In a museum depot in the Dutch town of Heerlen, a flattened limestone slab carved with intersecting lines has sat for decades, cataloged but not fully understood. Archaeologists agreed it looked like a Roman game board, yet no one could say how it was played or even what game it represented. Now artificial intelligence is being lined up as the tool that could finally turn this mute object into a playable set of rules, transforming a local curiosity into a test case for how we read the ancient world.
The emerging picture is that of a blocking game, tentatively dubbed Ludus Coriovalli, that may predate familiar Roman pastimes and echo the tactical mindset of the empire’s soldiers. If current research plans bear out, AI will not simply “solve” a puzzle, it will force historians to rethink how they reconstruct lost cultures, from the mechanics of leisure to the logic of military training.
The slab in Heerlen and a century of guesswork
The story begins in the Roman town of Coriovallum, the settlement that underlies modern Heerlen near the crossroads of important imperial roads in the province of Germania Inferior. Excavations there uncovered a worn limestone block etched with a crisscross of grooves, one long straight line, and a grid that strongly suggested a playing surface. The object was intriguing enough to be preserved, but not distinctive enough to be matched to any known Roman game, so it slipped into the background of local collections.
For years, specialists floated familiar labels such as Ludus Latrunculorum, the well attested “game of brigands” that appears in literary sources and on other boards. Yet the pattern on the Coriovallum stone did not quite fit the standard layouts associated with that game, and no inscription spelled out its name. Reports on ancient playthings, including work on Roman board games more broadly, underscored how often archaeologists are left with only partial clues: a grid here, a handful of counters there, and a wide interpretive gap in between.
Dueling AIs and the birth of “Ludus Coriovalli”
Into that gap steps a new generation of computational archaeology, with researchers at Leiden University preparing to treat the Heerlen slab as a laboratory for game-playing algorithms. According to a project description from Leiden researchers, the plan is to encode the board’s geometry, then let artificial agents explore possible move sets that respect the carved lines and plausible Roman playing conventions. The working label for the reconstructed pastime, Ludus Coriovalli, anchors the game in its findspot while signaling that it is not simply a clone of better known titles.
The technical approach borrows from reinforcement learning, where software agents learn by trial and error. Coverage of similar projects describes teams programming two AI opponents to play against each other under more than 100 different rule sets, then ranking which variants produce balanced, engaging contests rather than trivial wins or stalemates. In that work, a researcher identified as Crist has emphasized that the goal is not just to find any playable system, but to converge on one that feels consistent with what is known of Roman tastes and daily life.
From carved rock to blocking game
What emerges from these simulations is not a dice-driven race like backgammon, but a positional struggle that looks closer to checkers or abstract strategy games. Reports on the Coriovallum stone describe Thousands of AI test runs that favor interpretations where players move identical pieces on an 8×8 grid, trying to trap opponents rather than capture them outright. The proposed mechanics, in which pieces advance and hem in rivals, have been likened to pawns in chess, but with the emphasis on encirclement instead of promotion.
One striking claim in early analyses is that this blocking game, named Ludus Coriovalli, may predate other European examples of similar mechanics by centuries. Commentators who have examined the AI outputs argue that the game’s focus on cutting off lines of retreat rather than eliminating pieces outright hints at a distinct design lineage. If that holds up under scrutiny, it would push back the timeline for when such abstract positional games took root in the Roman world.
Military logic on a tabletop
The most provocative hypothesis is that Ludus Coriovalli did more than pass the time in taverns or bathhouses. The blocking tactics that the algorithms favor, where players build walls and pockets to immobilize enemy pieces, echo descriptions of encirclement and fortification in Roman military manuals such as Vegetius’ De Re Militari, even if that text is not directly cited in the current research. Coverage of the AI work suggests that the reconstructed rules reward players who think in terms of zones of control and layered defenses, a mindset that would have been second nature to legionaries stationed along the frontier.
That connection is still speculative, and it is important to flag it as such. Yet it gains plausibility when set alongside other finds that tie games to martial elites, such as the 1,700-year-old Roman board game discovered in a burial mound in western Norway. Archaeologists there linked the set, found with gaming pieces and dice, to high-status individuals who moved in imperial trading and military networks. If officers on the empire’s northern edge were buried with such objects, it is not a stretch to imagine soldiers in Coriovallum honing strategic instincts on a carved stone between patrols.
How AI changes the rules of archaeological interpretation
What makes the Coriovallum project so consequential is less the specific game and more the method. Traditional scholarship on ancient games has relied on slow, comparative work, matching boards and pieces to scattered literary references. AI-driven approaches invert that process: they start from the physical layout, generate a vast space of possible rules, and then let simulated play winnow the field. One recent study of Roman gaming used machine learning to decode inscriptions and infer mechanics for Ludus Latrunculorum, showing that algorithms can spot patterns in fragmentary evidence that human readers might miss.
In the Coriovallum case, accounts describe Dueling AI agents that “learn” to exploit weaknesses in each other’s strategies, effectively stress testing candidate rule sets at a scale no human play group could match. Another report on the same line of work notes that the carved limestone from Coriovallum, with its crisscross of grooves and single straight line, was fed into a model that had already been trained on other Roman-era boards, allowing it to recognize when a proposed rule set produced familiar-looking patterns of play. This is not just a faster version of old methods, it is a qualitatively different way of reasoning from artifacts to behavior.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
