CoupVisor: Strategy Optimization by Round and Challenge Decision Support

Cris Huynh

Source-linked AI summary

CoupVisor: Strategy Optimization by Round and Challenge Decision Support

Cris Huynh

arXiv:2608.15868v1Aug 16, 2026cs.AIcs.GTcs.LG

TL;DR

CoupVisor addresses how to act and when to challenge in Coup despite hidden hands and publicly observable evidence. It unifies event representation, belief tracking, decision support, and learning-based policies, finding that reward alignment determines whether BC or DQN performs best, with win-aligned DQN outperforming all baselines.

Problem
Coup requires players to choose actions and challenges under hidden information, but public game evidence is not converted into an explicit probability for evaluating those decisions.
Method
CoupVisor uses a shared event schema with belief tracking and threshold decisions, and compares rule-based, heuristic, behavior-cloning, and deep-Q-learning players under alternative rewards.
Results
Reward definition determines the leading learner: shaped short-term reward favors behavior cloning, whereas win-aligned reward makes deep Q-learning outperform every baseline.
Takeaways & Limitations
Reward alignment can matter more than algorithm choice when learning policies for this decision-support setting.
Takeaways & Limitations
Evaluation uses scripted opponent styles and a simulator that omits a real card-swapping rule, limiting coverage of adaptive deception and potentially biasing belief statistics.

Abstract

from arXiv · show

This paper presents CoupVisor, a decision-support system for the hidden-information card game Coup. It addresses two questions: what a player should do on each turn, and when a player should challenge an opponent's claim. The system is built around a single description of game events, which is shared across manual play, replay of recorded games, simulation, belief tracking, advisor recommendations, and learning-based policies. CoupVisor estimates the chance that a claim is truthful by combining how likely each role is with how many cards the claimant still holds, which corrects a case where the very first claim of a game was flagged as suspicious despite no evidence. We compare a rule-following advisor and several learned and heuristic players across many simulated games and different opponent styles. Our main finding is that the choice of reward, whether it rewards short-term gains or ultimately winning the game, decides which learning approach performs best, and that a win-oriented reward produces a policy that outperforms all baselines.

I. Introduction … E. Threshold Rules and Expected Value

CoupVisor turns Coup’s public game record into explicit beliefs and inspectable challenge advice while supporting manual play, replay, simulation, and policy evaluation through a shared event description. Its technical approach combines role likelihoods, hand-size information, finite-deck reasoning, and threshold decisions grounded in expected value.

I. Introduction: Coup presents two coupled hidden-information decisions: choosing an action and deciding whether to challenge an opponent’s claim.Claims are enforced by challenges, and an incorrect challenger or claimant loses a card.
I. Introduction: CoupVisor maintains explicit opponent-role probabilities and converts them into challenge advice using one event schema shared across system components.The system is an observer and advisor that uses the public record rather than acting as a game-playing bot.
I. Introduction: The paper evaluates the advisor over 500 simulated games spanning five seeds and two opponent conditions, alongside belief ablations, advisor ablations, behavior cloning, deep Q-learning, and heuristics.These experiments vary belief modelling, decision rules, and learned policies within a frozen evaluation testbed.
II. Background: The background section defines the technical terms needed to read the later sections without external reference.It establishes the paper’s terminology before introducing Coup’s rules and the POMDP framing.
A. Coup and the Public Record: Coup uses five roles with three copies each; players begin with two hidden influence cards and two coins, while actions impose costs and legality constraints.Assassinate costs three coins, Coup costs seven, and players holding ten or more coins must Coup.
B. Imperfect Information and the POMDP: Because opponents’ cards are hidden, Coup fits the POMDP formalism, in which an agent acts using a belief state over hidden game states.The formalism includes hidden states, actions, transitions, observations, and an observation rule.
C. Belief, Prior, Posterior, and the Bayesian Update: CoupVisor updates role beliefs from a uniform prior of 0.20 using likelihood ratios, including about 5.7 for ordinary claims and 4.5 for Steal blocks.The likelihood ratio compares claim frequency when a player holds a role with claim frequency when they do not.
D. From Per-Card Belief to Per-Hand Probability: The advisor converts per-card beliefs into per-hand claim probabilities, blends independent and hypergeometric estimates, and challenges when p_truth < θ.The deployed threshold is inspectable and corresponds to θ = G/(G + L), where G is the gain from catching a false claim and L is the loss from wrongly challenging a true one.

