Researchers at Leiden University used two AI agents playing thousands of matches against each other to reconstruct the rules of a lost Roman-era board game carved into a limestone slab found at Coriovallum, the ancient settlement beneath modern Heerlen in the Netherlands. The artifact, a small stone tablet made from white Jurassic limestone quarried in northeast France, had sat in a museum collection for years with its purpose only partially understood. By feeding the board’s grid into a general game system and letting automated players test competing rule sets, the team identified a blocking-style contest that predates chess by centuries and may share structural DNA with game boards scratched into walls across the eastern Roman Empire.
What is verified so far
The core findings appear in a peer-reviewed paper on the Coriovallum board published in the journal Antiquity. The artifact carries museum accession number 04433 at Het Romeins Museum in Heerlen. Its material has been identified as white Jurassic limestone from quarries near Norroy in northeast France, a sourcing detail that ties the slab to known Roman-period supply routes running through the region. The find location, Coriovallum, was a Roman road station and bathhouse settlement along the route from Cologne to the coast, making it a plausible hub for the circulation of portable leisure objects such as game boards.
The research team modeled the slab’s grid inside Ludii, a general game system designed to represent many different board games through modular building blocks called ludemes. The Ludii architecture, outlined in an arXiv preprint, allows researchers to define a game’s components, then run AI agents against each other to evaluate whether a given rule set produces meaningful strategic play or degenerates into trivial outcomes. For the Coriovallum board, the agents cycled through candidate rule sets, discarding those that ended too quickly, produced forced wins, or failed to generate the kind of back-and-forth decision-making that would make a game worth playing repeatedly.
The surviving rule sets describe a territorial blocking game. Players place or move pieces to restrict an opponent’s options, and the contest ends when one side can no longer act. That structure is older than the move-and-capture logic familiar from chess, which emerged centuries later in South Asia. The Leiden team’s approach treats the physical traces on the stone-the grid lines, intersection points, and any visible wear patterns-as constraints that any valid rule set must satisfy, then uses self-play tournaments to explore which hypothetical games remain fun and balanced under those constraints.
A separate line of evidence connects the Coriovallum board to similar game grids documented far to the east. At Aphrodisias in modern Turkey, a carved grid catalogued as iAph 2007 entry 6.3 was cut into the Basilica floor during the late Roman or Byzantine period. Academic work on game boards at Aphrodisias has shown that geometric layouts with blocking-game features appear across multiple structures at the site. The overlap in grid geometry between Heerlen and Aphrodisias suggests that Roman soldiers, traders, or administrators carried these games across vast distances, scratching boards into whatever surface was available when purpose-made slabs were not at hand.
What remains uncertain
Several questions sit beyond the reach of current evidence. No primary excavation field notes or find-context photographs from the original Heerlen dig have been cited beyond the museum accession number. Without that documentation, the exact archaeological layer and associated dating material for the slab remain difficult to pin down independently. The published study discusses a late-Roman-period dating window based on stylistic and contextual arguments, but the precision of that estimate depends on supporting evidence that has not been reproduced in full for outside reviewers.
The Ludii game-description files and exact search parameters used in the self-play runs have not been released as a separate dataset, which means independent researchers cannot yet replicate the AI tournaments step by step. The team’s rationale for discarding specific rule variants is available only through the published paper and secondary summaries rather than through raw experimental logs or code repositories. This is not unusual for computational archaeology, where projects often rely on custom pipelines, but it limits the speed at which other groups can verify or extend the results.
The comparison between Coriovallum and Aphrodisias is suggestive but not conclusive. Geometric similarity between two game boards does not prove they were played under the same rules or that one descended from the other. Comparative measurements and precise dating details for the Aphrodisias boards have been referenced in the literature but not reproduced side by side with the Heerlen slab’s dimensions in a single, standardized analysis. A stronger test would involve running identical Ludii self-play tournaments on every documented Roman-era grid and comparing how often the same rule sets emerge as viable, but that work has not yet been done and would require a coordinated corpus of digitized boards.
Even the core classification of the Coriovallum artifact as a game board, while widely accepted, is not absolutely airtight. In principle, any carved grid might have served some other function-ritual, educational, or practical. The authors argue that the layout, scale, and parallels at other Roman sites make a gaming interpretation the most economical explanation, but alternative readings cannot be ruled out unless new contextual evidence surfaces.
How to read the evidence
The strongest evidence here is the physical artifact itself, a stone object with a fixed grid that constrains what games could have been played on it, combined with the reproducible capacity of modern AI tools to test large families of rule sets against that constraint. The Ludii simulations do not prove that Roman players at Coriovallum used the exact rule set favored by the AI search, but they show that certain plausible patterns of play are compatible with the board while many others are not. That negative filtering-eliminating games that are clearly ill-suited to the grid-is as important as the positive proposal of a specific blocking game.
Readers should therefore treat the reconstructed rules as a best-fit hypothesis grounded in both material culture and computational modeling, not as a recovered instruction manual. The artifact demonstrates that some grid-based game was played in late Roman Heerlen on imported limestone. The AI experiments demonstrate that certain types of blocking contests work well on that grid and produce the kind of balanced, replayable experience likely to sustain a gaming tradition. The historical link between those two facts is persuasive but inferential.
The broader implication is methodological. By combining detailed archaeological recording with flexible game-modeling software and self-play tournaments, researchers can explore the design space around fragmentary or ambiguous boards more systematically than before. Similar approaches could be applied to other carved grids from Roman forts, marketplaces, and religious complexes, gradually building a comparative picture of how rule sets, board geometries, and social contexts interacted across the empire.
For non-specialists, the key is to distinguish between what the stone and its documented context can tell us directly and what the AI reconstructions add as a probabilistic layer on top. Museum records, petrographic analysis, and parallels at other sites anchor the object in time and space. Algorithmic self-play explores the landscape of possible games that might have animated that object in daily use. Both strands are valuable, but they speak with different degrees of certainty.
Access to the underlying article and supplementary materials depends on journal subscriptions and platform tools; readers navigating the Antiquity publication can find technical and access details through the Cambridge support pages, which explain how to reach full texts and related content. As more datasets and modeling files become openly available, the Coriovallum project may serve as a template for how AI can illuminate not only how ancient people fought wars or built roads, but also how they passed the time between those larger events-hunched over a small stone board, trying to block a friend’s final move.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
