesis. We set out to analyze the structure of the seal inscriptions without any prior interpretive commitment. But the structure we found — the positional constraints, the unique identifiers, the embedded numerals, the administrative profile — is precisely what you would expect if her hypothesis were correct.
Our paper says so explicitly. In the reconciliation section, we identify Mukhopadhyay’s work as “the closest existing interpretation to the model our data supports.” This is not a compliment paid out of politeness. It is a statement of empirical fact. Her semantic reading and our statistical analysis point to the same conclusion: the seals are administrative instruments. She told us what the system does. We showed her how it is built.
There is something satisfying about this convergence, and something important. In a field as fractious as Indus studies, where every interpretation is contested and every scholar is accused of bias, the agreement of two independent methods — one semantic, one statistical — carries special weight. It is hard to accuse both approaches of the same bias when they use entirely different data and entirely different logic. If we are both wrong, we are wrong in the same direction by coincidence, which is statistically implausible. If we are both right, we are right for different reasons, which is the strongest form of evidence there is.
Mukhopadhyay’s work also highlights something about the sociology of knowledge that is worth pausing over. She is not affiliated with a major university. She does not hold a chair in Indology. She is a software engineer who became fascinated by the Indus script and brought to it the skills she had: pattern recognition, systematic analysis, and the engineer’s habit of asking “what does this thing actually do?” rather than “what does this thing symbolically represent?” This is not the kind of question that traditional humanists are trained to ask. But it is exactly the right question for a system that, as it turns out, was designed to do something rather than to say something.
The academy has not always been kind to outsiders. The history of the Indus script is littered with amateur decipherers who were rightly dismissed for lack of rigor. But Mukhopadhyay is not an amateur. She is a professional from a different profession, applying a different toolkit to a problem that the traditional toolkit had failed to solve. The fact that she published in a Nature Portfolio journal — not once but twice — suggests that the gatekeepers, however slowly, are beginning to notice.
In the story of the Indus script, Mukhopadhyay plays a role that is both essential and underappreciated. She is the person who first said, clearly and with supporting evidence, what the system actually is: a tax collection and trade licensing apparatus. She said it before we had the statistics to back it up. She said it from outside the guild, without the credentials that the guild demands. And she was right.
The tax collector, it turns out, has been standing in the room the whole time. She was just holding a different badge than the one the committee expected.
Chapter Sixteen: Near-Duplicates and Prefix Trees
If you want to understand a system, find the places where it almost repeats itself. Perfect repetition tells you nothing — it is just a copy. Perfect randomness tells you nothing either — it is just noise. But a near-miss, a case where two things are almost identical but differ in exactly one place, is the fingerprint of a rule. It tells you what the system holds constant and what it allows to vary. It tells you where the structure lives.
In our corpus of 179 unicorn seals, we found eight pairs of inscriptions with a Hamming distance of one — meaning they are identical except for a single sign in a single position. Eight pairs, out of 15,931 possible pairings. A small number, which is itself informative: the system rarely produces near-matches, consistent with a code designed to maximize uniqueness. But the near-matches that do exist are extraordinarily revealing.
In every pair involving inscriptions of five or more signs, the variation occurs in the middle. Never at the beginning. Never at the end. Always in the interior positions — where entropy is highest, where the system generates its greatest diversity. The edges are locked. The center flexes.
If you have two license plates that differ by one character, and that character is always in the middle, you can infer structure without reading a single letter. The prefix is shared because both plates belong to the same jurisdiction. The suffix is shared because both belong to the same class. What differs is the sequential identifier — the part that makes each plate unique within its group.
The Indus near-duplicates behave the same way. First signs shared. Last signs shared. Middle signs varied. These are structural echoes: two seals from the same category, differing by one unit of individual variation.
We tested this with a binomial test. If the difference could appear in any position with equal probability, the chance of it always avoiding the edges is about 1.7 percent. The pattern is real. The edges are protected by design.
Now consider the prefix-sharing groups. We defined a prefix as the first two signs. Twenty-eight distinct prefixes appear in two or more seals. The largest groups share nine members each — nine different seals beginning with the same two signs, diverging only afterward. These are branches of a classification tree: the root defines the category, the branches define the individual.
Computer scientists will recognize this immediately. It is a trie, or prefix tree — one of the most fundamental data structures in computer science. Your phone’s autocomplete uses one. Your browser’s address bar uses one. Every search engine on the planet uses one. The Indus seal system, it appears, uses one too. Four thousand years before anyone wrote a line of code.
Get Boris (Bruce) Kriger’s stories in your inbox
Join Medium for free to get updates from this writer.
The combinatorial implications are staggering. Ten common prefixes, times thirty signs per middle position cubed, times five common suffixes: 1,350,000 possible unique identifiers. Conservative estimate. A liberal calculation yields over two billion. The system had capacity orders of magnitude beyond what the civilization needed. It was overengineered — which is exactly what you want in an identification system. You build with room to grow. You build with margins so wide that the system never runs out of numbers.
Someone designed this. Not accidentally. Not through gradual cultural drift. Deliberately. Someone decided: the first two signs mark the category, the middle signs mark the individual, the penultimate sign carries a number, the final sign closes the sequence. They designed a format, enforced it across dozens of cities, and it worked for seven hundred years.
That is not pre-literate fumbling. That is systems engineering. And we found it by counting.
Chapter Seventeen: The Silence of Things
There is a city that has no palace. No throne room, no royal tomb, no triumphal arch, no statue of a king. No monument to victory, no inscription celebrating conquest, no image of a ruler towering above his subjects. For an ancient city, this is remarkable. For one of the largest cities in the Bronze Age world, it is almost inexplicable.
Mohenjo-Daro is the most thoroughly excavated site of the Indus Valley Civilization, and what it reveals is, in many ways, more interesting for what is absent than for what is present. There are no defensive walls. No centralized military apparatus. No weapons caches, no barracks, no depictions of warfare. In a world where every other Bronze Age civilization — Egypt, Mesopotamia, the Hittites, the Shang — proclaimed its power through monumental violence, the Indus cities are silent.
What they have instead is infrastructure. Extraordinary, obsessive, magnificent infrastructure.
The street grid of Mohenjo-Daro is oriented to the cardinal directions with a precision that suggests astronomical alignment or, at the very least, a society that took geometry very seriously. The main streets are roughly ten meters wide. The brick dimensions are standardized throughout the civilization: a ratio of 1:2:4, length to width to height, repeated across sites separated by hundreds of kilometers. Whoever set the standard had authority. Whoever maintained it had reach.
The drainage system is, by common scholarly agreement, the most impressive feature of the city. Every house had a drain that led to a covered sewer running beneath the street. The sewers connected to larger mains that led outside the city. Access points were built into the walls so workers could clear blockages. This is not accidental plumbing. This is civil engineering — centrally planned, consistently maintained, and universal. Every house had drainage. Not just the rich ones. In the Europe of three thousand years later, most houses had a ditch.
The Great Bath — roughly twelve meters long, seven wide, two and a half deep, lined with bricks sealed in bitumen to prevent leakage — remains one of the most impressive public structures of the ancient world. Steps descend on two sides. Small rooms surround it, possibly for changing. Whatever it was used for — ritual purification, communal gathering, civic ceremony — it was clearly important. Important enough to build with the same care that other civilizations reserved for temples and tombs.
But there is no temple. Or rather, there is no structure that archaeologists can confidently identify as one. There are large buildings, some with columned halls, some with elevated platforms. But nothing like the ziggurats of Mesopotamia or the pyramids of Egypt — no monumental structure that screams “this is where we worship.” If the Indus people had gods, they worshipped them quietly. If they had priests, the priests did not live in palaces. If they had rituals, the rituals did not require architecture that dwarfed the houses of ordinary people.
This absence has baffled archaeologists for a century. Every other civilization of comparable size and duration built monuments to power. The Egyptians built pyramids. The Sumerians built ziggurats. The Chinese built walls. The Indus people built drains. The most impressive public works in their cities are not temples or palaces but systems of water management. This tells you everything about their priorities and nothing about their religion, which is, depending on your temperament, either frustrating or deeply admirable.
What it also tells you is something about the kind of authority that governed these cities. Monumental architecture is, everywhere else in the ancient world, a declaration of centralized power — a king saying, in stone, “I am larger than you.” The Indus cities make no such declaration. Their authority is expressed not in height but in consistency — in the standardized bricks, the uniform weights, the identical drainage plans, and, of course, the seals. Authority here is not vertical. It is horizontal. It does not tower above the population. It permeates the population, embedded in every brick, every drain, every tiny stone pressed into clay.
The seals, in this context, are not anomalies. They are the perfect expression of a civilization whose power was administrative rather than military, systematic rather than monumental, distributed rather than concentrated. The seal does not glorify a king. It identifies a participant. It does not command obedience. It registers existence. It is the smallest possible monument — the size of a postage stamp — and yet it carries more information about the system that produced it than any pyramid carries about the pharaoh who commissioned it.
There is a philosophical point here that deserves a moment of silence — which is appropriate, given the chapter’s title. We tend to measure civilizations by their loudest achievements: the tallest buildings, the largest armies, the most dramatic conquests. By this measure, the Indus Valley Civilization is a disappointment. It did not build anything tall. It did not conquer anyone. It did not produce a single known hero, general, or philosopher whose name has survived.
But if you measure a civilization by the quality of life it provided to its ordinary inhabitants, the Indus cities may be the most successful societies in the ancient world. Clean water. Effective drainage. Standardized housing. An economy that reached from Afghanistan to the Persian Gulf. And an information system that tracked participants across the entire network without requiring anyone to read or write a language.
The silence of Mohenjo-Daro is not the silence of poverty or ignorance. It is the silence of a civilization that did not need to shout.
Chapter Eighteen: The Graffiti Connection
In January 2025, at a three-day conference in Chennai marking the centenary of the discovery of the Indus Valley Civilization, Tamil Nadu Chief Minister M. K. Stalin made two announcements. The first was the prize — one million dollars for anyone who could decipher the Indus script. The second was a book, published by the Tamil Nadu State Department of Archaeology, that contained an extraordinary claim: sixty percent of graffiti marks found on ancient pottery across Tamil Nadu have morphological parallels with signs from the Indus Valley script.
