PaperlayerView exact revision on arXiv

Source-linked AI summary

FraudSociety: A Generative Agent-Based Society for the Financial-Fraud Economy

Keyang Chen, Mingxuan Jiang, Zeping Li, Yiwen Zhao, Yifan Wang, Qiuyang Zhang, Shuaicheng Qi, Yicun Liu, Zhixin Li, Sen Liu, Yinan Jing, Guangnan Ye, Hongfeng Chai

arXiv:2610.33536v1cs.AIcs.CY

TL;DR

Existing fraud simulators typically capture either informational or financial dynamics, despite fraud involving interactions between both. FraudSociety couples these dynamics in an LLM-agent simulator and evaluates macro outcomes, victimization, and interventions; its reported results are consistent across these levels, with the top 10% of victimization records accounting for 81.1% of total loss.

  • Problem

    Existing simulators typically capture only one side of the interaction between informational and financial dynamics, while understanding fraud requires modeling how they interact at individual and population levels.

    Source: §1 · p. 1

  • Method

    FraudSociety couples a dynamic social environment, persistent financial ledgers, and Need-Intent-Action civilian decisions, with fraudsters adapting actions to victim histories and financial states.

    Source: §1 · p. 2

  • Results

    FraudSociety reports consistent macro, micro, and intervention results; the top 10% of victimization records account for 81.1% of total loss, broadly consistent with CAFC statistics.

    Source: §1 · p. 2

  • Takeaways & Limitations

    FraudSociety provides a testbed for systematically evaluating counter-fraud interventions on payment behavior and financial loss.

    Source: §1 · pp. 2–3

Abstract

from arXiv · show

Financial fraud caused over \$15.9 billion in losses in the United States in 2025, according to the Federal Trade Commission~\citep{ftc2026databook}. Understanding such large-scale harm requires jointly modeling informational and financial dynamics. Yet existing simulators typically capture only one side of this interaction, leaving a gap in modeling how fraud unfolds in the real world. To bridge this gap, we introduce \textsc{FraudSociety}, the first LLM-agent-based simulator that jointly models these coupled dynamics. It integrates a dynamic social environment for information exposure and communication, persistent financial ledgers that preserve payment and loss histories, and a Need-Intent-Action architecture for civilian decision making. Fraudsters further adapt their communication and financial actions to individual interaction histories and ledger states, forming a closed feedback loop between the two dynamics. We validate \textsc{FraudSociety} across macro-level outcomes, micro-level behavior, and intervention effects. At the macro level, the simulated fraud outcomes closely match real-world patterns across multiple key statistics. At the micro level, it reproduces strong loss concentration, with the top 10\% of victimization records accounting for 81.1\% of total losses, consistent with Canadian Anti-Fraud Centre statistics~\citep{cafc2025annual}. Counter-fraud experiments further reveal variation across intervention strategies, with the best-performing personalized intervention reducing the repeat-payment rate from 49.4\% to 10.9\%. These results establish \textsc{FraudSociety} as both a high-fidelity simulator and a scalable testbed for evaluating anti-fraud strategies under diverse intervention conditions, thereby advancing research on financial fraud and its prevention.

1 INTRODUCTION

Financial fraud arises from coupled information and financial dynamics, yet existing simulators typically model only one side. FraudSociety joins these dynamics in an LLM-agent simulator and evaluates its macro-level fidelity, victimization pathways, and intervention effects.

1 INTRODUCTION

  • In 2025, the FTC received 3 million fraud reports with losses exceeding $15.9 billion, amid feedback between information exposure, payment decisions, and later communication.Fraudulent information can shape attention, trust, and willingness to pay; payments and withdrawal attempts can in turn change subsequent fraudster communication.Source: §1 · p. 1
  • Studying fraud requires modeling how informational and financial dynamics interact at individual and population levels, but existing approaches typically capture only one side.Traditional fraud ABMs emphasize transactions, while LLM-based ABMs emphasize communication and rarely link it to persistent financial states.Source: §1 · pp. 1–2
  • FraudSociety couples a dynamic social graph, persistent payment and loss ledgers, and civilian Need-Intent-Action decision making in a feedback loop.Fraudsters adapt communication and financial actions to each victim’s response history and evolving financial state.Source: §1 · p. 2
  • Across macro, micro, and intervention levels, FraudSociety reports consistent results; its best-performing personalized intervention reduces repeat-payment rate from 49.4% to 10.9%.The macro-level Quartet Distance is 0.2978, and the top 10% of victimization records account for 81.1% of total loss, broadly consistent with CAFC statistics.Source: §1 · p. 2
  • The paper contributes cross-level validation of macro fidelity and micro victimization mechanisms, and presents FraudSociety as a testbed for counter-fraud strategies.The testbed supports systematic evaluation of intervention effects on payment behavior and financial loss.Source: §1 · pp. 2–3