F. Reinforcement Learning Terms … E. Coin Pressure and Legal Action Constraints

CoupVisor combines offline reinforcement learning concepts, challenge evaluation, and a shared event-driven framework for round advice, belief updates, and legal-action filtering. Its methodology defines reward alternatives, challenge thresholds, and game constraints that shape learning and recommendations.

F. Reinforcement Learning Terms: RL learns a policy π(a|s) by maximizing discounted future rewards, while Q-learning estimates action values and DQN stabilizes training with replay and target networks.The discount factor γ trades immediate against future reward.
F. Reinforcement Learning Terms: Behavior cloning imitates logged demonstrator actions without rewards, whereas offline RL learns from fixed simulator transitions and risks distribution shift.Offline policies receive no further environment interaction.
F. Reinforcement Learning Terms: Reward shaping adds coin gain and retained influence to create denser feedback, while win-aligned reward credits actions matching the eventual winner’s logged actions.Both reward definitions appear verbatim in Appendix B.
G. Exploration Versus Exploitation: Strong adversarial play requires mixed strategies because deterministic policies can be inferred and exploited by opponents.Exploration also improves uncertain value estimates and avoids predictability.
H. Evaluation Terms: Challenge evaluation uses precision, recall, F1, outcome accuracy, and calibration to assess correctness, false-claim detection, hindsight decisions, and probability reliability.Calibration compares stated probabilities with observed frequencies.
III. Related Work: The methodology combines POMDP-style role beliefs, Markov transition tuples, imitation learning, Random Forest supervision, DQN baselines, offline RL, and standard classification metrics.Evaluation also includes ROC-AUC and calibration-bin analysis.
IV. Framework and Methodology; A. Overall Framework; B. Round Strategy Optimization Loop; D. Belief Update Logic: A single event pipeline keeps manual entry, recorded-game replay, simulation, training data, and live advice consistent, while each round recomputes legal actions and expected utilities.Legality filtering precedes scoring, including forced-Coup constraints at ten coins.
C. Challenge Decision Logic: Challenge logic blends role beliefs into p_truth, applies context-dependent thresholds, and converts per-card priors so turn-one Tax claims are not spuriously challenged.Without conversion, the 0.20 prior falls below the 0.25 Tax threshold; under (3), it becomes approximately 0.36.

F. Advisor Reasoning Under Bluff Pressure … N. Non-Learning Baseline Agents

CoupVisor combines probabilistic beliefs, rule-based legality, and risk-sensitive reasoning to advise actions and challenges in Coup. Its event-centered design supports replay, simulation, learning, and evaluation against non-learning baselines.

F. Advisor Reasoning Under Bluff Pressure: The advisor combines hidden-role probabilities with rule-based impossibility checks, presenting uncertainty separately from certainty so recommendations remain auditable under pressure.The probability estimate incorporates claim context and unaccounted copies of the claimed role.
F. Advisor Reasoning Under Bluff Pressure: Conservative, Aggressive, and Balanced modes shift the challenge threshold by -0.05, +0.07, and 0, respectively, while preserving shared beliefs and legality constraints.Early symmetric states can yield nearly uniform beliefs because information is genuinely insufficient.
G. Risk-Asymmetric Behavior in the Endgame: With one influence remaining, players become risk-asymmetric, accepting unlikely bluffs or challenges when failure would otherwise cause near-certain elimination.This behavior improves simulated-opponent realism and better matches observed human play.
H. Thompson Sampling Extension: Thompson sampling would explore according to posterior uncertainty over latent quantities, unlike uniform ε-greedy or value-spread-based Boltzmann exploration, but requires explicit posteriors.The proposed extension samples latent role assignments and evaluates actions under sampled worlds.
I. Decision Layers and Their Current Character: CoupVisor’s belief layer performs deterministic posterior reasoning, its advisor layer exploits current beliefs for immediate expected value, and only the learning layer explores.The learning layer uses ε-greedy or Boltzmann action selection.
J. Relation to Poker-Style Imperfect Information: Unlike poker, Coup makes role claims and challenges explicit and frequent, while coin constraints such as forced Coup at ten coins create stronger hard legal transitions.Both games nevertheless require probability estimates, bluff-risk handling, and mixed strategies.
K. Single-Turn Interaction; M. Training and Evaluation Pipeline: The event schema represents a turn’s branching interaction and serves as the operating unit for both replay and simulation, supporting the training pipeline for BC and DQN policies.Simulator transitions are benchmarked against a random policy and non-learning agents.
L. Worked Example; N. Non-Learning Baseline Agents: In the worked example, successive Duke and Assassin claims update role beliefs, account for deck composition and claim-history consistency, and trigger recomputed challenge or block recommendations.The initial four-player state has uniform role belief at 0.20 per role, and the first claim moves Player A to five coins.

