Reverse-Engineering Latent Balancing Rules in a Collectible Card Game: A Hedonic Pricing and Order-Statistics Approach to Riftbound

Collectible card game (CCG) designers price every card with an internal, unpublished cost model. We ask whether that model can be recovered from static card data alone. Treating a card as a bundle of effect attributes and its printed cost as the designer's price for that bundle, we cast the problem as an inverse hedonic pricing problem and solve it for Riftbound, Riot Games' League of Legends trading card game, using 923 cards from five sets. Prices for the game's core effects and keywords are identified through anchor-pair elimination, simultaneous equations, a piecewise-linear might budget, and least-squares / minimax fits, with robust statistics for small samples. Three mechanism-level models extend the linear core: a mental-accounting model of hand cards (play is rent-free, stock is priced) that reconciles draw, discard, and the Flow keyword; an order-statistics model of selection premia that prices Predict, look-N-take-1, and modal "choose one" effects with one calibrated dispersion parameter per mechanism; and a cost-controlled regression of on-play effects on units that finds a strong low-cost discount (−0.63, p < 0.001) and a weaker combat-trick discount (−0.27, p = 0.057). The recovered price table exposes several latent design regularities: contingent value is booked at about one half (median per-card fraction 0.50, lower for large triggers), play-timing keywords carry no detectable price (95% upper bound 0.36 C), and cheap units receive on-play effects at a fraction of spell price. Parameters calibrated on one group of cards predict held-out cards (e.g. discard cost predicted at 0.83–1.04 C versus 1.00 C observed). Across 497 fully priced cards the ratio of rule-booked value to cost has median 1.00 (interquartile range 0.90–1.16); on the 246 cards that are neither vanilla nor cited as anchors it has median 0.94 (0.75–1.10), and it falls by about 0.03 per unit of cost, a residual cost gradient that the model does not yet explain. We discuss transfer of the framework to other CCGs and its use as a constraint for procedural card generation. J. C. Lee and J. Chen contributed equally to this work.