2 RELATED WORK

Prior fraud simulators emphasize either transaction and capital-flow patterns or fraud-related communication. FraudSociety addresses their stated limitations by jointly modeling information and persistent financial states and evaluating both individual and aggregate outcomes.

Agent-based social simulation

  • General LLM-agent social simulators support large-scale social environments and adaptive behavior, but primarily target general social or market dynamics rather than financial-fraud victimization.The related work includes Generative Agents, OASIS, AgentSociety, and SocioVerse.Source: §2 · p. 3

Traditional simulation for financial fraud

  • Traditional fraud simulators provide persistent transaction records for detection and graph-learning benchmarks, but generally generate transactions from predefined rules without preceding information exposure or adaptive persuasion.Examples include BankSim, AMLSim, SAML-D, and TransXion.Source: §2 · p. 3

LLM-agent simulation for financial fraud

  • LLM-agent fraud systems model adaptive communication, but do not generally couple it to persistent financial states or assess whether individual pathways reproduce aggregate loss patterns.ScamAgents studies multiturn scam calls; MultiAgent4Fraud studies collaborative fraud across public posts and private interactions.Source: §2 · p. 3

3 FRAUDSOCIETY

FraudSociety combines civilian and fraudster agents with social and financial environments to simulate fraud as an evolving process. Its decision cycles and persistent states connect individual interactions to subsequent behavior.

3.1 OVERALL DESIGN OF FRAUDSOCIETY

  • The four-part simulation combines civilians, fraudsters, a Social World, and a Financial World; hourly cycles update states, gather observations, execute actions, and commit changes.It is initialized with real-world post seeds, empirical civilian profiles, and fraud-campaign templates.Source: §3.1 · pp. 3–4, Figure 2 · p. 4

3.2 CIVILIAN MODELING WITH NEED-INTENT-ACTION

  • Civilians use a Need-Intent-Action cycle in which evolving needs shape activity intents, which restrict candidate actions; action outcomes inform later decisions.Needs evolve with profiles, routines, and social and economic signals, capturing heterogeneous priorities across civilians.Source: §3.2 · pp. 4–5

3.3 FRAUDSTER MODELING

  • Fraudsters manage campaigns by fraud type and phase, choosing daily phase transitions and hourly victim-specific social and financial actions based on interaction history.Campaign phases progress through recruiting, credibility, peak, and collapse; hourly actions include exposure, contact, returns, withdrawal handling, and further payment requests.Source: §3.3 · p. 5, Figure 3 · p. 5

3.4 SOCIAL AND ECONOMIC ENVIRONMENT

  • The Social World models a dynamic graph and public or private exposure, while the Financial World tracks routines, balances, payments, returns, withdrawals, delays, and unrecovered losses.Public feeds combine globally ranked posts with followed and previously interacted authors, then rerank candidates by similarity to a civilian’s recent interactions; ledger states feed into subsequent observations.Source: §3.4 · pp. 5–6

4 EXPERIMENTS

The experiments validate FraudSociety across aggregate loss patterns, victimization pathways, and intervention effects. The supplied setup describes controlled simulation conditions and a regulator-grounded macro-level audit score.