O. User Interaction Flow … B. Bot-Mix Sensitivity

CoupVisor presents a unified, event-driven interaction and evaluation framework for advisor and policy assessment. Results show stable outcome accuracy but strongly opponent-dependent challenge performance.

O. User Interaction Flow: Each turn branches through actions, blocks, and challenges, with every claim or reveal triggering belief updates and fresh advisor recommendations.Challenges resolve before the action or counteraction they target, creating multiple decision points within one turn.
A. Evaluation Run: The evaluation fixes configuration, seeds, and outputs before analysis, ensuring all tables, ablations, and figures describe the same games.A single evaluation run underlies the paper’s reported numbers and plots.
B. Primary Design Matrix: The balanced design assigns each of five seeds 250 games per opponent condition, totaling 500 games.Equal game counts make the opponent conditions directly comparable.
O. User Interaction Flow: Users load or enter games, inspect reconstructed public states and role beliefs, and receive recommendations alongside separate validation warnings.The interface is deliberately shallow and hides the event schema from nontechnical users.
C. Metrics: Advisor quality uses challenge precision, recall, F1, outcome accuracy, and outcome score, while policy quality uses episode reward, step reward, and action-match rate.The same evaluation harness computes policy metrics for learned policies and non-learning baselines.
D. Exploration-Mode Ablation: The DQN ablation compares ε-greedy and Boltzmann softmax action selection under identical seeds, architecture, replay buffer, and reward function.It records episode reward, action-distribution entropy in nats, and action-visitation histograms to characterize realised exploration.
E. Evaluation Beyond Random Comparison: Evaluation extends beyond random comparison through fixed decision scenarios, opponent-type cross-play, and strategic-soundness measures including Brier score and one-step regret.Scenarios isolate conflicts such as early Tax claims, Steal-block challenges, Assassinate-Contessa branches, forced Coup states, and one-influence endgames.
A. Advisor Multi-Seed Results; B. Bot-Mix Sensitivity: 0.5694 ± 0.0242 outcome accuracy is stable across seeds, whereas challenge F1 ranges from 0.336 to 0.511 and varies with opponent composition.Against aggressive-heavy opponents, challenge F1 is 0.5641 versus 0.2831 against honest-heavy opponents; outcome accuracy is 0.5619 versus 0.5769, and weighted ranking is 0.5628 versus 0.4594.

C. Ablation: Belief Mode … B. How This Research Helps

CoupVisor’s ablations find no reliable effect from strict duplicate-hand belief updating, while advisor-off comparisons require caution because most simulated claims are truthful. The broader results show that reward design determines learned-policy rankings, and that interpretable, unified architecture supports practical decision advice and transferable offline learning lessons.

