A stone board game carved during the Roman occupation of what is now the Netherlands has been identified as a blocking game, a type of strategy contest that was not previously documented in Europe for that period. Researchers from Dutch, Belgian, and Australian institutions used two AI agents to simulate thousands of possible rule sets against physical wear evidence on the artifact, which is housed in Het Romeins Museum in Heerlen. The result is a peer-reviewed reconstruction of a lost game they call Ludus Coriovalli, and it suggests that path-blocking mechanics reached the Roman frontier centuries earlier than existing records indicated.
A stone board, two AI players, and groove marks that told the story
The artifact at the center of this work is a stone slab incised with a grid of lines, recovered from the Roman settlement of Coriovallum, the ancient name for Heerlen. Researchers conducted use-wear analysis and 3D scanning of the board’s surface, documenting differential groove wear along specific lines. Some grooves were worn deeper than others, a pattern consistent with game pieces being dragged repeatedly along preferred paths rather than placed randomly.
That physical evidence became the filter for an AI-driven experiment. The research team, working under the Digital Ludeme Project funded by European Research Council Consolidator Grant No. 771292, programmed two AI agents to play candidate rule sets derived from historic Northern European board games. Each simulated match generated its own pattern of piece movement across the grid. The team then compared those simulated wear distributions against the actual groove wear on the Heerlen stone.
Only one family of rules produced wear patterns that matched the physical evidence: a blocking game in which players win by restricting an opponent’s movement rather than capturing pieces. The Antiquity article published by Cambridge University Press details how simulations combined with use-wear data converged on this blocking interpretation, ruling out capture-based and race-based alternatives. In the successful rule family, pieces advance along preferred lanes, jockeying for positional advantage until one player’s options collapse.
In practice, the AI agents served as tireless test players. They ran through thousands of games under each candidate rule set, producing heat maps of how often each line segment on the virtual board was used. These heat maps were then compared to the real stone’s 3D wear profile. When the researchers tried race-style or capture-heavy games, the predicted wear spread more evenly across the grid than the artifact shows. Only the blocking-style rules produced the distinct corridors of heavy wear and adjacent zones of relative inactivity that the scans revealed.
What remains uncertain about the Coriovallum game
Several gaps persist in the reconstruction. The full set of candidate rule sets tested by the AI simulations has not been published beyond high-level descriptions. Readers looking for raw 3D scan datasets or exact differential groove measurements will not find them in the current paper. The researchers describe their method and results but have not released the underlying digital models of the artifact for independent replication or re-analysis.
The provenance chain of the stone itself also lacks direct institutional documentation in the published record. The artifact’s Roman-era context in Coriovallum is established in the Antiquity paper, and a summary on Phys.org confirms the involvement of Dutch, Belgian, and Australian teams. But no separate museum catalog entry or archaeological field report has been cited to trace the stone’s discovery, excavation circumstances, or accession history in detail. For now, the contextual dating rests on the broader archaeological understanding of the Heerlen site rather than a fully published chain of custody for this specific object.
The blocking-game identification also raises a question the paper does not fully resolve: how did this game mechanic arrive at a Roman military outpost in the Low Countries? Blocking games are well attested in later medieval Scandinavian and North African traditions, but the Coriovallum artifact, if correctly dated to the Roman period, would predate those records by several centuries. Whether the game traveled along trade routes, with military units, or emerged independently remains open. The authors outline these possibilities but acknowledge that, without textual mentions or additional boards of similar design, any proposed transmission route is speculative.
Even within the blocking framework, details remain fuzzy. The starting arrangement of pieces, the exact number of tokens per player, and whether there were special moves or jump mechanics cannot be read directly from the grooves. The AI-assisted reconstruction offers a plausible, internally coherent ruleset, but it is best understood as a historically informed model rather than a definitive recovery of the original players’ experience.
Separating physical proof from algorithmic inference
Two distinct types of evidence support the blocking-game conclusion, and they carry different weights. The primary physical evidence, the differential groove wear documented through 3D scanning, is direct and observable. Deeper wear along certain lines means pieces moved along those paths more often. That pattern is not an interpretation; it is a measurable feature of the artifact’s surface.
The second layer, the AI simulation results, is inferential. The researchers built a model, fed it candidate rules, and checked which rules produced wear patterns closest to the real stone. This approach is powerful but depends on assumptions: that the candidate rule sets were broad enough to include the actual game, that piece movement is the primary cause of groove wear, and that the simulation accurately models how stone wears under repeated contact. The paper addresses these assumptions but cannot eliminate all alternatives, especially unusual play styles, house rules, or post-depositional damage that might mimic play wear.
What makes the study distinctive is the combination. Neither the groove wear nor the simulations alone would be sufficient. Groove wear without simulation could support many different games. Simulations without physical evidence would be speculative modeling. Together, they narrow the possibilities to a specific game type, though the exact rules, including starting positions and win conditions, are reconstructed rather than recovered from any written source. The authors present their preferred ruleset as the most parsimonious explanation that fits both the archaeological data and the AI-generated movement profiles.
The broader significance extends beyond a single Roman-era board. The Digital Ludeme Project, which funded this work through its ERC grant, aims to model the evolution of traditional games across cultures and centuries. The Coriovallum stone is one test case for a method that could be applied to dozens of other unidentified game boards sitting in museum collections across Europe and the Mediterranean. If AI-driven wear analysis can recover lost rules from physical traces alone, it opens a new channel for understanding how ancient communities spent their leisure time, what strategic thinking they valued, and how ideas traveled across frontiers.
For researchers and readers interested in delving deeper into the methodology, Cambridge University Press provides support resources through its Cambridge Core help pages, which explain access options for journals like Antiquity. Those seeking to ask about data availability or permissions related to the Ludus Coriovalli study can also use the publisher’s contact information portal to reach the appropriate editorial or technical teams. As more archaeological AI projects come online, similar channels for sharing scans, models, and code will be crucial to testing how far algorithms can go in reconstructing the games people once played on carved stone in distant provinces of the Roman world.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
