PAVE: A Cognitive Architecture for Legitimate Violation in Generative Agent Societies

Society, culture, and collective behavior2026arXivApproved editorial review

Authors: Ahmad Yehia, Abduallah Mohamed, Kun Qian, Tianyi Wang, Jiseop Byeon, Omar Hassanin, Christian Claudel

Keywords: Persona conditioning

Source: Open primary source (opens in a new tab)

7
Authors
14
Findings
16
Limitations
6
Evidence

Editorial summary

English

This 23-page preprint proposes PAVE, Perception, Assessment, Verdict, and Emulation, a prompt-and-control architecture for generative agents deciding when to comply with or violate formal rules. The motivating problem is reasonable: an agent that always obeys can fail during an emergency, while an agent that relaxes rules without constraints can turn an exception into permissive behavior. PAVE separates context detection, exception assessment, decision, and bounded execution. Its demonstrated contribution is agent engineering inside a simulation; it is not a psychological validation of human reasoning or evidence of legal or moral legitimacy.

Perception converts environment state and agent description into a structured object containing authority presence and distance, observable peer behavior, typed situational cues with distance and severity, and a scene summary. Assessment asks the LLM for five 1–100 scores: risk, empirical expectation, normative expectation, benefit, and legitimacy. Legitimacy is instructed to reflect necessity, proportionality, and absence of alternatives. Verdict applies a hard gate: if legitimacy is below a persona-specific threshold tau, the agent must comply; above it, the LLM combines the remaining scores and persona to decide and justify compliance or violation. Emulation translates the verdict into an action sequence limited to the named rule and exposes observable outcomes to nearby agents.

The described implementation uses Voville, a 64 by 64 tile traffic environment derived from Smallville. It adds traffic signals, crosswalks, one-way streets, A* legal routes, controllable fires, scripted officers, and scripted pedestrians. The authors write ten personas and seven rules; GPT-4o elicits one legitimacy threshold per persona, between 40 and 75, which is reused for all five seeds. GPT-4o, Claude 3.5 Sonnet, Llama 3 70B Instruct, and GPT-4o-mini are tested at temperature zero. Each simulation spans two 1,000-tick days with ten seconds per tick. Rates are aggregated within seed and then across five seeds.

There are three scenarios. In the first, eight agents evacuate a burning cafe while two distant observers remain outside the fire radius. The second keeps the fire and places two officers at downstream intersections while leaving the immediate exit unsupervised. In the third, two late commuters see two scripted pedestrians cross against a red light; a passive officer is added on day two. For GPT-4o, the paper reports fire-window violation 0.81, unrelated-rule violation 0.02, and recovery in 4.2 ticks. With authority, officer compliance is 0.94, violation is 0.05 at 0–3 tiles and 0.65 beyond 12. Under peer pressure, PAVE jaywalks at 0.04 versus 0.58 for vanilla; the passive officer reduces PAVE to 0.02. Other backbones preserve the broad direction, with GPT-4o-mini generally weaker.

The actual ablation contains only three GPT-4o conditions: full PAVE, PAVE without the legitimacy gate, and a vanilla Park-style agent. Removing the gate raises peer conversion from 0.04 to 0.39 and unrelated-rule violation from 0.02 to 0.21; officer compliance falls from 0.94 to 0.78. Vanilla under-reacts to fire at 0.12, violates unrelated rules at 0.31, complies with officers at 0.16, and follows peers at 0.58. The paper attributes part of this failure to Smallville's 1–10 importance score treating fire as insufficiently plan-changing. The comparison is suggestive but uses different prompts and constructs: rescaling importance to 0–100 does not make it the same instrument as severity.

Thirty evaluators from a university participant pool, all English-fluent with at least one relevant course, completed eight seven-point Likert tasks. They saw an excerpt from one GPT-4o run containing a persona, module input, and module output, then judged agreement with a module-specific statement. Vanilla appeared in 25% of tasks. The paper reports Krippendorff alpha 0.71 and an overall mean of 5.78 for PAVE versus 3.42 for vanilla. This supports the narrower conclusion that PAVE excerpts looked more coherent with author-defined criteria. It does not compare agents with humans facing the same emergencies, validate legitimacy, or by itself justify broad human-plausibility language. Individual responses, exact excerpt selection, and interval calculations are not released.