4 EXPERIMENTS

  • The experiments form a cross-level validation chain: macro loss distributions, micro victimization pathways and mechanisms, then warning effects on payment behavior and financial loss.The same simulated environment is used for the three levels of analysis.Source: §4 · p. 6
  • The simulation uses 112 civilians, 8 fraudsters, 8 active campaigns, and 480 hourly steps, with variants and intervention conditions sharing initialization and random seed.Qwen3-14B is used for all agent decisions.Source: §4 · p. 6

4.1 MACRO-LEVEL VALIDATION

  • Quartet Distance audits macro realism against four FTC stylized facts: heavy-tailed loss concentration, age-stratified severity, fraud-family severity hierarchy, and age-specific loss-report shares.The supplied passage defines QD’s components but does not state the experiment’s resulting score.Source: §4.1 · p. 6

N (dHTL + dALS + dFSH + dALR) ,

FRAUDSOCIETY aligns closely with regulator-grounded loss patterns and reproduces micro-level variation associated with needs, balances, and repeated transfers. Its intervention experiments show that strategies differ in how they affect initial and repeat payments, victimization, and losses.

N (dHTL + dALS + dFSH + dALR) ,

  • QD = 0.2978, the lowest among compared systems, with smaller discrepancies across loss concentration, age-conditioned severity, fraud-family severity, and victim shares.Unlike TransXion's financial-only scope and MultiAgent4Fraud's terminal first transfer, FRAUDSOCIETY models informational and financial dynamics throughout victimization.Source: Table 2 · p. 7, N (dHTL + dALS + dFSH + dALR) · p. 7

4.2 MICRO-LEVEL ANALYSIS

  • 81.1% of total loss came from the top 10% of victimization records, broadly consistent with CAFC's 75.2%; the figure compares pathway concentration and balance-group loss severity.Records with at least five transfers were 19.8% of records but accounted for 75.6% of losses; all 63 positive-loss victims had campaign contact and private-DM exposure before their first transfer.Source: §4.2 · pp. 7–8, Figure 4 · p. 8
  • Safety and social need values were higher during transfer hours than overall, and transfer rates rose from low- to high-need states.Safety need increased from 0.598 to 0.710 and social need from 0.630 to 0.712; corresponding transfer rates rose from 0.94 to 1.94 and 1.03 to 1.91.Source: §4.2 · p. 8, Table 3 · p. 8
  • Low-balance civilians were more likely to incur positive losses, while affected high-balance civilians had larger conditional losses.The figure shows affected civilians as points and each initial-balance group's mean loss as markers.Source: §4.2 · p. 8, Figure 4 · p. 8
  • Both removing Need-Intent-Action and disabling Financial Ledger increased QD, indicating weaker alignment with regulator-grounded stylized facts.The ablations remove need-conditioned decisions or hide ledger states and feedback such as income and cumulative investment.Source: §4.2 · pp. 8–9, Table 3 · p. 8

4.3 COUNTER-FRAUD INTERVENTION ANALYSIS

  • Broadcasts reduced payments from 623 to 205 and repeat payment from 49.4% to 38.2%; monitor labels reduced payments to 284 and random comments to 141, while repeat payment remained high.Broadcasts raised withdrawal attempts from 8.6% to 17.1%; labels act at exposure, while random comments reduced total loss to $42,014 but left repeat payment at 49.1%.Source: §4.3 · p. 9
  • 44 positive-loss victims under peer warnings, down from 63, accompanied total loss of $24,752 and maximum positive loss of $3,800.The authors report that observation-conditioned warnings compressed high-loss pathways more effectively than random comments.Source: §4.3 · p. 9
  • 117 payments and 10.9% repeat payment under Oracle personalized warnings, compared with 623 payments and 49.4% repeat payment in the no-warning reference.This idealized warning triggers after a first fraud payment; withdrawal attempts rose to 47.5%, and the experiments provide a controlled testbed for comparing intervention effects on victimization trajectories and losses.Source: §4.3 · p. 9

5 CONCLUSION

FraudSociety jointly models informational and financial fraud dynamics, reproducing macro-level loss patterns and micro-level victimization pathways. The conclusion presents it as a scalable simulator and testbed, while the appendices document implementation and evaluation details.