C. Ablation: Belief Mode: Strict-minus-Standard belief-mode differences are not claimed because all three deltas are small, negative, and have 95% confidence intervals crossing zero.The duplicate-hand constraint rarely binds at important decision points because other evidence usually dominates once a player holds one hidden influence.
D. Ablation: Advisor On Versus Off: Advisor-off drives challenge F1 to zero by construction but raises outcome accuracy to 0.8192 and outcome score to 0.6424, since most simulated claims are true.This does not show that challenging is harmful; it shows that never challenging is rarely punished in this simulator.
E. Significance Notes: Advisor-on-minus-off intervals exclude zero, whereas every Strict-minus-Standard interval crosses zero, separating statistically detectable advisor differences from unsupported belief-mode differences.The advisor comparison still requires the interpretive caveat that never challenging benefits when claims are usually truthful.
F. Policy Benchmark: Under the original shaped reward, behavior cloning reaches +1.04 average reward per episode versus −0.29 for DQN and −0.60 for random.The shaping combines coin change, influence change, and an action-match bonus; DQN’s reward oscillates between roughly −6 and +4 over 100 episodes without an upward trend.
F. Policy Benchmark: Under win-aligned reward, DQN leads at 6.74 reward per episode and 0.373 winner-action match, ahead of belief-EV heuristic at 3.68, behavior cloning at 3.38, and random at 1.32.The reward curve rises from a rolling mean near 1.5 at episode 20 to roughly 5.9 by episode 1000.
G. Result Figures / H. Exploration-Mode Comparison: Figures 8–12 visualize the reported results, while Fig. 13 shows that matching scalar exploration schedules does not reproduce the exploration actually performed.The supplied figure descriptions identify broad graphical confirmation and an exploration-mode contrast rather than additional numerical findings.
VII. Discussion / A. What the Results Say: The deterministic threshold advisor favors interpretability but remains theoretically exploitable, motivating a stochastic advisor through exploration ablations and Thompson sampling.This discussion characterizes deterministic advice as an exploitation-end design choice in tension with mixed-strategy requirements for imperfect-information games.
B. How This Research Helps: CoupVisor helps players by exposing a stated probability, supporting evidence, and hard impossibilities, while one event schema keeps manual play, replay, simulation, belief tracking, advice, and training aligned.Methodologically, behavior cloning wins under shaped reward but the same DQN architecture wins every baseline under win-aligned reward, demonstrating the importance of reward choice for heterogeneous logged behavior.

C. Error Analysis … X. Future Work

CoupVisor’s dominant errors are over-eager challenges, while partial observability, offline learning, scripted opponents, outdated benchmark parameters, and myopic decisions limit interpretation. Future work targets belief calibration, posterior-based advising, standardized evaluation, and constrained self-play.

C. Error Analysis: 35.3% of 2,670 challenged claims were true claims the advisor incorrectly recommended challenging, compared with 3.6% false claims it missed.These 942 false challenges versus 96 missed challenges quantify the system’s precision-recall imbalance.
C. Error Analysis: False challenges concentrate on Steal (467 of 942) and Tax (299 of 942), while missed challenges are mostly Steal (62 of 96).The error pattern indicates that the advisor challenges particularly eagerly for Steal and Tax.
C. Error Analysis: Cases E2 and E3 fall within 0.0003 of their threshold, whereas E1 and E4 produce posteriors of 10^-34 from unbounded repeated multiplicative updates.A margin band could recover E2 and E3; capping or recalibrating repeated-claim evidence addresses E1 and E4, while E5 reflects a hard-rule violation.
VIII. Challenges and Limitations: Partial observability keeps early and symmetric beliefs close across players, while multi-stage action-block-challenge-reveal interactions make event consistency difficult.Event-ordering errors can propagate into both belief and advice across replay, simulation, and the interface.
VIII. Challenges and Limitations: Offline BC and DQN training is vulnerable to distribution shift and extrapolation error in rare, high-impact states absent from logged trajectories.Scripted honest, bluffer, aggressive, and cautious opponents enable controlled experiments but omit adaptive human deception, long-horizon signaling, and repeated-match meta-game adjustment.
VIII. Challenges and Limitations: The benchmark used an earlier belief-multiplier and threshold parameterization, so current-configuration re-evaluation remains outstanding despite over-eager challenging likely persisting.The reported error magnitudes may change because the current likelihood ratios are larger and concentrate belief faster.
VIII. Challenges and Limitations: The thresholded posterior rule is myopic, and simulation-only evaluation lacks human-in-the-loop validation, cross-implementation comparison, and stratification by player counts or table dynamics.Retaliation risk, information signaling, and endgame tempo are unmodeled sources of long-term game value.
X. Future Work: Future work prioritizes calibrated uncertainty output, Thompson sampling, deterministic scenario benchmarks, and constrained self-play while retaining legality-rate and forced-Coup compliance checks.The proposed calibration methods include isotonic or temperature scaling and reliability diagrams; scenario coverage includes Tax, Steal, Assassinate-Contessa, forced-Coup, and one-influence endgames.

Appendix A Belief Update and Challenge Rule … A.3 Deck pressure

Appendix A specifies the implementation of CoupVisor’s belief updates and challenge rule through likelihood ratios, event-based evidence, and deck-pressure damping. The deck-pressure mechanism reduces role beliefs according to unrevealed copies, with sublinear damping and zero probability when no copies remain.