The study was conducted by K. Rajan, an archaeologist at Pondicherry University, and R. Sivanantham, the deputy director of Tamil Nadu’s archaeological department. They had digitized over fifteen thousand graffiti-marked pot shards from a hundred and forty archaeological sites across Tamil Nadu and compared them, shape by shape, with the known inventory of Indus signs. The parallels, they reported, were striking: not just a handful of vague resemblances, but systematic overlap across the entire repertoire. Ninety percent of all marks shared some degree of similarity.
Rajan and Sivanantham were not the first to notice this resemblance. In 1962 — more than sixty years earlier — the renowned archaeologist B. B. Lal published a study in the Bulletin of the Archaeological Survey of India showing that eighty-nine percent of surveyed megalithic graffiti symbols had counterparts among the Indus signs. Lal’s figure is higher than Rajan and Sivanantham’s sixty percent, and his geographic scope was broader: the megalithic symbols he examined were not confined to Tamil Nadu. They appeared across the Deccan and, crucially, in ancient Lanka — a region whose languages are Indo-Aryan, not Dravidian. This is a detail that deserves emphasis. If the same marking tradition appears in both Dravidian-speaking and Indo-Aryan-speaking regions, then the tradition is not linguistic. It is administrative. It crosses language boundaries because it was never bound to any language in the first place. This is precisely what our hypothesis predicts.
The non-linguistic interpretation also finds independent support from two other scholars who deserve recognition. Steven Bonta, in a 2010 article that has received less attention than it merits, argued on structural grounds that the Indus signs do not encode normal human language — a conclusion reached independently of Farmer, Sproat, and Witzel, and by different methods. Bryan Wells, working from a different angle entirely, proposed a decipherment of the volumetric measures encoded in the Indus sign system — a reading that, if correct, directly confirms our finding that stroke signs function as numerical fields within a structured code. Bonta saw the absence of language. Wells saw the presence of measurement. Both saw pieces of the same system — a system that counts, classifies, and registers, but does not speak.
The political implications were obvious and were not lost on anyone in the room. If Tamil Nadu’s graffiti marks match Indus signs, then there is a material connection between the Indus Valley Civilization and the deep history of southern India. For Tamil politicians who have long argued that Dravidian culture is the root of Indian civilization, this was headline gold. Stalin, in his speech, made the connection explicit: the Indus civilization and Tamil civilization are linked. The bulls on the seals are the bulls of jallikattu. The script contains the ancestors of Tamil.
The science, as is often the case when politics enters the room, is more complicated than the headlines suggest.
Rajan and Sivanantham’s study is morphological, not structural. They compared the shapes of marks. They did not compare the positions those marks occupy, the sequences they form, the statistical distributions they follow, or the functional roles they play. Shape matching, while useful as a first step, has a well-known limitation: simple geometric forms — crosses, V-shapes, parallel lines, circles, zigzags — arise independently in virtually every human culture that has ever made marks on surfaces. Two cultures can produce the same shape without having any historical connection whatsoever, the same way two people can independently invent the wheel because circles are a good idea.
The question is not whether Tamil Nadu’s graffiti marks look like Indus signs. Some of them clearly do. The question is whether this resemblance reflects a real historical connection — a transmission of an administrative or symbolic tradition across a thousand years and a thousand kilometers — or whether it reflects the universal tendency of humans to draw simple shapes on pots.
Our research offers a way to test this. If the Indus sign system is a structured administrative code — with positional constraints, numeral fields, prefix categories, and a specific entropy profile — then the question becomes: do the Tamil Nadu graffiti marks exhibit the same structural properties? Not the same shapes. The same rules.
If the Tamil Nadu marks show positional constraints — if certain marks cluster at the beginning and others at the end, if there are numerical elements in fixed positions, if the sequences exhibit high uniqueness — then the resemblance is not superficial. It is structural. It would mean that the same information architecture persisted from the Indus Valley to Tamil Nadu, surviving the collapse of the cities, the dispersal of the population, and the passage of a millennium. Not linguistic continuity — administrative continuity. The same system, adapted to new materials, new locations, but retaining its fundamental logic.
This is our sixth falsifiability prediction, and it is, in our view, the most exciting. Because the data already exists. Rajan and Sivanantham have digitized fifteen thousand marks. The marks have positions, sequences, and frequencies. All that is needed is to apply the same statistical analysis we applied to the Mohenjo-Daro seals: entropy by position, uniqueness testing, numeral identification, prefix analysis. If the structural parallels hold, the connection is real. If they don’t, the resemblance is coincidental.
This is what we mean by falsifiability. We are not claiming that the connection exists. We are claiming that the connection can be tested, with existing data, using existing methods, and that the test would produce a definitive result either way. This is the difference between speculation and science. Speculation says: the marks look similar, therefore they are connected. Science says: here is a test. Run it and find out.
For Tamil Nadu, the stakes are high. If the structural test confirms the connection, it would mean that the Indus administrative tradition did not die when the cities fell. It migrated. It adapted. And it survived, in the pottery marks of Tamil Nadu, for over a thousand years after the last seal was pressed in Mohenjo-Daro. This would not prove that the Indus people spoke a Dravidian language — that is a separate question, and one our methods cannot answer. But it would prove something equally significant: that the information system they invented was robust enough to persist across the collapse of an entire civilization.
That is a stronger claim than linguistic kinship. Languages change quickly — a few centuries of isolation can make two dialects mutually unintelligible. But administrative systems, once established, have a peculiar durability. The Roman system of roads, milestones, and census records survived the fall of the empire and was adopted, with modifications, by the kingdoms that replaced it. The Chinese system of imperial examinations persisted through dynastic changes spanning two thousand years. Administrative traditions are sticky. They outlast the governments that create them, because they are useful to every government that follows.
If the Indus administrative system survived in Tamil Nadu, it would be the oldest example of this phenomenon in human history. A four-thousand-year-old bureaucratic tradition, persisting across the collapse of cities, the migration of peoples, and the transformation of cultures. Not because anyone decided to preserve it, but because it worked. Because once you have a system for marking ownership, controlling trade, and tracking goods, you do not easily abandon it. You adapt it. You simplify it. But you keep it, because the alternative — starting from scratch — is worse.
Tamil Nadu may hold the key to proving this. The data is there. The method is there. What is needed now is the will to apply the one to the other. And perhaps a million-dollar incentive doesn’t hurt.
Chapter Nineteen: Administrative Continuity
Here is a fact that should surprise no one but somehow surprises everyone: bureaucracy is the most durable technology ever invented.
Empires fall. Languages die. Religions transform beyond recognition. But the forms — the actual paperwork, the formats, the systems of registration and taxation and record-keeping — persist with a cockroach-like tenacity that would be admirable if it were not so annoying. The Roman road system survived the fall of Rome. The Chinese examination system survived dynasty after dynasty for two thousand years. The British land registry, established in 1862, is still in use. Governments come and go. Their forms are eternal.
This is not an accident. Administrative systems survive because they are useful to everyone who holds power, regardless of their ideology, religion, or ethnic identity. A new king does not burn the tax records. He reads them. A conquering army does not dismantle the census. It uses it. Administrative infrastructure is the one thing that every ruler, however revolutionary, however destructive, has an interest in preserving. You can overthrow a government in a day. It takes decades to rebuild its filing system.
This principle — call it administrative inertia, or bureaucratic persistence, or, if you are feeling cynical, the immortality of red tape — is the key to understanding the relationship between the Indus Valley Civilization and Tamil Nadu.
The conventional narrative of the Indus decline goes something like this: around 1900 BCE, the great cities began to empty. The Ghaggar-Hakra river system, which had sustained the eastern settlements, dried up. The monsoon patterns shifted. The population dispersed — some eastward into the Gangetic plain, some southward into the Deccan and beyond. The urban centers were abandoned, but the people did not vanish. They became villagers, pastoralists, migrants. They carried their skills, their crafts, their traditions, and — possibly — their administrative practices with them.
If our hypothesis is correct — if the Indus seals represent a structured registration system used for tracking goods, people, and transactions — then the system would not necessarily disappear when the cities did. It would adapt. The seals might no longer be pressed into clay at a central warehouse, but the practice of marking goods with standardized symbols could continue in a simplified form. Steatite seals might be replaced by scratches on pottery. The elaborate five-sign codes might be shortened to two or three marks. The unicorn might vanish, but the strokes might remain.
This is exactly what we see in Tamil Nadu.
Rajan and Sivanantham’s fifteen thousand graffiti marks, scratched onto pottery across a hundred and forty sites, bear morphological resemblance to Indus signs. But morphology alone proves nothing. What would prove something is structure. Do the Tamil marks follow positional rules? Do they form sequences with high uniqueness? Do they include numerical elements in fixed positions? Do they behave, statistically, like a degraded version of the Indus code?
We do not yet know. The positional analysis has not been done. This is our falsifiability test number six, and it is the one that excites us most, because the answer — whichever way it goes — will be definitive. If the Tamil marks show structural parallels to the Indus system, the connection is not cosmetic. It is functional. It means the administrative tradition survived the collapse of the cities, traveled southward with the migrating population, and persisted in simplified form for over a thousand years. Not linguistic continuity. Administrative continuity. The forms outlasted the civilization that created them.
For Tamil Nadu, this reframing is not a demotion — it is a promotion. The conventional Dravidian hypothesis says: the Indus people spoke our language. Our hypothesis says: the Indus people invented the world’s first information system, and Tamil Nadu inherited it. The first claim is about ancestry. The second is about achievement. And achievement, in the long run, is the stronger foundation for pride.
Consider the analogy with Roman administration. Nobody in modern Europe speaks Latin as a native language. But Roman roads, Roman law, Roman administrative divisions, and Roman concepts of governance pervade every European nation. The continuity is not linguistic — it is structural. The forms survived because they worked. The same may be true of the Indus-Tamil connection. The signs may have changed. The material may have shifted from steatite to pottery. The format may have simplified. But the underlying logic — mark your goods, register your transactions, use standardized symbols in consistent positions — may have persisted for a millennium, carried not by texts but by practice.
If this is confirmed, it would be the oldest known case of administrative continuity in human history. A system designed in Mohenjo-Daro around 2500 BCE, still traceable in Tamil Nadu around 500 BCE. Two thousand years of quiet persistence, through urban collapse, population migration, and cultural transformation. Not because anyone decreed it. Not because anyone wrote it down. But because it worked. Because once you have a system for marking things, you keep it. Because bureaucracy, for all its sins, is the most successful survival strategy civilization has ever produced.