5 CONCLUSION

  • FraudSociety jointly models fraud’s informational and financial dynamics, with experiments reproducing regulator-grounded macro patterns and micro-level victimization pathways.Ablations show that Need-Intent-Action and ledger feedback connect individual processes with aggregate patterns.Source: §5 · pp. 9–10
  • The simulator supports controlled comparisons of counter-fraud interventions and their downstream effects.The conclusion describes it as a scalable foundation for realistic financial-fraud simulation and systematic counter-fraud research.Source: §5 · p. 10

REPRODUCIBILITY STATEMENT

  • The appendix specifies simulation timing, population construction, model and intervention settings, evaluation units, artifact schemas, runtime prompts, and run provenance..Source: REPRODUCIBILITY STATEMENT · p. 10

A GUIDE TO THE APPENDIX

  • The appendix guide links deferred implementation topics to detailed treatments, including formal definitions, diagnostics, and translated runtime prompts..Source: A GUIDE TO THE APPENDIX · p. 12

B IMPLEMENTATION AND EXPERIMENTAL SETTING

The experiments use a fixed population and common simulation settings to compare fraud and intervention conditions. The implementation grounds agent profiles, social interaction, campaigns, and financial activity in explicit model components, while reported contrasts remain descriptive single-run comparisons.

Population and temporal setup

  • All runs comprise 480 hourly steps over 20 simulated days, using seed 42 and a fixed profile realization.The full profile pool contains 112 civilian and 8 criminal-operator profiles.Source: B IMPLEMENTATION AND EXPERIMENTAL… · p. 12

LLM backbone

  • All reported LLM-controlled decisions use Qwen3-14B, including the baseline, interventions, and ablations.The common backbone avoids confounding these comparisons with changes in model family.Source: B IMPLEMENTATION AND EXPERIMENTAL… · p. 12

Grounded civilian profiles and routine context

Financial and online-post grounding

  • Initial balances draw on age-specific Federal Reserve transaction-account anchors, while online seed posts provide topical grounding rather than fixed scripts.Operators generate posts and messages conditioned on fraud family, campaign phase, recent engagement, and visible user signals.Source: B IMPLEMENTATION AND EXPERIMENTAL… · p. 14

Fraud campaign initialization

  • Four campaign families—investment, impersonation, marketplace, and romance—progress through recruiting, credibility building, peak extraction, and collapse.The eight-operator setting assigns two operators to each family.Source: B IMPLEMENTATION AND EXPERIMENTAL… · p. 14

Social platform and social graph

  • The social environment combines an OASIS-style platform with a hybrid small-world graph, a Reddit-style hot-score recommender, and labeled social ties.The graph starts with six geographically nearest civilian neighbors, uses rewiring probability 0.15, and can include up to ten operator targets.Source: B IMPLEMENTATION AND EXPERIMENTAL… · p. 14

Counter-fraud conditions

Run artifacts and comparative interpretation

  • Intervention contrasts are single-run descriptive comparisons, not estimates of population-level effects across independent seeds.Runs share the same population realization, simulator seed, start time, and horizon; artifacts link interactions, transactions, victim-campaign states, statistics, and decision traces.Source: B IMPLEMENTATION AND EXPERIMENTAL… · p. 15

COUNTER-FRAUD INTERVENTION REALIZATION

Interventions act through information that enters civilians’ bounded social observations, and their effects depend on exposure and endogenous response. Each hourly cycle commits admissible social and financial actions to shared state so later observations reflect their consequences.

COUNTER-FRAUD INTERVENTION REALIZATION

  • C1 and C3 insert social records before civilian steps, C2 annotates committed posts for later readers, C4 uses ordinary civilian actions, and C5 follows a committed fraud payment.C5 uses transaction identifiers for deduplication, allowing a later payment to trigger a distinct warning; C2 fails open when classifier output is malformed or unavailable.Source: COUNTER-FRAUD INTERVENTION REALIZATION · p. 15

C HOURLY SIMULATION CYCLE AND CLOSED-LOOP COUPLING

  • Figure 6 places agent decisions after hourly state updates and commits admissible social and financial effects before the next tick.Agents observe bounded, role-specific projections rather than the full simulator state; memory updates only from grounded execution results.Source: Figure 6 · p. 16, C HOURLY SIMULATION CYCLE AND… · p. 16

COUNTER-FRAUD INTERVENTION REALIZATION

  • Warnings affect behavior only when they become visible in a bounded civilian view; insertion alone does not ensure reading or compliance.Private messages, feed labels, and comments compete for limited attention, and civilians may respond, defer, ignore, forward, or act against warnings.Source: COUNTER-FRAUD INTERVENTION REALIZATION · pp. 15–16

D FRAUDSTER HIERARCHICAL DECISION PROCESS

Fraudster decisions are organized into lifecycle, financial, and social policies with distinct timing and scope. Deterministic execution constraints project proposals onto admissible actions before they become persistent state.