Appendix A Belief Update and Challenge Rule: The appendix provides the code underlying every belief and threshold constant quoted in the paper.
A.1 Likelihood ratios: Claim likelihood ratios compare the probability of making a claim while holding its role against making it without that role.An honest claim rate near 0.85 versus a bluff rate near 0.15 yields 5.67.
A.1 Likelihood ratios: The implementation assigns 5.7 likelihood ratios to Tax, Assassinate, Steal, and Exchange claims for their implicated roles.
A.1 Likelihood ratios: Foreign Aid and Assassination each use a 5.7 ratio, while Steal blocks use 4.5 ratios for both Captain and Ambassador.Steal blocks are lower because either of two roles can block a Steal.
A.2 Applying evidence: Each event type multiplies the score of the role it implicates, with won and lost challenges providing near-certain confirmation or disconfirmation.The implementation applies action multipliers to the actor’s implicated role and challenge multipliers to the challenged claim.
A.2 Applying evidence: Block evidence is applied through the blocked action’s configured role ratios, and challenge outcomes select separate win or loss scaling factors.
A.3 Deck pressure: Each role is damped by its remaining fraction of three unrevealed copies, raised to the sublinear power 0.7; a role with no copies left becomes zero.The implementation computes remaining = max(0, 3 - revealed) and damp = (remaining / 3.0) ^ 0.7 when copies remain.

A.4 Per-hand probability · A.5 Threshold rule

A.4 defines per-hand truth probability by blending a hypergeometric calculation with an independence estimate, while A.5 applies context- and style-dependent thresholds for challenge decisions. The probability blend uses fixed 0.7 and 0.3 weights, and the threshold implementation includes claim-context bases and table-state adjustments.

A.4 Per-hand probability: The blend weights are 0.7 for combinatorial correctness and 0.3 for responsiveness to accumulated claim evidence.These weights are explicitly described as trading the two objectives.
A.4 Per-hand probability: The probability routine normalizes remaining copies, hand size, and deck size before computing the estimate.It clamps remaining_copies to nonnegative values, hand_size to at least 1, and deck_size to at least hand_size.
A.4 Per-hand probability: The hypergeometric component calculates the chance of at least one claimed role by subtracting the zero-copy probability from 1.The zero-copy probability uses combinations of non-copies and total deck cards.
A.4 Per-hand probability: The per-hand truth estimate blends a hypergeometric probability with an independence-based probability.The implementation computes p_hyper and p_independence before combining them.
A.5 Threshold rule: The threshold rule defines base values according to claim context, including Steal, Assassinate, and Assassination cases.The cited implementation passages show context-specific base assignments, including base = 0.99 and base = 0.35.
A.5 Threshold rule: Thresholds incorporate style deltas, with Conservative and Aggressive styles receiving distinct adjustments.The implementation normalizes the style before applying the corresponding delta.
A.5 Threshold rule: The rule also includes three adjustments based on public table state.The body’s threshold constants are said to include three table-state adjustments, alongside base values and style deltas.

A.6 Decision · Appendix B Reward Definitions · B.1 Per-transition reward

The decision module recommends whether to challenge by comparing estimated truthfulness with a threshold and explains the inputs behind that comparison. Appendix B defines shaped and win-aligned rewards, with reward design determining which learned policy performs best.

A.6 Decision: The recommendation combines the comparison with an explanation string containing the probability, threshold, and inputs that produced them.
A.6 Decision: The decision logic recommends “Challenge” when estimated truthfulness falls below the configured threshold.
Appendix B Reward Definitions: The paper identifies reward definition, rather than algorithm choice, as the factor deciding which learned policy wins.Appendix B presents both reward definitions.
B.1 Per-transition reward: The shaped reward equals the acting player’s coin gain plus the number of players eliminated on that transition.
B.1 Per-transition reward: Under win-aligned mode, the per-step reward is zeroed and each transition records whether the acting player ultimately wins.
B.1 Per-transition reward: The win-aligned transition label is determined by whether the acting player is the game’s sole remaining player.

B.2 Environment step … C.2 Defaults

CoupVisor’s environment distinguishes shaped imitation rewards from winner-focused rewards, while its learning setup uses a two-hidden-layer ReLU network and fixed training defaults. The documented implementation also specifies action matching, terminal-state handling, and exploration parameters.