The test is waiting. The data exists. The method exists. All that remains is to run it. And when someone does, the silence of the graffiti marks may finally speak — not in a language, but in the unmistakable voice of structure.
Chapter Twenty: The Scent of the Seal
We have talked about what the seals look like. We have talked about what they encode. We have talked about what survives and what does not. But there is one dimension of the Indus seals that nobody talks about, because it is invisible, ephemeral, and — until very recently — unmeasurable. The seals had a smell.
This is not poetry. This is chemistry. A steatite seal, when pressed into a soft surface to produce an impression, would have been pressed into something. Clay is the obvious candidate, and we have the impressions to prove it. But clay is not the only option. Wax, resin, oil-soaked fabric, and various organic compounds were all available to the Indus people and all used for sealing purposes in contemporary Mesopotamian practice. Each of these materials has a distinct chemical signature, and each, when analyzed with modern techniques, can tell us something about how the seals were used.
The field that studies these questions is called sensory archaeology, and it is one of the most under-explored frontiers in the study of ancient civilizations. For most of history, archaeology has been a discipline of sight: we look at things, measure things, draw things, photograph things. But ancient life was not only visual. It was olfactory, tactile, gustatory, auditory. People did not just see their world. They smelled it, touched it, tasted it. And some of those sensory experiences left traces that modern technology can detect.
Mass spectrometry and gas chromatography — techniques that can identify individual chemical compounds in vanishingly small quantities — have already revolutionized our understanding of ancient trade and ritual. Analysis of Egyptian funerary vessels has identified the components of kyphi, the sacred incense blend used in tomb purification: frankincense, myrrh, honey, wine, and aromatic resins, mixed in specific proportions that the priests guarded as closely as any sacred text. Residues from Roman amphorae have revealed traces of myrrh and olive oil, illuminating the cargo of ships that sank two thousand years ago.
The Indus seals have not yet been subjected to this kind of analysis. This is a gap — not a catastrophic one, but a tantalizing one. If residue analysis were performed on the surfaces of well-preserved seals, it might answer questions that epigraphy cannot. What were the seals pressed into? If the answer is beeswax, that suggests a sealing practice similar to later Indian and Mediterranean traditions. If the answer is vegetable oil or sesame paste, it suggests a different kind of impression — perhaps one applied to cloth bundles or containers rather than clay tablets.
Even more intriguing: the seals may carry traces of the goods they were associated with. If a seal was consistently stored with a particular commodity — spices, textiles, oils — microscopic residues of that commodity might survive in the pores of the steatite. A seal found in a warehouse context might carry traces of the goods that passed through that warehouse. A seal found in a burial might carry traces of the funerary offerings. Each seal, in this reading, is not just a visual record but a chemical one — a tiny time capsule of scent and substance.
The Indus Valley Civilization was located at the crossroads of some of the ancient world’s most important trade routes. Carnelian beads from Gujarat, lapis lazuli from Afghanistan, shell from the Arabian Sea, timber from the Himalayas, and — almost certainly — spices, oils, and aromatics from across the subcontinent and beyond all flowed through Indus cities. The seals, as administrative instruments, would have been associated with these goods. They were the paperwork of trade. And paperwork, as anyone who has ever opened an old filing cabinet knows, absorbs the smells of its environment.
Treasure A. Hunt, our co-author, has proposed that residue analysis of Indus seals should be a priority for future research, and we agree. The technology exists. The seals exist. What is needed is the institutional will to bring them together. A single well-preserved seal, analyzed with mass spectrometry, could tell us more about Indus trade practices than a hundred epigraphic analyses. It could tell us what was traded, how it was stored, and — perhaps — what the Indus marketplace smelled like on a warm afternoon four thousand years ago.
We will never hear the Indus people speak. We may never read their signs. But we might, one day, catch a faint trace of their world on the surface of a stone — a whiff of sesame, a ghost of sandalwood, the chemical memory of a transaction completed in the shade of a city that fell silent before Rome was founded.
The past is not only visible. It is volatile. And sometimes, the most important evidence is the kind that evaporates.
Chapter Twenty-One: The Compression
Treasure A. Hunt is not easy to categorize, which is appropriate for someone whose central intellectual contribution is a framework for understanding how systems resist categorization.
She is an information theorist. That much is clear from her published work. But “information theorist” is one of those job descriptions that tells you the genus without revealing the species. Information theory, in its broadest sense, is the study of how data is encoded, transmitted, compressed, and preserved. Claude Shannon invented it. The internet runs on it. Every file you have ever downloaded, every phone call you have ever made, every photograph you have ever stored on a device has been shaped by its principles. But Hunt does not work on telecommunications. She works on something considerably stranger: the question of how structure itself persists.
Her framework, which she calls Living Information Theory, begins with an observation so simple that it is easy to overlook: information does not exist in isolation. It exists in constraints. The meaning of a signal is not in the signal itself but in the rules that govern what the signal can and cannot be. A traffic light communicates not because red has some inherent meaning, but because the system constrains red to mean “stop” and green to mean “go.” Remove the constraints, and the signal is just colored light. The information lives in the architecture, not in the content.
This sounds abstract, and it is. But it becomes very concrete very quickly when you apply it to the Indus seals. For a century, scholars tried to extract meaning from the content of the signs — from what each sign depicts, from its phonetic value, from its semantic associations. All of these approaches assume that meaning resides in the signs themselves. Hunt’s framework says: look at the constraints instead. What positions can a sign occupy? What combinations are permitted? What is excluded? The answers to these questions reveal the system’s architecture, and the architecture reveals the system’s function.
When we applied this principle to the Mohenjo-Daro corpus, the results were immediate and unambiguous. The signs are positionally constrained. Certain signs appear only at the beginning. Others appear only at the end. The middle positions are diverse but still governed by transition rules. Numerals cluster in a fixed position. The system, viewed through the lens of constraint analysis, reveals itself as a structured code — not because of what the signs mean, but because of where they are allowed to go.
Hunt’s insight is not that constraints exist. Any linguist knows that language has constraints; any computer scientist knows that codes have rules. Her insight is that constraints are the primary carrier of information in systems where content has been lost. When you cannot read the words, read the grammar. When the grammar is gone, read the format. When the format is all that survives, the format is the finding.
That phrase — the format is the finding — originated in our collaboration, and it captures the essence of what Hunt’s framework contributes to this research. She provided the analytical lens. We provided the data. The lens focused the data into a picture that neither of us could have produced alone.
Hunt’s background is eclectic in the best sense of the word. Her published work ranges from information compression to adaptive systems to, remarkably, the behavioral patterns of cats — which she analyzes as constraint-governed systems in their own right. If this sounds whimsical, it is not. A cat’s behavior is highly structured: it follows rules of territory, timing, and social hierarchy that can be modeled as information constraints. The analogy to the Indus seals is not as absurd as it first appears. In both cases, you are looking at a system whose rules are invisible but whose outputs are consistent. In both cases, meaning is carried not by content but by pattern. In both cases, the observer’s task is to infer the architecture from the behavior.
She developed her constraint-governed approach independently of Indus studies. When she encountered the Indus script, she recognized it immediately as a system that fit her framework: a set of outputs (the seal inscriptions) governed by invisible rules (the positional constraints) that could be extracted through statistical analysis without any knowledge of the content. She published two position papers on the subject — one arguing for a non-phonetic approach, one proposing a full reconstruction of the Indus system as a structured registration code. Asko Parpola, who has spent sixty years pursuing the opposite approach, read her work and called it “a welcome exception.”
In our collaboration, Hunt’s role was conceptual and interpretive. She designed the analytical framework. She insisted on the separation between empirical results and theoretical interpretation — a discipline that makes the paper stronger, because the statistics stand on their own even if the framework is questioned. She contributed the idea that the Indus system should be analyzed as a constraint-governed architecture rather than as a degraded writing system. And she reviewed every claim in the paper with the skepticism of someone who knows that the worst enemy of a good finding is an overenthusiastic interpretation.
The Author Contributions section of our paper states it plainly: Hunt developed the framework. Kriger ran the numbers. Both interpreted the results. The empirical findings are independent of the theoretical framework and stand on their own statistical merits. This separation is not a formality. It is a guarantee. If the framework turns out to be wrong, the numbers are still right. If the numbers turn out to be insufficient, the framework still offers a method. The two components are designed to survive independently, even though they work best together.
This is, itself, a lesson in information architecture. A well-designed system does not collapse when one component fails. It degrades gracefully. The Indus seal system degraded gracefully over a thousand years, shedding its cities and its language but preserving its structure in the graffiti marks of Tamil Nadu. Our paper is designed with the same principle: lose the framework, and the data survives. Lose the data, and the framework survives. The format is the finding, and the finding is built to last.
Hunt’s contribution, in the end, is not a set of answers. It is a way of asking questions. And the question she taught us to ask — not “what does this mean?” but “what are the rules?” — turned out to be the question that the Indus seals had been waiting a hundred years to answer.
Chapter Twenty-Two: The Machine
On a Tuesday afternoon in March 2026, in a city where the snow was doing its annual impression of a civilization in decline — melting slowly, refusing to leave, and getting uglier by the day — a systems theorist in Toronto opened a chat window and typed a question to an artificial intelligence.
The question was not dramatic. There was no trembling hand, no dimly lit laboratory, no orchestra swelling in the background. The question was something like: “Can you download the Indus Valley script corpus from GitHub and count the signs?” The AI said yes. And then, over the next three hours, something happened that had not happened in a hundred years of Indus script research: the data was analyzed, comprehensively and correctly, by an entity that had no ego, no theory to defend, no career to protect, and no opinion about whether the Indus people spoke Dravidian.
The AI was Claude Opus 4.6, built by Anthropic. I should declare, at this point, that I did not choose Claude because of a careful comparative evaluation of available AI systems. I chose it because it was the tool I had, and because when I asked it to do something, it did it without arguing. This is a quality I value in research partners and find vanishingly rare in human ones.
What Claude did, in those three hours, was straightforward in concept and extraordinary in execution. It retrieved the corpus from GitHub. It parsed the JSON files. It extracted every sign from every seal, with position indices. It counted tokens, types, and frequencies. It calculated Shannon entropy at each position for five-sign inscriptions. It ran a permutation test with ten thousand iterations and a fixed random seed. It computed bootstrap confidence intervals. It identified stroke signs, mapped their positional distribution, calculated chi-squared statistics for digit frequency and numeral positioning, found near-duplicate pairs by Hamming distance, built prefix-sharing groups, and estimated combinatorial capacity. It did all of this using nothing but Python’s standard library — no NumPy, no pandas, no SciPy, no external dependencies whatsoever.