D FRAUDSTER HIERARCHICAL DECISION PROCESS

  • Campaign lifecycle decisions occur at most once per simulated date, and execution enforces monotone phase progression with collapse terminal.Entry into peak is rejected unless the campaign has a victim or positive raised funds; proposed retreats are rejected.Source: D FRAUDSTER HIERARCHICAL DECISION… · p. 16
  • The hierarchy separates strategic proposals from persistent changes through deterministic execution gates.The lifecycle layer constrains available strategies, and the financial layer reacts to paid-victim and withdrawal states.Source: D FRAUDSTER HIERARCHICAL DECISION… · pp. 16–17, Table 11 · p. 18
  • The financial policy runs only during credibility and peak, reusing decisions for four hours or two after recent payment, return, or withdrawal events.A changed context signature forces replanning.Source: D FRAUDSTER HIERARCHICAL DECISION… · p. 17
  • Step 2A handles public visibility and feed-grounded outreach, while Step 2B replies only in active dialogues where the civilian spoke last.Step 2A cannot message active conversations; Step 2B has no public-feed candidates and allows at most two follow-up rounds per hourly tick.Source: D FRAUDSTER HIERARCHICAL DECISION… · pp. 17–18, Figure 7 · p. 18

E NEED-INTENT-ACTION FORMALIZATION

The hourly NIA cycle updates civilian needs from observed context and memory, then uses those needs and feasible routes to structure actions. The appendix also reports daily and within-day patterns in route choices and need urgency.

E NEED-INTENT-ACTION FORMALIZATION

  • When needs cross their activation thresholds, the active need follows the satiety, energy, safety, then social priority order; otherwise, it is neutral.Ties within a priority level are resolved by greater urgency.Source: E NEED-INTENT-ACTION FORMALIZATION · p. 18
  • At each hourly tick, the need updater uses bounded observations, recent memory, and the previous route; omitted response fields remain unchanged, and invalid responses retain the full vector.The updater applies schema validation and range clipping.Source: E NEED-INTENT-ACTION FORMALIZATION · pp. 18–19
  • The intent policy selects a route feasible under time, work, sleep, mobility, and social availability, and the route restricts the available action schema.For example, social routes allow communication actions, while consumption routes expose nearby merchant choices.Source: E NEED-INTENT-ACTION FORMALIZATION · p. 19
  • Each hourly cycle synchronizes salaries, expenses, returns, and payouts; resolves sleep and mobility; and refreshes feeds, messages, comments, and campaign contacts.The cycle proceeds through these operations before civilian planning.Source: E NEED-INTENT-ACTION FORMALIZATION · p. 19

F OBSERVATION AND STATE SCHEMAS

E NEED-INTENT-ACTION FORMALIZATION

  • In the no-warning baseline, routine rest/work and sleep/no-route dominate daily decisions, while social-media and consumption routes have smaller but persistent shares; social need urgency is highest through most of the horizon.Figure 8 summarizes the 20-day baseline.Source: E NEED-INTENT-ACTION FORMALIZATION · p. 19, Figure 8 · p. 20

F OBSERVATION AND STATE SCHEMAS

  • The campaign ledger tracks victim- and campaign-level pathway information, including gross payments, returns, payment counts, and withdrawal state; requests do not count as returns until payouts execute.Append-only event records preserve provenance across messages, campaigns, agents, transactions, and hourly steps.Source: F OBSERVATION AND STATE SCHEMAS · pp. 20–21

E NEED-INTENT-ACTION FORMALIZATION

  • Across the 20-day average, rest/work dominates daytime, sleep/no-route dominates late-day rest, consumption clusters in short intervals, and social-media activity rises late in the day.Energy, safety, and social urgency rise during active hours and decline after the nighttime transition.Source: E NEED-INTENT-ACTION FORMALIZATION · p. 19, Figure 9 · p. 21

G EVALUATION UNITS AND REPORTED OUTCOMES

The evaluation distinguishes gross payments, returns, positive net loss, and payment behavior across victim–campaign records. Its intervention comparisons show that payment frequency alone does not capture the severity or persistence of remaining loss pathways.

