How do you reconstruct rules for a game nobody has played in 1,500 years? Let AI figure it out

Leiden University / Maastricht University / Flinders University

What do you do with a carved stone that might be a game board but might not, from a civilization that left no instruction manual? For decades, archaeologists stared at a limestone slab from Roman-era Heerlen in the Netherlands - a flat surface scored with intersecting lines and marked by uneven wear - and could not agree on what it was or how it worked.

Now a team of researchers has tried something no one had attempted before: they handed the problem to artificial intelligence and asked it to play.

A stone scored with lines and worn by centuries of fingers

The object itself is modest. A carefully shaped piece of limestone, engraved with a geometric pattern of intersecting lines, recovered from what is now Heerlen in the southern Netherlands. The site was part of the Roman province of Germania Inferior, and the stone dates to the Roman period of occupation.

Two features mark it as something more than decoration. First, the pattern of lines forms a grid-like structure consistent with board game layouts known from other ancient cultures. Second, and more importantly, the stone shows visible wear - not uniform wear from age or handling, but concentrated wear along specific lines and at specific intersections. The pattern suggests game pieces being slid repeatedly across the surface over extended periods.

Dr. Walter Crist, an archaeologist at Leiden University who specializes in ancient games, puts it simply: the wear pattern points strongly to repeated play rather than any other purpose. But knowing something is a game board and knowing how the game was played are very different problems.

Simulating hundreds of rule sets to match the wear

Most everyday Roman games were drawn in dust or carved into wood - materials that do not survive the centuries. Written rules for Roman board games are scarce, and the few that exist tend to describe well-known games like ludus latrunculorum (a tactical capture game) that do not match this stone's layout. The Heerlen board represents something else, something with no surviving documentation.

The research team, which included scholars from Maastricht University, Leiden University, the Universite Catholique de Louvain, Flinders University, and the Roman Museum in Heerlen, turned to the Digital Ludeme Project's AI system called Ludii. Ludii is a general game-playing platform designed to model and simulate a vast range of board games, both historical and hypothetical.

The approach was methodical. The researchers programmed two AI agents to play against each other on a digital version of the Heerlen board, using rule sets drawn from documented ancient European board games - including haretavl from Scandinavia and gioco dell'orso from Italy. They ran the simulations repeatedly, adjusting rules each time, and tracked where the simulated game pieces moved most frequently.

The question was precise: which rule sets produce simulated wear patterns that match the actual wear on the stone?

A blocking game, not a capture game

The simulations pointed strongly to a specific genre: a blocking game. In blocking games, the objective is to trap your opponent's pieces by preventing movement, rather than capturing them by jumping or displacement. Players maneuver their pieces to restrict the opponent's options until one side cannot make a legal move.

This finding carries a historical implication that extends beyond one stone. Blocking games are scarcely documented before the Middle Ages. The earliest known examples come from medieval Scandinavia and continental Europe. If the Heerlen stone is indeed a blocking game board from the Roman period, it pushes the documented history of the genre back by several centuries.

Dr. Matthew Stephenson, a computer scientist at Flinders University who worked on the AI simulations, emphasizes that the match between simulated and actual wear patterns does not constitute proof that the stone was used for one specific game with one specific rule set. Multiple blocking game variants produced similar wear patterns. What the AI analysis establishes is the genre of play - blocking rather than capture, pursuit, or race - and it does so with considerably more rigor than visual inspection alone could provide.

What AI-driven archaeology can and cannot do

The method has real power and real limits, both worth stating clearly.

The power lies in the systematic nature of the approach. A human archaeologist looking at wear patterns on a stone makes qualitative judgments: this area looks more worn than that one. The AI approach quantifies those judgments. It generates precise predictions about where wear should concentrate under different rule sets and compares those predictions against measured wear on the actual artifact. It can test hundreds of rule combinations that would take a human researcher years to work through manually.

The limits are equally important. The AI can only test rule sets that researchers think to program into it. If the actual game followed rules entirely unlike any documented historical game, the simulations would not find it. The method assumes that the wear patterns on the stone are primarily the result of gameplay, but other activities - scoring, ritual use, or simply idle scratching - could contribute to wear in ways that overlap with game-related patterns.

The stone is also a single artifact. Without additional examples of similar boards from similar contexts, it is difficult to assess how representative the Heerlen stone is. It could reflect a widely played Roman game that simply left few traces, or it could be an oddity - a local invention or an import from a culture where such games were common but whose gaming traditions are poorly documented.

The study also does not reconstruct a complete, playable rule set. It identifies the likely genre of play and narrows the range of plausible rules, but the specific mechanics - how many pieces each player starts with, whether pieces can move backward, how the game ends - remain uncertain.

Other mysterious artifacts, waiting for their turn

The research, published in Antiquity, grew out of the Digital Ludeme Project, a European Research Council-funded initiative at Maastricht University that uses artificial intelligence to reconstruct ancient games. The project has built a database of historical game rules and a computational framework for testing them against archaeological evidence.

Crist and Stephenson both see the Heerlen stone as a proof of concept rather than an endpoint. Archaeological collections around the world contain objects that might be game boards but cannot be identified because no rules survive. Stones with incised grids, pottery fragments with geometric patterns, wooden boards with mysterious markings - many of these objects currently sit in museum storage, catalogued but unexplained.

The AI-driven approach offers a new way to interrogate them. By generating testable predictions about how different games would produce different patterns of wear, damage, or piece placement, the method transforms static artifacts into sources of behavioral information. It does not replace traditional archaeological analysis. It adds a layer of quantitative rigor that was previously unavailable.

Somewhere in a museum basement, there may be a stone with grooves worn smooth by Roman fingers playing a game whose rules have been lost for a millennium and a half. We now have a tool that might be able to figure out what they were playing.