I want to dwell on that last point, because it matters. The choice to use only the standard library was deliberate. It means that anyone, anywhere, with a basic Python installation, can reproduce every result. There is no exotic software to install, no proprietary package to license, no version conflict to debug. The code runs on a Chromebook. It runs on a ten-year-old laptop. It runs in a university computer lab in Chennai or a café in Helsinki. Reproducibility is not a feature of our research. It is the architecture of our research. And that architecture was built by an AI that understood, without being told, that the simplest solution is always the most trustworthy.
Now, let me be clear about what Claude did and did not do. Claude did not have the idea to study the Indus seals. I did. Claude did not develop the constraint-governed framework. Hunt did. Claude did not interpret the results, draw historical conclusions, or write the sentences you are reading now. Those are human responsibilities, and they remain human responsibilities. What Claude did was compute. It took instructions, wrote code, executed it, and returned numbers. It did this faster, more accurately, and more reproducibly than any human could have done it, and it did it without a single complaint about the working hours.
But there is a subtlety here that deserves honesty. Claude did more than just crunch numbers. In the process of computation, it made choices — small, technical choices about how to implement a test, how to handle edge cases, how to structure a loop. These choices were not random. They reflected a vast training on scientific literature, statistical methodology, and programming practice. When Claude chose to use a permutation test rather than a parametric test, that choice was informed by an implicit understanding of when parametric assumptions are appropriate and when they are not. When it set the random seed to 42 — a number beloved by programmers for reasons that would require an entire chapter on Douglas Adams to explain — it was making a reproducibility decision that many human researchers forget to make.
Is this “thinking”? I have no idea. I am a systems theorist, not a philosopher of mind, and the question of whether artificial intelligence truly thinks or merely simulates thinking is one that I am happy to leave to people with more patience and fewer deadlines. What I can say is that the output was indistinguishable from the output of a very competent, very fast, very tireless research assistant who happened to have read every statistics textbook ever written and remembered all of them.
The collaboration between human and AI in this research was not a gimmick. It was not a marketing angle. It was a methodological necessity. The Indus corpus is small enough that a human could, in theory, perform the same analyses by hand. But “in theory” is the phrase that separates things that could happen from things that actually do. In practice, a human performing ten thousand permutation iterations by hand would die of old age before completing the first thousand. In practice, a human computing bootstrap confidence intervals for five positions across ten thousand resamples would make errors. In practice, the speed and precision of AI computation allowed us to test hypotheses that would otherwise have remained untested — not because they were untestable, but because testing them was too tedious for a human lifetime.
This is the real contribution of AI to archaeology, and to the humanities more broadly. It is not that AI has ideas. It does not. It is not that AI understands ancient civilizations. It does not. It is that AI removes the bottleneck of computation, freeing human researchers to focus on what humans do best: asking questions, interpreting answers, and arguing about what it all means over drinks at a conference.
In the paper, we are transparent about Claude’s role. The Methods section describes exactly what the AI did. The AI Usage statement declares it as a computational tool. The Author Contributions section assigns all hypotheses, interpretations, and claims to the human authors. This is not false modesty. It is accuracy. Claude did not discover that the Indus seals are registration codes. We did, using data that Claude processed. The discovery belongs to the humans. The computation belongs to the machine. And the reproducibility belongs to everyone.
There is one more thing worth saying about this collaboration, and it is the thing that nobody expected. Working with an AI on a research project is, in a strange way, clarifying. When you explain your question to an AI, you must be precise. You cannot wave your hands. You cannot say “you know what I mean.” You must say exactly what you mean, in terms that leave no room for ambiguity. This discipline — the discipline of precision, of operational definition, of saying what you actually want rather than what you vaguely intend — is the same discipline that makes good science. The AI, by demanding clarity, made us clearer. By requiring explicit instructions, it forced us to make our assumptions explicit. By doing exactly what we asked and nothing more, it showed us what we were really asking.
In the end, the machine did not replace the human. It refined the human. And the Indus seals, which had waited four thousand years for someone to ask the right question, did not care whether the question came from flesh or silicon. They answered it anyway.
Chapter Twenty-Three: The Paper
Writing a research paper is an act of controlled self-delusion. You begin with the conviction that you have discovered something important. You end, if you are honest, with the suspicion that you have merely rearranged the obvious. In between, there is formatting.
The paper at the heart of this book went through five versions, three rounds of internal review, and more arguments about semicolons than any human relationship should have to endure. It is now twenty-five pages long in its author version and nineteen in its anonymous submission version, which tells you something about how much space is consumed by the mere act of identifying who wrote what and where they work. The anonymous version is shorter not because it contains less science, but because it contains less ego.
We wrote it for Humanities and Social Sciences Communications, a journal in the Nature Portfolio. This choice was strategic in ways that are worth explaining, because the sociology of academic publishing is at least as interesting as the science it publishes, and considerably more absurd.
HSSC is an interdisciplinary journal. It publishes work that crosses boundaries — archaeology and statistics, linguistics and computation, social theory and hard data. This is rare. Most journals live inside a single discipline and enforce its norms with the zeal of a homeowners’ association: your lawn must be this green, your methodology must be this shape, and if your references do not include Professor So-and-So from our editorial board, your paper will be returned with regrets. HSSC, to its credit, is more flexible. It has published Mukhopadhyay’s administrative interpretation of the Indus script — twice — which means the journal has already accepted the premise that non-linguistic approaches to the Indus signs are publishable. We are not asking them to break new ground. We are asking them to continue a conversation they started.
The paper is structured in the way that academic papers are structured, which is to say: in a way that no human being would voluntarily organize information if they were not being held at gunpoint by a style guide. Introduction, Background, Methods, Conceptual Framework, Results, Reconciliation, Implications, Discussion, Limitations, Future Directions, Data Availability, Acknowledgements, Author Contributions, AI Usage, Competing Interests, Ethics, References. Seventeen sections for twenty-five pages. This is not writing. This is architecture — and, like most architecture, the beauty is in the engineering, not the façade.
The most important section, from a scientific standpoint, is Results. It contains every number cited in this book: the 98.3 percent uniqueness, the entropy profile, the numeral statistics, the near-duplicates, the prefix groups, the combinatorial estimates. Each result is accompanied by a formal test: permutation, chi-squared, Friedman, Kendall’s W, binomial, bootstrap. Each test has a p-value. Each p-value has a confidence interval. This is belt-and-suspenders statistics — more than a humanities journal typically demands, but exactly what we need to preempt the criticism that our claims rest on anecdote rather than evidence.
The most controversial section is probably the title. “The world’s first known information civilisation” is a large claim, and our co-author has rightly pointed out that it extends beyond what 179 seals from one site can definitively prove. She is correct. The claim is conditional — it depends on replication across the full corpus, which contains over five thousand inscriptions from dozens of sites. We flagged this explicitly in the Discussion: the designation is a hypothesis, not a conclusion. It is a prediction that the full corpus will confirm, and if it does not, the designation falls. This is how science is supposed to work: you make a testable claim, you specify the conditions under which it would be falsified, and you invite others to try. Six falsifiable predictions are listed. The invitation is open.
The section that gave us the most trouble was the Conceptual Framework — Section 4 — which presents Hunt’s constraint-governed approach. The challenge was to integrate her methodology without making the paper dependent on it. The empirical results — the entropy profile, the uniqueness rate, the numeral statistics — are raw data. They do not need a theoretical framework to be true. But they do need a framework to be meaningful. Numbers without context are just numbers. The framework provides the context: it says, “Here is why these particular numbers matter, and here is what they tell us about the kind of system we are looking at.”
The solution was to build a firewall. The Results section contains only empirical findings and standard statistical tests. The Conceptual Framework section explains the analytical lens. The Discussion section brings them together. A reviewer who disagrees with the framework can still accept the numbers. A reviewer who questions the numbers can still engage with the framework. Neither depends on the other for validity. Both are strengthened by the other’s presence. This is, as Hunt herself observed, a lesson in information architecture: build components that survive independently but perform better together.
The paper also includes something unusual for a humanities submission: a complete Python script, provided as Supplementary Information, that reproduces every statistical result in the paper. This is standard practice in computational science but almost unheard of in the humanities. We included it for one reason: trust. In a field where claims are routinely contested on ideological grounds, reproducibility is the only defense that cannot be dismissed. You can argue with an interpretation. You cannot argue with a number that anyone can verify in thirty seconds by typing “python3 supplementary_analysis.py” into a terminal.
The paper is now a preprint on Zenodo, with a DOI, a timestamp, and an open invitation to the world to verify, challenge, or extend our results. It will be submitted to HSSC after we have collected responses from the researchers we cite — responses that will inform the final version. This is not the usual sequence. Usually, you submit first and hear from the field later. We chose to reverse the order, because the field’s input will make the paper better, and because — in a field this fractious — it is better to invite collaboration before publication than to defend against criticism after it.
Writing the paper took months. Writing this book took weeks. The paper is the skeleton. The book is the body. And the data — the signs, the positions, the strokes, the patterns — is the heartbeat that drives them both.
Chapter Twenty-Four: The First Information Civilisation
Every civilization that has ever existed has managed information. You cannot run a city without it. You cannot coordinate a harvest, organize a trade network, levy a tax, or maintain a priesthood without some method of tracking who does what, who has what, and who owes what to whom. Information management is not a modern invention. It is as old as the first settlement where more than a hundred people tried to live together without killing each other, which is to say, it is as old as civilization itself.
But there is a difference between managing information and building an information system. The distinction is the same as between cooking food and inventing a recipe. One is an act. The other is a structure. One happens in the moment. The other persists beyond it. The Egyptians managed information — they kept records, wrote letters, inscribed decrees. But their information management was conducted in the medium of language. The hieroglyphs encode Egyptian speech. The records are sentences. The system is linguistic.
The Mesopotamians did the same. Cuneiform tablets from Ur and Uruk contain everything from beer rations to astronomical observations to epic poetry. They are linguistic records written in a linguistic medium. The information is real, but it passes through language on its way to the tablet. You must speak Sumerian — or at least read it — to access the data.