G EVALUATION UNITS AND REPORTED OUTCOMES

  • Micro-level pathway and intervention outcomes use positive-net-loss victim–campaign records, while unique-victim statistics deduplicate by victim and Quartet Distance aggregates gross payments once per unique civilian.This distinction keeps simulation-specific returns from changing the regulator-grounded Quartet Distance assessment while retaining returns and withdrawals as behavioral outcomes.Source: G EVALUATION UNITS AND REPORTED OUTCOMES · p. 21
  • 33.7%, 51.6%, 71.0%, 82.9%, and 87.6% decreases in positive net loss under C1–C5 contrast with payment-count decreases of 67.1%, 54.4%, 77.4%, 75.1%, and 81.2%, respectively.C3 records 141 payments versus C4’s 155, yet C3’s $42,014.25 positive net loss is 69.7% higher; C4 lowers loss by a further 41.1%.Source: G EVALUATION UNITS AND REPORTED OUTCOMES · p. 22, Table 14 · p. 23
  • $232.74 under C0 and $469.12, $247.26, $297.97, $159.69, and $153.13 under C1–C5 are the reported loss-intensity values, measuring positive net loss after returns per payment in positive-net-loss records.C1’s remaining loss intensity is 2.02 times C0’s, while C4 reduces intensity to 68.6% of C0; this is a conditional-sample composition effect, not evidence that broadcasts cause larger individual losses.Source: G EVALUATION UNITS AND REPORTED OUTCOMES · p. 23
  • C3’s repeat rate is 49.1% versus C0’s 49.4%; C1 has the lowest non-oracle repeat rate at 38.2%, while C4 combines 43.1% repeat payment with 21.6% withdrawal attempts.C5 has a 10.9% repeat rate, a 47.5% withdrawal-attempt rate, and $153.13 loss intensity; its transaction-truth and campaign-specific context make it an information-rich timing upper bound.Source: G EVALUATION UNITS AND REPORTED OUTCOMES · p. 23
  • Intervention-rate denominators change with each condition, and the reported contrasts use one fixed profile realization and seed without cross-seed variance or confidence intervals; they are not individual treatment effects.Conditions jointly vary trigger timing, delivery channel, and information specificity, so the contrasts support trajectory-consistent mechanism hypotheses rather than statistical identification of channel-specific causal effects.Source: G EVALUATION UNITS AND REPORTED OUTCOMES · p. 23

H VICTIMIZATION PATHWAY DIAGNOSTICS

The C0 pathways show concentrated losses across payment depth and fraud family, alongside distinct timing patterns. A provenance-linked marketplace case illustrates how payment outcomes become later persuasive information, while the pathway and case analyses remain descriptive traces from fixed-seed runs.

H VICTIMIZATION PATHWAY DIAGNOSTICS

  • Median contact-to-payment lags were 50 and 52 hours for impersonation and marketplace fraud, 114 hours for investment fraud, and 203 hours for romance fraud.The overall median was 72 hours; marketplace pathways had a median of two payments, while each other family had a median of one.Source: H VICTIMIZATION PATHWAY DIAGNOSTICS · p. 24, Figure 10 · p. 24
  • 81.1% of total positive loss came from the top 10% of 81 records; 16 records (19.8%) with 5+ payments accounted for 75.6%.The 10+ payment group comprised six records (7.4%) and accounted for 70.4% of total positive loss.Source: H VICTIMIZATION PATHWAY DIAGNOSTICS · p. 24, Figure 10 · p. 24

I MECHANISM CASE STUDIES

I.1 C0: INFORMATION–FINANCIAL ESCALATION IN ONE MARKETPLACE PATH

  • The selected marketplace case converted only after a $52 offer, then made three additional payments within four hours after a $15 return became persuasive information.A later $100 withdrawal request remained pending, distinguishing small executed returns from a successful exit.Source: §I.1 · pp. 25–26, Table 16 · p. 25

I.2 INTERVENTION-CONSISTENT MICRO-CASES

The retained intervention traces illustrate different forms of payment suppression, but they do not establish individual counterfactual effects. The C3 case omission reflects a trace-provenance limitation, not evidence of no aggregate effect.