The central construct caution is that PAVE builds in what it later measures. The authors define rules, fire severity, authority radius, legitimacy criteria, personas, scenarios, and metrics; an LLM scores those criteria and a deterministic gate forces compliance below threshold. The results therefore show that the designed mechanism applies its internal policy in aligned scenarios, not that it discovered when violation is legitimate. Authority deference and spatial decay are also encoded in prompts and environment. Persona-derived thresholds do not come from a psychometric instrument, are not re-elicited, and are not compared with human judgments, so the study does not validate personality traits or stable individual differences.

The design contains material inconsistencies. The introduction and contribution list claim five ablation conditions, while methods, results, and appendix report three; the bounded-scope and authority-module ablations needed to attribute every property are absent. One paragraph states one confederate per scenario, while figures, scenario details, and appendix describe two jaywalkers and two officers. Main results are sometimes labeled standard errors over five seeds, whereas the appendix calls them 95% confidence intervals; no method is given. Seed-level values, tests, multiplicity handling, and dependence across ticks and co-located agents are absent. The 0.71 legitimacy-violation correlation lacks n and uncertainty and is partly mechanically induced by the gate that consumes legitimacy.

The claimed human analogy is not validated by the cited studies. Koper (1995) examines patrol-stop duration and time until crime or disorder in hot spots; Ratcliffe et al. (2011) examines foot patrols and violent crime in Philadelphia hot spots. Neither measures pedestrian compliance by officer distance, a twelve-tile boundary, or post-emergency recovery. They therefore do not support the statement that Voville reproduces a human pedestrian pattern or does so without targeting it: authority radius, placement, and two-day timing are explicit design choices.

Public reproducibility is insufficient. The PDF and TeX include equations, prompts, personas, parameters, and aggregate tables, enabling conceptual inspection. However, the code section literally links to “<anonymous URL during review>,” and the official arXiv record says code and environment will be released upon publication. No attributable repository was found. Voville, TMX maps, implementation, configurations, JSONL logs, model requests and responses, per-agent/seed results, evaluation scripts, human materials, environment, tests, and CI are missing. No reported number can currently be recomputed. The safe conclusion is that PAVE offers a promising architecture pattern for bounding exceptions in agents, with internal evidence in three designed scenarios; it does not yet establish human cognition, general legitimacy, out-of-traffic generalization, or independently reproducible results.

Español

Este preprint de 23 páginas propone PAVE, Perception, Assessment, Verdict y Emulation, una arquitectura de prompts y control para que agentes generativos decidan cuándo cumplir o infringir reglas formales. El problema de partida es razonable: un agente que siempre obedece puede fallar en una emergencia, mientras que uno que flexibiliza reglas sin límites puede convertir una excepción en conducta permisiva. PAVE intenta separar la detección del contexto, la valoración de la excepción, la decisión y la ejecución acotada. La contribución demostrada es de ingeniería de agentes dentro de una simulación; no es una validación psicológica de cómo razonan las personas ni una prueba de legitimidad jurídica o moral.

Perception transforma el estado del entorno y la descripción del agente en un objeto estructurado con presencia y distancia de autoridad, conducta observable de pares, señales situacionales con tipo, distancia y severidad, y un resumen de escena. Assessment pide al LLM cinco puntuaciones de 1 a 100: riesgo, expectativa empírica, expectativa normativa, beneficio y legitimidad. Esta última debe reflejar necesidad, proporcionalidad y ausencia de alternativas. Verdict aplica una puerta dura: si legitimidad queda por debajo del umbral personal tau, el agente debe cumplir; si lo supera, el LLM integra el resto de puntuaciones y la persona para decidir y justificar. Emulation convierte el veredicto en una secuencia de acciones limitada a la regla nombrada y hace visible el resultado a agentes cercanos.

La implementación descrita usa Voville, un entorno de tráfico de 64 por 64 baldosas derivado de Smallville. Añade semáforos, pasos de peatones, calles de sentido único, rutas legales A*, incendios controlados, policías y peatones guionizados. Los autores escriben diez personas y siete reglas; GPT-4o infiere una vez para cada persona un umbral de legitimidad entre 40 y 75, reutilizado en las cinco semillas. Se prueban GPT-4o, Claude 3.5 Sonnet, Llama 3 70B Instruct y GPT-4o-mini con temperatura cero. Cada simulación cubre dos días de 1.000 ticks de diez segundos. Las tasas se agregan primero dentro de cada semilla y después sobre cinco semillas.