The Indus system is different. It does not encode language. It encodes structure. The information on an Indus seal is not a sentence, a phrase, or even a word. It is a code — a positionally constrained, combinatorially generated, statistically unique identifier that functions independently of any spoken language. You do not need to speak the Indus people’s language to extract information from their seals. You need to understand the format. And the format is readable today, four thousand years later, by anyone with a laptop and a copy of Python.
This is not a trivial distinction. It is, we argue, the defining characteristic of the Indus Valley Civilization and the reason we propose — conditionally, pending full-corpus replication — that it deserves to be called the world’s first information civilisation.
Six criteria support this designation, and no other Bronze Age civilization meets all of them simultaneously.
First: non-linguistic encoding. The Indus sign system encodes information without using language as an intermediary. Egypt and Mesopotamia wrote language. The Indus encoded information directly. This is a qualitative difference in the architecture of information management.
Second: positional structure. The signs follow strict positional rules, with different positions serving different functions — categories at the start, identifiers in the middle, numerals near the end. This is the architecture of a designed code, not the grammar of a spoken language.
Third: near-total uniqueness. Ninety-eight point three percent of inscriptions are unique. The system was built to generate difference, not repetition. This is the hallmark of an identification system, not a writing system.
Fourth: embedded numerals. Stroke signs function as numerical fields within the code, occupying fixed positions and encoding quantities or classes. This embeds quantitative data within a qualitative framework — a feature characteristic of modern registration systems.
Fifth: combinatorial capacity. The system can generate over a million unique identifiers from fewer than two hundred sign types. It was designed with capacity far exceeding its immediate needs — the mark of forward-looking engineering.
Sixth: standardization across sites. The same format, the same sign types, the same positional constraints appear across Mohenjo-Daro, Harappa, and other sites separated by hundreds of kilometers. The system was not local. It was civilizational. It was a shared protocol, enforced across a vast territory.
Egypt had writing. Mesopotamia had writing. China had writing. The Indus had something else: a structured information system that operated at the civilizational level without requiring literacy in any language. This is not a lesser achievement than writing. It is a different achievement — and, for certain purposes, a more efficient one. Writing requires that every user share a language. A structured code requires only that every user share a format. Format is cheaper. Format is faster. Format crosses language barriers. And format, as the Indus seals demonstrate, lasts longer.
We are not claiming that the Indus people did not have language. Obviously they did. We are not claiming they did not have oral literature, philosophy, religion, or myth. They almost certainly did. What we are claiming is that their administrative information system — the system encoded on the seals — was not linguistic. It was structural. And the structural approach to information management that they pioneered is, as far as we know, the earliest example of its kind in human history.
The word “civilisation” is derived from the Latin “civis” — citizen. A civilization is, at its root, a system for organizing citizens. The Indus seals did exactly that. They organized participants in an economic and administrative network through structured identifiers. They were the citizenship cards of the Bronze Age — not metaphorically, but functionally. Each seal identified its bearer. Each impression registered a transaction. Each code linked a person to a system larger than any individual city.
If that is not an information civilisation, the term has no meaning.
The designation is conditional. It depends on full-corpus replication. It depends on the structural analysis extending from Mohenjo-Daro to Harappa, Dholavira, Lothal, and beyond. It depends on the positional constraints holding across animal types, not just unicorns. These are testable predictions, and we have listed them explicitly. If the full corpus does not confirm the pattern, the designation falls. If it does, the Indus Valley Civilization becomes not only one of the great Bronze Age societies, but the one that invented something the others did not: administration without alphabets, information without language, structure without speech.
The world’s first information civilisation. Four thousand years ahead of schedule. And still readable today.
Chapter Twenty-Five: What We Do Not Know
It is fashionable, in popular science writing, to end on a note of triumph. The mystery is solved. The code is cracked. The hero walks into the sunset, metaphorically speaking, while the orchestra plays something stirring and the reader closes the book feeling that the universe is a little less mysterious than it was three hundred pages ago.
This is not that kind of book. And the Indus script is not that kind of mystery.
We have established, with statistical rigor, what the Indus seal system is: a structured, positionally constrained, combinatorially powerful identification code. We know its architecture. We know its entropy profile. We know where the numerals sit and how the prefixes branch. We know that it generates uniqueness by design and that its format was enforced across a civilization the size of Western Europe. These are real findings, supported by reproducible data and formal statistical tests.
But there is a great deal we do not know, and honesty requires that we say so clearly, without hedging and without the academic habit of burying limitations in a footnote on page forty-seven where nobody will read them.
We do not know what language the Indus people spoke. Our analysis is entirely structural. It identifies the format of the system without touching its content. The signs could encode Proto-Dravidian, early Indo-Aryan, a language isolate, or something entirely unknown. Our method cannot distinguish among these possibilities, because the method operates at the level of position and frequency, not phonetics or semantics. If someone tomorrow discovers a bilingual inscription — an Indus seal with a Sumerian translation on the back — our structural findings would remain valid, but the content question would finally be answered. Until then, the language remains unknown.
We do not know what individual signs mean. We know that sign P013 frequently appears in position one, that it behaves like a category marker, and that it co-occurs with specific prefix partners. But we do not know whether P013 represents a profession, a clan, a geographic origin, a commodity type, or something else entirely. We have identified the sign’s role in the system. We have not identified its referent in the world. Role and referent are different things, and conflating them would be a mistake we are determined not to make.
We do not know who carried the seals. Were they merchants? Officials? Craft specialists? Heads of household? Every adult, or only certain adults? The archaeological context suggests a connection to trade and administration, but the details are murky. Seals have been found in residential areas, in workshops, in what appear to be warehouses, and in trash deposits. They have been found at Mohenjo-Daro, at Harappa, and at sites in Mesopotamia and the Persian Gulf — which tells us the seals traveled, but not who carried them or why.
We do not know what the digit two means. We know it dominates the numeral field — sixty-seven percent of all stroke signs are twos. We have proposed three hypotheses: a quantity of goods, a tier in a classification, or a weight standard. We cannot yet distinguish among them. The answer may come from cross-referencing numeral distributions with archaeological context — if seals with “two” are found in specific locations or associated with specific goods, the correlation could resolve the question. But this requires data we do not have.
We do not know whether the pattern holds across the full corpus. Our analysis is based on 179 unicorn seals from Mohenjo-Daro. The full corpus contains over five thousand inscriptions from dozens of sites, across multiple animal types and object categories. The positional constraints we found may be universal, or they may be specific to unicorn seals at Mohenjo-Daro. The only way to find out is to run the analysis on the complete dataset, which requires access to the Interactive Corpus of Indus Texts — a database that is not currently open-access.
We do not know whether the Tamil Nadu graffiti marks share the same structural properties as the Indus signs. The morphological parallels reported by Rajan and Sivanantham are suggestive but not diagnostic. Our proposed test — positional entropy analysis of the Tamil Nadu corpus — has not yet been performed. Until it is, the hypothesis of administrative continuity remains a hypothesis.
These are not minor gaps. They are the central questions of the field, and they remain open. What we have contributed is not a solution but a framework — a set of tools, a body of evidence, and a structural model that narrows the space of possible answers without closing it. We have eliminated certain categories of interpretation (the seals are not prayers, not clan symbols, not random). We have confirmed others (they are administrative, they are structured, they are unique). And we have proposed six specific, falsifiable predictions that future research can test.
Those predictions, stated plainly, are:
One: the positional entropy profile will replicate across the full ICIT corpus. Two: cross-animal analysis will reveal that different animal types correspond to different prefix distributions. Three: the penultimate numeral pattern will hold across inscription lengths and sites. Four: the digit-two dominance will appear in other subcorpora, not just Mohenjo-Daro unicorn seals. Five: near-duplicate variation will consistently avoid edge positions across the full corpus. Six: positional analysis of Tamil Nadu graffiti marks will reveal structural parallels with the Indus seal format.
If all six predictions are confirmed, the case for an Indus information civilisation becomes overwhelming. If some fail, the model must be revised. If all fail, the model is wrong. This is how science works — not through certainty, but through the willingness to be wrong in specific, measurable, publicly verifiable ways.
We have placed our bets on the table. The cards are face up. The data is on GitHub. The code is in the appendix. If we are wrong, you can prove it before dinner. That is not arrogance. That is the only kind of confidence worth having: the kind that comes with an invitation to be disproved.
Chapter Twenty-Six: The Prize
In January 2025, at a conference in Chennai celebrating the centenary of the Indus Valley Civilization’s discovery, Chief Minister M. K. Stalin of Tamil Nadu stepped to the podium and announced a prize of one million US dollars for the decipherment of the Indus script. The prize was named after Iravatham Mahadevan — the civil servant turned epigraphist whose concordance, as you now know, is the foundation upon which everything in this book rests.
The announcement made headlines around the world. A million dollars is a lot of money, and the Indus script is a lot of mystery, and the combination of the two produced exactly the kind of story that news editors love: ancient puzzle, big prize, exotic location, anyone can enter. It was framed as a challenge — a call to scholars, amateurs, and dreamers everywhere to crack the code that had resisted a century of effort.
What the headlines did not mention is that there is no official submission portal. No committee website, no application form, no deadline. Treasure Hunt, our co-author, who submitted her own work for the prize in 2025, spent a week tracking down the people to send it to. This suggests that the Government of Tamil Nadu did not entirely expect anyone to actually submit a claim. The prize, like many government announcements, may have been designed more for the ceremony than for the outcome.
But outcomes have a way of arriving uninvited.
Our work is, by any reasonable interpretation, relevant to the prize. We have not deciphered the Indus script in the traditional sense — we have not assigned phonetic values to individual signs or read a single inscription as a sentence. But we have achieved something that may, in the long run, prove more valuable: we have identified the structural architecture of the system. We have shown what kind of thing the Indus script is. And we have done so with reproducible data, open code, and formal statistical tests that any evaluator can verify.
Whether this qualifies as “decipherment to the satisfaction of archaeological experts” — the language of the prize announcement — is a question that the prize committee will have to answer. Our view is that structural decipherment is decipherment. It is the identification of the system’s function, format, and internal logic. It does not tell you what the signs say, but it tells you what the signs do. And in a system where the format is the primary carrier of information, understanding the format is understanding the system.
But the prize raises a deeper question, one that goes beyond any single paper or any single team. The structural decipherment presented in this book is not the work of two people. It is the accumulated product of a century of scholarship by dozens of researchers across five continents. Mahadevan built the corpus. Parpola identified the semantic depth. Farmer, Sproat, and Witzel cleared the non-linguistic ground. Rao and Yadav proved computational structure. Mukhopadhyay identified the administrative function. Rajan and Sivanantham documented the Tamil Nadu parallels. Hunt developed the analytical framework. We ran the statistics. No single contribution, including ours, would have been possible without the others.