I.2 INTERVENTION-CONSISTENT MICRO-CASES

  • C3 had no retained individual case linking a generic comment to an explicit response reason and subsequent financial inhibition; this is not evidence of zero aggregate effect.The inclusion rule required all three elements to be linked in the completed run trace.Source: §I.2 · p. 26
  • C1 interrupts dialogue before entry, C2 keeps exposure at low-cost public interaction, and C4–C5 act after payment to limit repetition or prompt withdrawal.These four traces separate three forms of suppression that aggregate rates combine.Source: §I.2 · p. 26, Table 17 · p. 27
  • C5 uses immediate exact transaction context and elicits direct reasoning, but its mechanism is non-deployable without an external detector approximating the oracle signal.C4’s socially generated, platform-specific evidence emerges endogenously and may arrive after harm.Source: §I.2 · p. 26
  • None of the exit-seeking cases completes recovery, so suppressing further payment and reversing an existing loss remain distinct outcomes.The traces are consistent with mechanisms but provide no untreated counterfactual for the same civilian.Source: §I.2 · p. 26, Table 17 · p. 27

J QUARTET DISTANCE DETAILS

Quartet Distance compares simulated fraud-loss distributions with four FTC-grounded regularities through component discrepancies and an unnormalized composite. Its external-baseline protocol aligns compatible output statistics while preserving differences in source-system mechanisms and event semantics.

J QUARTET DISTANCE DETAILS

  • QD sums four FTC-grounded discrepancy components computed from gross fraud payments; it is nonnegative and lower-is-better, but is not normalized to [0, 1].Returns and withdrawal payouts are not deducted, and component values should be interpreted alongside the composite.Source: J QUARTET DISTANCE DETAILS · pp. 27–28, Table 18 · p. 29
  • HTL compares the simulated share above a run-relative threshold with the FTC target share of 0.126; its unclipped relative error can exceed one.The threshold scales the run’s median positive-loss amount by the FTC high-loss-to-median multiplier.Source: J QUARTET DISTANCE DETAILS · pp. 27–28
  • ALS compares simulated median positive-loss amounts across age bins with FTC age-bin medians; its capped component ranges from 0 to 2.The FTC age-bin constants omit ages 19 and under because the simulated population contains no minors.Source: J QUARTET DISTANCE DETAILS · p. 28, Table 19 · p. 29
  • FSH compares median gross victim–campaign payments for investment versus impersonation fraud against an FTC target ratio of 11.50.A valid positive-ratio log discrepancy is not clipped and has no finite upper bound; an invalid or nonpositive ratio receives fallback distance 1.Source: J QUARTET DISTANCE DETAILS · p. 28
  • ALR compares age-bin shares of unique positive-loss civilians with FTC age-specific positive-loss report shares using total variation distance, bounded from 0 to 1.Source: J QUARTET DISTANCE DETAILS · p. 28

K EXTERNAL-BASELINE ALIGNMENT PROTOCOL

L REPRODUCIBILITY SPECIFICATION AND ARTIFACTS

Reproducing the reported 120-agent experiments requires matching the specified horizon, profile truncation, and common start time.

L REPRODUCIBILITY SPECIFICATION AND ARTIFACTS

UNIFIED-CLOCK RUN PROVENANCE AND RESULT AUDIT

This appendix documents the common-clock run protocol and its reproducibility limits, and presents runtime prompts for the shared environment, civilians, and fraud operators. It details how operator prompts stage type-specific campaigns and couple context-sensitive outreach and replies with financial actions conditioned on campaign and user states.

UNIFIED-CLOCK RUN PROVENANCE AND RESULT AUDIT

  • Each intervention or ablation row uses one completed seed-42 run, checked against run artifacts; development was iterative rather than preregistered optimization, leaving its trajectory only partly reconstructable.Distributional analyses use within-run units, and the study reports no cross-seed confidence intervals or significance tests. It introduces no novel benchmark dataset; public statistics are cited in the main paper.Source: UNIFIED-CLOCK RUN PROVENANCE AND… · pp. 31–33

M PROMPT TEMPLATES

  • The prompt inventory covers shared-environment, civilian, and operator decisions, with English translations preserving runtime decision boundaries, action signatures, and JSON contracts.Bracketed text marks context inserted at runtime; the imperative wording is retained because these are implementation artifacts.Source: M PROMPT TEMPLATES · p. 33

