The Fidelity Ladder Agentic Monte-Carlo Playtesting for Tabletop-RPG Encounter Calibration An experience report

Balancing a tabletop role-playing game (TTRPG) combat encounter increasingly relies on an abstract Monte-Carlo simulator: thousands of dice rolls under a scripted policy, read off as a party win rate. We estimate one level-9 Level Up: Advanced 5E boss fight at four rising levels of difficulty-estimation fidelity: (i) the static Challenge-Rating / encounter-XP budget; (ii) an abstract, positionless Monte-Carlo model with scripted "competence-band" policies; (iii) an agentic simulation in which each player character is driven by a large-language-model (LLM) policy on a 2D grid, coordinating with allies, while a deterministic engine resolves all dice; and (iv) a human playtest. Our central observation: raising fidelity does not merely shift the difficulty level, it re-ranks party compositions. Holding the boss fixed, the positionless model's confident ordering is overturned -- the composition it rates third (an archer trio, 24%) the agentic model rates at 70% (tied for second), and that trio won at the table -- because low-fidelity models cannot credit value that is positional. We use the agentic tier to calibrate the boss along two axes, lethality and table-time, to a designer target; the final design produced a clutch table win (two of three characters down, one at 10 HP), confirming the estimate and surfacing a further mechanism the simulation under-modeled. We claim no novelty for the agentic architecture, the balancing loop, or the fidelity-hierarchy idea -- each is cited prior art. The contribution is the conjunction: a concrete four-tier instantiation for TTRPG encounter design, a measurement of how far and in which direction estimates move between tiers, and a reusable open artifact.