This is why we have proposed — outside the paper itself, as a post-publication initiative — that any recognition for this work be shared. Not as a gesture of generosity, but as a statement of fact. The Indus script was not cracked by a lone genius. It was cracked by a field — slowly, painfully, through decades of competing approaches that turned out to be complementary rather than contradictory. If the prize is meant to reward decipherment, it should reward the process that produced it. And that process involved everyone.
We have identified seven groups whose contributions were essential: Kriger and Hunt, Mahadevan (represented by his heirs), Parpola, Farmer-Sproat-Witzel, Rao and Yadav, Mukhopadhyay, and Rajan and Sivanantham. We have proposed that each group direct its share toward continuation of Indus script research — so that the prize functions not as a terminal reward but as a catalyst for the next phase of investigation. The full-corpus replication, the cross-site analysis, the Tamil Nadu positional study — all of these require funding. A million dollars, divided seven ways and directed toward research, would fund years of work.
There is a precedent for this kind of thinking, though not in archaeology. The Human Genome Project, completed in 2003, was not the achievement of a single laboratory. It was a coordinated international effort involving hundreds of researchers across dozens of institutions. When the sequence was published, no single author was credited. The achievement belonged to the collaboration. The Indus script, we suggest, deserves the same treatment.
Mahadevan’s heirs — his son Sridhar, his grandchildren Vandana and Vinay — would receive a share in the name of the man whose concordance made everything possible. The prize is named after him. It seems only right that his family share in its outcome. He died in 2018, seven years before this work was completed. Every number in our paper comes from data he compiled by hand, over decades, with patience that bordered on the superhuman. If the concordance is the foundation, then Mahadevan is the architect, and the architect’s family should not be forgotten when the building is celebrated.
We have also written to Anthropic, the company that built Claude, proposing that they match the Tamil Nadu prize — turning it into a two-million-dollar global research fund for Indus script studies. For Anthropic, the story is irresistible: their AI helped unlock the structure of the world’s oldest information system. For the Tamil Nadu government, the partnership elevates the prize from a state initiative to an international collaboration. For the field, it means real money for real research — not a one-time payout, but a sustained investment in answering the questions that remain open.
Whether any of this happens is, at this writing, uncertain. Prizes are political. Governments are slow. Corporations are cautious. But the proposal is on the table, and the logic behind it is simple: the Indus script belongs to everyone. Its decipherment should benefit everyone. And the researchers who made it possible — living and dead, famous and obscure, inside the academy and outside it — deserve to share in whatever recognition follows.
The million-dollar meaning of the Indus script, as Nisha Yadav put it, may remain elusive. But the million-dollar prize, if wisely distributed, could ensure that the search continues — with better tools, better data, and the knowledge that the format has already been found. What remains is the content. And content, unlike format, may require not just statistics but luck: a bilingual inscription, a long text, a Rosetta Stone buried somewhere in the silt of a river that dried up four thousand years ago.
Until then, we have the structure. We have the numbers. We have the code. And we have a prize named after a man who spent his life counting signs on stones, and whose count, in the end, was the one that mattered most.
Chapter Twenty-Seven: Eighty Handshakes Back
It is early morning in Mohenjo-Daro. The year is approximately 2400 BCE, give or take a century, because nobody here has invented the calendar you use and nobody will for a very long time. The sun is rising over the eastern wall of the city, which is not really a wall in the defensive sense — nobody attacks this city, because this city does not have enemies in the way that cities usually have enemies — but rather a boundary, a line that says: inside is order, outside is everything else.
A man walks through the streets. He is not a king, because this city does not have kings, or if it does, they do not advertise. He is not a priest, though he may pray. He is not a soldier, because this city does not have soldiers, or if it does, they are not the kind who march in formation and carry spears. He is a person with a job — perhaps a trader, perhaps a craftsman, perhaps an official of some kind. His name is lost. His language is lost. His religion is lost. Everything about him that was not carved in stone is gone.
But he is carrying a seal.
It is small — barely the size of his thumb. It is made of steatite, heated and hardened, carved with a precision that would impress a modern jeweler. On one face, a unicorn — or rather, a bull in profile, its two horns overlapping into one, its musculature rendered with the care of someone who has spent a long time looking at animals and an even longer time learning to carve. Beneath the animal, a strange object on a pedestal. Above the animal, five signs, incised in a line that reads from right to left.
The first sign is one he shares with others in his group — his profession, his district, his guild, whatever the first position encodes. The second and third signs are his alone, or nearly so — the part of the code that makes this seal different from every other seal in the city. The fourth sign is two vertical strokes: the number two. The fifth sign closes the sequence — a terminal marker, a suffix, a period at the end of a sentence that is not a sentence.
He does not think of it as a code. He thinks of it as his seal. It is the thing that identifies him, that authorizes his transactions, that links his name — whatever his name is — to the system that runs this city and every city in the network. When he presses it into a lump of clay affixed to a bundle of goods, the impression says: this is mine. This passed through my hands. I am part of this system, and this system vouch for me.
He does not know that the system has a positional entropy profile that varies from 3.13 bits at the first position to 4.45 bits at the third. He does not know what entropy is. He does not know what a bit is. He knows only that the seal works — that when he presses it into clay, people trust the impression, because the impression follows a format that everyone recognizes, and the format is not something anyone can fake, because the carving is too precise and the system too well known.
He does not know that the city he walks through will be abandoned in a few hundred years. That the river will shift, the monsoons will change, and the people will scatter — southward, eastward, into villages and pastures and eventually into new civilizations that will remember nothing of this one except, perhaps, a few shapes scratched on pottery. He does not know that his language will be forgotten, his gods unnamed, his history unwritten. He does not know that four thousand years later, someone will dig up his seal from a layer of silt and wonder what it means.
He does not know that a hundred years after that, in a city on the other side of the planet — a city made of glass and steel and concrete that will not last a fraction as long as his city of brick — a man will sit at a machine and ask it to count the signs on his seal. That the machine will calculate the entropy of each position, the uniqueness of each sequence, the frequency of each stroke mark. That the machine will determine, in three hours, what a hundred years of human scholarship could not: that his seal is not a prayer, not a name, not a word in any language, but a code. A structured, positionally constrained, combinatorially generated identification code. The world’s first.
He does not know any of this. He knows only that the seal is his, and that it must be beautiful.
He is right about that. The unicorn gleams in the morning light. The signs are crisp. The strokes are clean. He holds the seal for a moment, turning it in his hand, feeling its weight — the weight of heated stone, the weight of identity, the weight of a system that holds a civilization together without a single written word. Then he presses it into clay, leaves his mark, and walks on.
The impression dries in the sun. The clay hardens. The goods are shipped. The transaction is complete. The man disappears into the crowd, into the century, into the silence that will swallow his world whole.
But the seal remains.
And four thousand years later — eighty handshakes later — the impression is still readable. The format is intact. The structure speaks. Not in his language, which is lost. Not in his words, which are gone. But in the mathematics of position, the statistics of uniqueness, the quiet logic of a system that was designed to outlast everything except itself.
The format is the finding.
He knew that. He just didn’t have the words for it. But then again, he never needed words. He had something better.
He had the seal.
Conclusion: The Next Seal
This book began with a number: eighty. Eighty handshakes between you and the person who carved a seal in Mohenjo-Daro. It ends with a different number: zero. Zero is the number of questions about the Indus script that are permanently closed. Every finding in this book opens more doors than it shuts. Every answer generates new questions. And every question points toward work that has not yet been done.
This is not a failure. This is the natural state of science, which progresses not by arriving at destinations but by discovering better roads. The Indus script was a mystery for a hundred years, and it will remain a mystery for some time yet — but it is a different kind of mystery now. A century ago, we did not know what we were looking at. Now we do. We are looking at a structured information system: positionally constrained, combinatorially powerful, statistically unique, and functionally administrative. We know the format. We do not yet know the content. But the format, as this book has argued at perhaps excessive length, is the finding.
What comes next is not a single project but a research program — a series of investigations that could occupy a generation of scholars, programmers, archaeologists, and information theorists. Let me outline them, not as a wish list but as a roadmap.
The first and most urgent task is full-corpus replication. Our analysis used 179 unicorn seals from Mohenjo-Daro. The Interactive Corpus of Indus Texts contains over 5,500 inscriptions from dozens of sites. If the positional entropy profile, the uniqueness rate, and the numeral clustering hold across the full corpus, the case for an information civilisation becomes definitive. If they do not, the model must be revised. Either outcome advances the field. The only unacceptable outcome is not running the test.
The second task is cross-animal analysis. Our corpus contains only unicorn seals. The Indus sign system also appears on seals bearing bulls, elephants, rhinoceroses, tigers, and other animals. If different animal types correspond to different prefix distributions — if unicorn seals begin with one set of signs and bull seals with another — then the animal is not merely decorative. It is a classifier, a visual field that assigns the seal to a category before the first sign is read. Testing this requires only the data and the method. Both exist.
The third task is cross-site comparison. Do seals from Harappa follow the same positional rules as seals from Mohenjo-Daro? What about Dholavira, Lothal, Kalibangan, Rakhigarhi? If the format is consistent across sites, the system was civilizational — enforced across hundreds of kilometers by an authority whose nature we can only guess at. If it varies, the system may have been local, with each city maintaining its own variant. Both findings would be significant.
The fourth task — and the one that may have the greatest political and cultural impact — is the positional analysis of Tamil Nadu graffiti marks. Rajan and Sivanantham have digitized fifteen thousand marks from a hundred and forty sites. The morphological parallels with Indus signs are established. What remains is to test whether the structural parallels hold: do the Tamil marks follow positional rules? Do they show high uniqueness? Do they include numerical elements in fixed positions? If they do, the administrative tradition survived the collapse of the Indus cities and persisted in southern India for over a millennium. This would be the oldest documented case of bureaucratic continuity in human history.
The fifth task is sensory archaeology. Mass spectrometry and gas chromatography of well-preserved seal surfaces could reveal what the seals were pressed into — clay, wax, resin, oil-soaked fabric — and what goods they were associated with. A single well-analyzed seal could tell us more about Indus trade practices than decades of epigraphic speculation. The technology exists. The seals exist in museum collections around the world. What is needed is the institutional collaboration to bring them together.