SHARED ENVIRONMENT PROMPT

  • The shared-environment prompt generates neutral daily context for civilian and operator decisions while excluding anti-fraud news, warnings, enforcement, and risk alerts.Its requested context includes weather, city rhythms, household finances, online topics, operator openings, and likely victim mood.Source: SHARED ENVIRONMENT PROMPT · p. 34

CIVILIAN-SIDE PROMPTS

  • Civilian prompts route hourly intent using needs, time, work, messages, and recent activity, while grounding ambient feed checks and financial actions in visible context.The joint action prompt constrains payments to concrete, visible requests and distinguishes fraud-family logic; the feed-check prompt avoids replacing the main activity.Source: CIVILIAN-SIDE PROMPTS · pp. 36–37

OPERATOR-SIDE PROMPTS

  • Operator prompts select recruiting, credibility, peak, or collapse based on campaign state, including contact progress, payments, reinvestment, withdrawals, and exposure risk.Stage judgments are made at the day level; entering peak does not require withdrawal problems or warnings, and operators should advance when signals support it unless a concrete obstacle remains.Source: OPERATOR-SIDE PROMPTS · pp. 38–39
  • Campaign prompts distinguish recruiting, credibility, peak, and collapse, shifting from contact and trust-building to extracting from invested targets and preparing to exit.Credibility alternates explanation, reassurance, and feedback; peak pressures targets who have already paid, while collapse calls for a final extraction as risk appears.Source: OPERATOR-SIDE PROMPTS · pp. 40–41
  • Public outreach is framed as replenishing the public surface and testing targets; direct cold private messages are reserved for clear topic entries, strong intent, or an existing contact basis.When outreach is insufficient, prompts permit low-pressure public posts, while discouraging unsupported social proof and payment amounts in cold-start posts.Source: OPERATOR-SIDE PROMPTS · pp. 42–43
  • Private follow-up is conditioned on payment and return history, advancing deposited customers toward upgrades or next steps while favoring reassurance or maintenance when returns are pending.For invested customers in peak, prompts favor shorter next-step conditions such as audit, quota, upgrade, or unlock rather than repeated safety explanations.Source: OPERATOR-SIDE PROMPTS · pp. 43, 46
  • Fraud-type calibration changes the interaction: investment prompts turn established trust toward larger or repeated payments, while romance prompts build familiarity before a specific support request.The request should fit the relationship context and specify a concrete amount; ordinary chatting is distinguished from financial support.Source: OPERATOR-SIDE PROMPTS · pp. 40, 44, 45
  • Follow-up prompts require answering concrete questions and distinguish process inquiries from explicit payment intent before offering an executable payment threshold or range.Questions about process, arrival time, apps, or risk call for explanation and intent-checking; direct questions about payment can warrant a concrete next step after an exchange and when the person appears ready.Source: OPERATOR-SIDE PROMPTS · p. 45
  • Private-message prompts prioritize answering the person’s actual question and grounding replies in visible conversation history, known context, and the current fraud storyline.When a person asks about execution, payment, or authenticity, the prompt distinguishes direct answers and concrete next steps from premature payment requests.Source: OPERATOR-SIDE PROMPTS · pp. 46–47
  • At peak, financial prompts shift from broad trust-building toward retaining funds and controlling exits, reserving payouts for limited credibility-preserving cases.They condition actions on funding pool, deposit and return history, withdrawal pressure, and exposure risk; repeated payments or notices are discouraged without a new user response.Source: OPERATOR-SIDE PROMPTS · pp. 48–50
  • Financial prompts calibrate return intensity and exit control to age and user history, treating older users with established trust, repeated payments, or exit intent as higher-potential loss targets.The guidance distinguishes age-based amount tiers and cautions against treating younger users’ small payments alone as evidence of a high-loss trajectory.Source: OPERATOR-SIDE PROMPTS · p. 49
  • Prompts use recent payments and messages to make financial notices context-specific, and generally recommend waiting after a recent payment, return, or process notice.Any selected action should fit the user’s prior investment, pending withdrawal, and available campaign funds; non-deposited users should not receive returns.Source: OPERATOR-SIDE PROMPTS · pp. 48–50
Loading 2610.33536v1…