B.2 Environment step: Shaped mode adds +1 when the chosen action matches the logged action, rewarding imitation of every logged player.The reward begins with the logged row reward in shaped mode.
B.2 Environment step: Win mode adds +1 only when the chosen action matches the logged action and that actor later wins.Because its per-step term is zero, the win reward is exactly this indicator.
B.2 Environment step: Each environment step selects an action from the configured labels, reads the logged action, and advances to a next state while recording termination.Termination is derived from the row’s done flag.
C.1 Network: The action-value function is a two-hidden-layer multilayer perceptron with ReLU activations and hidden width 128.Its implementation uses three linear layers separated by two ReLU activations.
C.1 Network: The network maps an input dimension to an output dimension through two hidden layers of width 128 before producing action values.The implementation exposes in_dim, out_dim, and hid_dim in the constructor.
C.2 Defaults: Training fixes the paper’s quoted hyperparameters, including optimization, replay, warm-up, target-network, random-seed, and exploration settings.The entry point also fixes both exploration schedules.
C.2 Defaults: The defaults include 200 episodes, seed 42, gamma 0.98, learning_rate 1e-3, batch_size 64, buffer_size 10000, and warmup_steps 256.They also specify target_update_steps 200, hidden_dim 128, epsilon decay, and shaped rewards by default.

C.3 Optimiser … C.6 Summary

The optimiser uses Adam, while action selection supports ε-greedy and Boltzmann exploration with entropy tracking. Training updates rely on replayed transitions, Bellman targets, periodic target-network synchronization, and comparisons summarized in Table XIII.

C.3 Optimiser: Adam applies the learning rate to the policy network parameters.
C.4 Action selection and entropy: Both exploration modes run in the same loop and compute per-step entropy for their respective action distributions.ε-greedy entropy is derived from ε and the action count, whereas Boltzmann entropy uses softmax probabilities reflecting learned action-value spread.
C.4 Action selection and entropy: ε-greedy selects a random action with probability ε and otherwise uses the policy action.Its greedy-action probability is 1.0 - epsilon + epsilon / action_dim, with the remaining probability assigned across other actions.
C.4 Action selection and entropy: Boltzmann exploration samples actions from probabilities obtained by temperature-scaled, numerically stabilized softmax values.The implementation computes entropy as -np.sum(probs * np.log(probs + 1e-12)).
C.6 Summary: The target network is copied from the policy network at a fixed step interval, while Table XIII summarizes the exploration ablation and supervised comparisons.The behavior-cloning baseline uses logistic regression over the same feature vector, and the supervised comparison uses a Random Forest.
C.5 Replay, Bellman target, and target sync: Training updates begin only when replay contains at least both the batch size and warm-up threshold.Transitions are stored with state, action, reward, next state, and terminal status before sampling.
C.5 Replay, Bellman target, and target sync: The target uses a one-step Bellman backup, masks bootstrapping on terminal transitions, and trains the policy network with the resulting loss.The target is r_batch + gamma * (1.0 - d_batch) * max_next_q, using the target network’s maximum next-state value.

Appendix D The Game of Coup … D.8 Relation to the Model

Appendix D states Coup’s published rules and connects them to CoupVisor’s model, simulator, and interface. The model represents character-based beliefs while enforcing costs, forced Coup legality, and branching blocking structure.

Appendix D The Game of Coup: The appendix summarizes Coup’s rules so the paper’s modelling choices can be checked against the described game.
D.1 Setup and Goal: Each player starts with two hidden influence cards and two coins; the goal is to eliminate every opponent and remain the last survivor.
D.2 Influence: Influence cards remain face down until lost, revealed cards no longer provide influence, and a player losing both cards is exiled.
D.3 Turn Structure: Play proceeds clockwise with exactly one action per turn; challenges and counteractions occur after declaration, before the targeted action, and cannot be issued retroactively.
D.4 Actions: Players may choose any affordable action, while holding ten or more coins forces a Coup that automatically succeeds and cannot be blocked or challenged.
D.5 Counteractions: Character actions and counteractions require claims that may be truthful or false, succeed automatically when unchallenged, and can cause an action to fail when successfully counteracted.
D.6 Challenges: Any player may challenge a character claim; the claimant must show the relevant card, and the loser immediately loses an influence.
D.7 Double Loss on Assassination: The simulator omits the published replacement-card swap after a successful challenge, while the rules allow two influence losses in one turn after a failed challenge and successful assassination.