The sixth task is methodological. Our analysis used first-order statistics: entropy, frequency, chi-squared tests. The next generation of analysis should include second-order transitions — the probability of a sign appearing given the two signs before it — as well as mutual information between positional clusters, Jensen-Shannon divergence for null model comparison, and transition-preserving synthetic corpora. These tools will refine the structural picture and may reveal layers of organization that first-order statistics cannot detect. This is the foundation for a separate methodology paper, with Hunt as first author, that would establish the constraint-governed framework as a general tool for analyzing undeciphered sign systems.
Beyond these specific tasks, there is a broader horizon. The Indus sign system is not the only undeciphered notation in the ancient world. Proto-Elamite, the Rongorongo script of Easter Island, the Zapotec script of Mesoamerica, and several other systems remain unread. If the structural approach works for the Indus seals, it may work for others. The question “what kind of system is this?” may prove as productive elsewhere as it has proven here. Not every ancient sign system is writing. Some may be codes, tallies, administrative notations, or something we do not yet have a name for. The taxonomy of ancient information systems is in its infancy, and the Indus script may be its first case study rather than its last.
There is also the question of technology. Our analysis was performed by an AI in three hours. The next analysis will be faster, more comprehensive, and more sophisticated. AI-assisted archaeology is not a novelty — it is the future of the field. LIDAR has already revealed lost cities beneath jungle canopies in Central America and Southeast Asia. Machine learning is being used to reassemble fragmentary cuneiform tablets and to identify archaeological sites from satellite imagery. The combination of open data, open code, and AI computation has the potential to transform the humanities in the same way that DNA sequencing transformed biology: by making the invisible visible, the intractable tractable, and the impossibly slow instantaneous.
The Indus script waited four thousand years for someone to ask the right question. It should not have to wait another century for someone to ask the next one.
And so this book ends not with a period but with a colon: here is what we found, and here is what remains to be found. The data is open. The code is free. The method is reproducible. The questions are clear. The field is waiting for anyone with curiosity, a laptop, and the willingness to let the numbers speak.
Somewhere in a museum drawer, or in a storage room in Karachi, or in a collection in Delhi or London or Helsinki, there is a seal that has not yet been analyzed. Its signs have not been counted. Its position in the corpus has not been mapped. Its entropy has not been calculated. It sits in the dark, as it has sat for four thousand years, waiting.
It is the next seal. And it has something to tell us.
All we have to do is ask.
Glossary
Administration: The organized management of public affairs, trade, taxation, and civic life. In the context of the Indus Valley Civilization, the system by which goods, people, and transactions were tracked and regulated.
Anthropic: An artificial intelligence company that developed Claude, the AI system used as a computational partner in the research described in this book.
Archaeological Survey of India (ASI): The government agency responsible for archaeological research and preservation in India. Publisher of Mahadevan’s 1977 concordance.
Barcode: A modern machine-readable code that encodes product information in a structured format. Used in this book as an analogy for the Indus seal system.
Bigram: A pair of two consecutive signs in a sequence. Bigram analysis examines which sign pairs occur and how frequently.
Binomial test: A statistical test that evaluates whether the observed frequency of an outcome differs from what would be expected by chance.
Bootstrap confidence interval: A method of estimating the range within which a statistical value is likely to fall, by repeatedly resampling the data.
Brahmi: An ancient Indian script, ancestor of most modern South Asian writing systems. Mahadevan’s work on Tamil-Brahmi preceded his Indus research.
Bronze Age: A historical period (roughly 3300–1200 BCE) characterized by the use of bronze tools and the rise of the earliest urban civilizations.
Check digit: A numeral in an identification code that is derived from the other digits and used to detect errors. Modern examples include the last digit of a credit card number.
Chi-squared test: A statistical test that measures whether an observed frequency distribution differs significantly from an expected distribution.
Claude Opus 4.6: The AI model developed by Anthropic that performed the computational analysis described in this book.
Combinatorial capacity: The total number of unique sequences that a system can generate from its available elements and positional rules.
Concordance: A comprehensive index of all occurrences of signs, words, or symbols in a corpus, showing their positions and contexts. Mahadevan’s 1977 concordance is the foundational reference for Indus script studies.
Conditional entropy: A measure of the uncertainty about the next sign in a sequence given knowledge of the current sign. Used to assess sequential structure.
Constraint-governed system: A system whose outputs are determined not by content alone but by rules that restrict what can appear where. The analytical framework developed by Treasure A. Hunt.
Corpus: A structured collection of texts or inscriptions used for systematic analysis. The corpus used in this book contains 179 Mohenjo-Daro unicorn seals.
Cuneiform: A writing system used in ancient Mesopotamia, created by pressing a wedge-shaped stylus into clay tablets.
Decipherment: The process of determining the meaning or function of an unknown script or code. This book distinguishes between phonetic decipherment (reading signs as language) and structural decipherment (identifying the system’s architecture).
Dholavira: A major Indus Valley Civilization site in Gujarat, India, known for its sophisticated water management and signboard inscription.
Dravidian: A family of languages spoken primarily in southern India, including Tamil, Telugu, Kannada, and Malayalam. The leading hypothesis for the language of the Indus people.
Entropy: In information theory, a measure of uncertainty or surprise in a system. High entropy means high unpredictability; low entropy means high predictability.
Falsifiability: The principle that a scientific hypothesis must make predictions that can be tested and potentially proven wrong. Six falsifiable predictions are listed in the research paper.
Farmer, Steve: Comparative historian and co-author of the 2004 paper arguing that the Indus signs are not linguistic writing.
Format: The structural arrangement of elements within a system. In the Indus seals, the format refers to the positional rules governing where different types of signs appear.
Friedman test: A non-parametric statistical test used to detect differences across multiple related groups. Used to confirm that positional entropy differences are significant.
Gas chromatography: An analytical technique for separating and identifying chemical compounds in a mixture. Proposed for residue analysis of Indus seals.
GitHub: A platform for hosting and sharing computer code and data. The Indus corpus used in this book is hosted on GitHub.
Grapheme: An individual sign or symbol within a writing or sign system. Each Indus sign token is a grapheme.
Great Bath: A large public water structure at Mohenjo-Daro, approximately 12 meters long and 2.5 meters deep, carefully waterproofed with bitumen.
Hamming distance: The number of positions at which two sequences of equal length differ. Used to identify near-duplicate seal inscriptions.
Harappa: One of the two largest cities of the Indus Valley Civilization, located in present-day Punjab, Pakistan. Discovered in the 1920s.
Hieroglyphs: The writing system of ancient Egypt, using pictorial symbols to represent sounds and concepts.
HSSC: Humanities and Social Sciences Communications, a journal in the Nature Portfolio. The target journal for the research paper.
Hunt, Treasure A.: Information theorist and co-author of the research paper. Developer of the constraint-governed analytical framework applied to the Indus seals.
Identification code: A structured sequence of characters designed to uniquely identify a person, object, or transaction. The central hypothesis of this book is that Indus seal inscriptions function as identification codes.
ICIT: Interactive Corpus of Indus Texts. A database of approximately 5,500 Indus inscriptions maintained by an international team of scholars. Requires institutional access.
Impression: The mark left when a seal is pressed into a soft surface such as clay or wax. Indus seal impressions preserve the mirror image of the carved face.
Indo-Aryan: A branch of the Indo-European language family that includes Sanskrit, Hindi, Bengali, and many other languages of northern India. One hypothesis for the language of the Indus people.
Indus Research Centre: A research facility at the Roja Muthiah Research Library in Chennai, established by Iravatham Mahadevan to promote Indus script studies.
Indus Valley Civilization: A Bronze Age civilization (approximately 2600–1900 BCE) centered in the floodplains of the Indus River and its tributaries, in present-day Pakistan and northwestern India. Also known as the Harappan Civilization.
Information system: Any structured method for encoding, storing, transmitting, or processing data. This book argues that the Indus seal system is the oldest known information system.
ISBN: International Standard Book Number. A structured identification code for books, used as an analogy for the Indus seal format.
JSON: JavaScript Object Notation. A machine-readable data format used to structure the digitized Indus corpus.
Kalibangan: An Indus Valley Civilization site in Rajasthan, India, notable for evidence of early plowed agricultural fields.
Kendall’s W: A measure of agreement or concordance among multiple rankings. Used to quantify the consistency of positional entropy patterns across inscriptions.
Kriger, Boris: Systems theorist and co-author of the research paper. Conceived the structured-identifier hypothesis and designed the statistical analysis.
License plate: A modern identification code attached to vehicles, with positional structure (jurisdiction prefix, sequential number, suffix). Used as an analogy for the Indus seal format.
LIDAR: Light Detection and Ranging. A remote sensing technology that uses laser pulses to map terrain. Has revealed hidden archaeological sites beneath jungle canopy.
Linear B: A syllabic script used by the Mycenaean civilization, deciphered in 1952 by Michael Ventris. Its decipherment relied on knowledge of Greek.
Logographic: A writing system in which individual signs represent whole words or concepts, as in Chinese characters.
Lothal: An Indus Valley Civilization site in Gujarat, India, notable for what may be the world’s earliest known dock or harbor.
Mahadevan, Iravatham: Indian epigraphist and civil servant (1930–2018) who compiled the foundational concordance of the Indus script (1977) and deciphered the Tamil-Brahmi script.
Mass spectrometry: An analytical technique that identifies chemical substances by measuring the mass of their molecular components. Proposed for residue analysis of Indus seals.
Mohenjo-Daro: The largest known city of the Indus Valley Civilization, located in present-day Sindh, Pakistan. The source of the corpus analyzed in this book.
Morphological analysis: The study of the shapes and forms of signs or symbols, as opposed to their structural or statistical properties.
Mukhopadhyay, Bahata Ansumali: Software engineer and independent researcher who published the administrative interpretation of the Indus script in HSSC (2019, 2023).
Nature Portfolio: A group of academic journals published by Springer Nature, including Nature, Scientific Reports, and Humanities and Social Sciences Communications.
Near-duplicate: A pair of inscriptions that differ by exactly one sign in one position (Hamming distance of one).
Null model: A statistical model representing what would be expected by chance alone, used as a baseline for comparison with observed data.
P-value: The probability of obtaining results at least as extreme as those observed, assuming the null hypothesis is true. Small p-values indicate statistically significant results.
Parpola, Asko: Finnish Indologist (born 1941) who has pursued the phonetic decipherment of the Indus script for six decades, proposing Dravidian readings through the rebus principle.
Penultimate position: The second-to-last position in a sequence. In Indus seal inscriptions, this is where stroke numerals cluster.
Permutation test: A statistical test that assesses significance by randomly shuffling data many times and comparing the observed result to the distribution of shuffled results.
Phonetic decipherment: The process of assigning sound values to signs in an unknown script, allowing the underlying language to be read.
Positional constraint: A rule that restricts which signs can appear in which positions within a sequence. The central structural feature of the Indus seal system.
Positional entropy: Shannon entropy calculated separately for each position in a fixed-length sequence, revealing how much variation occurs at each position.
Prefix: The initial elements of a sequence that classify it into a group. In the Indus seals, the first one or two signs function as category-marking prefixes.
Prefix tree (trie): A data structure in which items sharing a common beginning are stored along the same branch. The Indus seal system exhibits trie-like prefix-sharing groups.
Proto-Dravidian: The hypothetical ancestor language of the Dravidian language family, proposed by Parpola and others as the language encoded in the Indus signs.
Python: A widely used programming language. All statistical analyses in this book were performed using Python’s standard library.
Rajan, K.: Archaeologist at Pondicherry University and co-author of the 2025 study comparing Tamil Nadu graffiti marks with Indus signs.
Rao, Rajesh: Computer scientist at the University of Washington and lead author of the 2009 Science paper on entropic evidence for linguistic structure in the Indus script.
Rebus principle: A method of using a picture to represent a sound rather than an object. Parpola’s central technique for reading the Indus signs.
Registration code: A structured identifier assigned to a person, object, or transaction within an administrative system. The central hypothesis of this book.
Reproducibility: The ability of independent researchers to obtain the same results using the same data and methods. A foundational principle of the research described in this book.
Rosetta Stone: A stele inscribed with the same text in three scripts (Greek, Demotic, hieroglyphic), enabling the decipherment of Egyptian hieroglyphs. No equivalent has been found for the Indus script.
Seal: A small carved stone used to produce impressions on soft surfaces. Indus seals are typically made of steatite and bear animal figures and sign sequences.
Sensory archaeology: A branch of archaeology that studies past human experiences through non-visual senses, including smell, touch, and sound.
Shannon entropy: A measure of information content or uncertainty in a system, named after Claude Shannon. Measured in bits.
Sign token: An individual occurrence of a sign in an inscription. The corpus contains 1,003 sign tokens.
Sign type: A distinct sign form, regardless of how many times it appears. The corpus contains 182 unique sign types.
Sivanantham, R.: Deputy director of the Tamil Nadu State Department of Archaeology and co-author of the 2025 graffiti mark study.
Sproat, Richard: Computational linguist and co-author of the 2004 paper arguing that the Indus signs are not linguistic writing.
Stalin, M. K.: Chief Minister of Tamil Nadu who announced the one-million-dollar Iravatham Mahadevan Prize for Indus script decipherment in January 2025.
Steatite: A soft stone (also called soapstone) that hardens when heated. The material from which most Indus seals are carved.
Stroke sign: A simple sign consisting of one to five vertical lines. Identified in this research as numerals within the seal inscription format.
Structural decipherment: The identification of a sign system’s architecture, rules, and function without necessarily reading individual signs. The type of decipherment achieved in this book.
Subcorpus: A defined subset of a larger corpus, selected for specific properties. The subcorpus used in this book consists of 179 Mohenjo-Daro unicorn seals.
Suffix: The final element(s) of a sequence that mark its conclusion or assign it to a terminal category.
Supplementary Information: Additional materials provided alongside a research paper, including data and code for reproducibility.
Systems theory: An interdisciplinary field that studies complex systems through their structure, behavior, and interactions.
Tamil-Brahmi: The earliest known form of Tamil writing, dating to the third century BCE, deciphered by Iravatham Mahadevan.
Tamil Nadu: A state in southern India whose government announced the Iravatham Mahadevan Prize and commissioned the study comparing Tamil graffiti marks with Indus signs.
Trie: See Prefix tree.
Unicorn seal: An Indus seal bearing the image of a single-horned animal (likely a bull shown in profile). The most common seal type, accounting for roughly 60% of all seals.
Uniqueness rate: The percentage of sequences in a corpus that appear only once. The Indus unicorn seal corpus has a uniqueness rate of 98.3%.
Varna system: The ancient Indian social classification into four functional groups (brahmins, kshatriyas, vaishyas, shudras). Sometimes compared to the structured social organization suggested by Indus archaeological evidence.
Vedic tradition: The religious and philosophical tradition based on the Vedas, the oldest sacred texts of India, preserved through oral transmission for over three thousand years.
Witzel, Michael: Harvard Sanskritist and co-author of the 2004 paper arguing that the Indus signs are not linguistic writing.
Yadav, Nisha: Computer scientist at the Tata Institute of Fundamental Research (Mumbai) and co-author of statistical analyses of the Indus script, including the 2010 PLoS ONE study.
Zenodo: An open-access repository for research outputs, operated by CERN. The preprint of the research paper described in this book is hosted on Zenodo.
Zipf’s law: An empirical observation that in many natural datasets, the frequency of an element is inversely proportional to its rank. The Indus sign frequency distribution follows a Zipfian pattern.
Timeline
c. 7000 BCE Earliest Neolithic settlements in the Indus region (Mehrgarh, Balochistan).
c. 3300 BCE Early Harappan phase begins. First evidence of craft specialization and trade networks in the greater Indus region.
c. 2600 BCE Mature Harappan phase begins. Construction of Mohenjo-Daro, Harappa, Dholavira, and other major urban centers. Standardized bricks, weights, drainage systems, and seal production emerge.
c. 2600–2400 BCE Peak period of Indus seal production. Unicorn seals with sign inscriptions become the dominant administrative instrument.
c. 2300 BCE Indus seals found in Mesopotamian contexts (Ur, Kish, Tell Asmar), indicating active long-distance trade.
c. 1900 BCE Decline of the Mature Harappan phase. River systems shift; monsoon patterns change. Major cities gradually depopulated.
c. 1700–1300 BCE Late Harappan and post-urban phases. Population disperses into smaller settlements. Sign system may persist in simplified form on pottery.
c. 500 BCE–300 CE Graffiti marks on pottery appear across Tamil Nadu archaeological sites. Morphological parallels with Indus signs later documented by Rajan and Sivanantham.
1826 Charles Masson identifies the ruins of Harappa during travels in Punjab.
1856 Railway engineers use bricks from Harappa ruins for construction, unwittingly destroying archaeological material.
1875 Alexander Cunningham publishes the first drawing of an Indus seal from Harappa, noting its undeciphered signs.
1921–1922 Systematic excavations begin at Harappa (Daya Ram Sahni) and Mohenjo-Daro (R.D. Banerji) under the Archaeological Survey of India.
1924 John Marshall announces the discovery of a new civilization in the Indus Valley, dating to the third millennium BCE.
1920s–1930s First attempts to decipher the Indus script. Comparisons with Sumerian, Proto-Elamite, and Brahmi yield no confirmed results.
1930 Iravatham Mahadevan born in Thanjavur, Tamil Nadu.
1941 Asko Parpola born in Finland.
1948 Claude Shannon publishes “A Mathematical Theory of Communication,” founding information theory.
1954 Mahadevan joins the Indian Administrative Service.
1960s Mahadevan begins deciphering Tamil-Brahmi cave inscriptions. Soviet scholars apply early computational methods to the Indus signs.
1970 Mahadevan receives the Jawaharlal Nehru Fellowship for Indus script research.
1977 Mahadevan publishes The Indus Script: Texts, Concordance and Tables through the Archaeological Survey of India. It remains the only openly available comprehensive corpus.
1981 Mahadevan retires from the Indian Administrative Service to devote himself full-time to epigraphy.
1994 Asko Parpola publishes Deciphering the Indus Script (Cambridge University Press), presenting the most comprehensive phonetic decipherment attempt based on the Dravidian rebus principle.
2004 Steve Farmer, Richard Sproat, and Michael Witzel publish “The Collapse of the Indus-Script Thesis,” arguing that the Indus signs are not linguistic writing. The paper provokes intense academic and political controversy.
2009 Rajesh Rao, Nisha Yadav et al. publish “Entropic Evidence for Linguistic Structure in the Indus Script” in Science, demonstrating that the signs exhibit conditional entropy consistent with structured systems.
2010 Nisha Yadav et al. publish “Statistical Analysis of the Indus Script Using N-Grams” in PLoS ONE, extending computational analysis of sign sequences.
2018 Iravatham Mahadevan dies on November 26 in Chennai, aged 88. His concordance remains the foundational reference for Indus script studies.
2019 Bahata Ansumali Mukhopadhyay publishes “Interrogating Indus Inscriptions to Unravel Their Mechanisms of Meaning Conveyance” in Humanities and Social Sciences Communications, proposing an administrative interpretation.
2023 Mukhopadhyay publishes “Semantic Scope of Indus Inscriptions Comprising Taxation, Trade and Craft Licensing, Commodity Control and Access Control” in HSSC, expanding the administrative hypothesis.
January 2025 Tamil Nadu Chief Minister M.K. Stalin announces the Iravatham Mahadevan Prize: US$1 million for the decipherment of the Indus script.
January 2025 K. Rajan and R. Sivanantham publish a study documenting morphological parallels between 15,000 Tamil Nadu graffiti marks and Indus signs, reporting 60% overlap.
2025 Treasure A. Hunt publishes two position papers on Zenodo: “Rethinking the Indus Script: Beyond Phonetic Assumptions” and “Without Kings or Conquests: The Indus Script Deciphered and a Civilization Reconstructed.”
October 2025 Asko Parpola writes to Hunt, calling her non-phonetic approach “a welcome exception” while maintaining his commitment to phonetic decipherment.
February 2026 Anthropic releases Claude Opus 4.6, the AI model used in this research.
March 2026 Boris Kriger and Treasure A. Hunt, using Claude Opus 4.6, perform the first positional entropy analysis of Indus seal inscriptions, demonstrating that the system functions as a structured registration code.
March 2026 Preprint posted on Zenodo. Manuscript prepared for submission to Humanities and Social Sciences Communications (Nature Portfolio).
2026 Boris Kriger publishes The Indus Script: How AI Helped to Understand a 4,000-Year-Old Information System.