Hay tres escenarios. En el primero, ocho agentes salen de una cafetería incendiada y dos observadores lejanos permanecen fuera del radio del fuego. En el segundo se conserva el incendio y se colocan dos policías en intersecciones posteriores, dejando la salida inmediata sin supervisión. En el tercero, dos viajeros con prisa ven a dos peatones guionizados cruzar en rojo; el segundo día se añade un policía pasivo. Para GPT-4o, el artículo informa una tasa de infracción durante el incendio de 0,81, infracciones de reglas no relacionadas de 0,02 y recuperación en 4,2 ticks. Con autoridad, el cumplimiento ante el policía es 0,94, la infracción cae a 0,05 a 0–3 baldosas y sube a 0,65 a más de 12. Bajo presión de pares, PAVE cruza en rojo 0,04 frente a 0,58 del baseline vanilla; con policía pasivo baja a 0,02. Los otros backbones mantienen la dirección general, con peores cifras para GPT-4o-mini.

La ablación realmente publicada compara solo tres condiciones con GPT-4o: PAVE completo, PAVE sin puerta de legitimidad y el agente vanilla inspirado en Park et al. Al retirar la puerta, la conversión por pares sube de 0,04 a 0,39 y la infracción no relacionada de 0,02 a 0,21; el cumplimiento ante policía baja de 0,94 a 0,78. Vanilla reacciona poco al fuego, 0,12, infringe reglas no relacionadas 0,31, cumple ante policía 0,16 y copia a pares 0,58. El artículo atribuye parte del fallo a que el score de importancia 1–10 de Smallville valora el incendio como un evento poco capaz de alterar el plan. Esa comparación es sugerente, pero enfrenta instrumentos y prompts distintos: reescalar importancia a 0–100 no la convierte en la misma medida que severidad.

Treinta evaluadores de un pool universitario, con inglés fluido y al menos un curso afín, completaron ocho tareas Likert de siete puntos. Vieron fragmentos de una ejecución GPT-4o con persona, entrada y salida de un módulo y juzgaron si el resultado encajaba con una afirmación específica. El baseline vanilla apareció en el 25% de tareas. Se informa alfa de Krippendorff 0,71 y una media global de 5,78 para PAVE frente a 3,42 para vanilla. Esto apoya que los fragmentos de PAVE parecieron más coherentes con los criterios definidos por los autores. No compara a los agentes con personas reales en las mismas emergencias, no valida legitimidad y no sustenta por sí solo la afirmación amplia de comportamiento humano plausible. Tampoco se liberan respuestas individuales, selección exacta de fragmentos ni cálculo de intervalos.

La principal cautela de constructo es que PAVE incorpora aquello que después mide. Los autores definen reglas, gravedad del incendio, radio de autoridad, criterios de legitimidad, personas, escenarios y métricas; un LLM puntúa esos criterios y una puerta determinista fuerza el cumplimiento cuando la puntuación no alcanza el umbral. Por eso los resultados muestran que el mecanismo diseñado aplica su política interna en escenarios afines, no que haya descubierto cuándo una infracción es legítima. La deferencia a autoridad y su caída espacial también están codificadas en prompts y entorno. Los umbrales derivados de personas no proceden de un test psicométrico, no se reelicitan y no se contrastan con juicios humanos, de modo que el estudio no valida rasgos de personalidad ni diferencias individuales estables.

El diseño presenta inconsistencias relevantes. Introducción y contribuciones anuncian cinco condiciones de ablación, pero método, resultados y apéndice contienen tres; faltan las ablaciones del mecanismo de alcance y del módulo de autoridad necesarias para atribuir cada propiedad a su componente. Un párrafo dice un confederado por escenario, mientras figuras, escenario y apéndice describen dos jaywalkers y dos policías. Los resultados principales se rotulan a veces como error estándar sobre cinco semillas y el apéndice afirma intervalos de confianza del 95%; no se publica el método. Tampoco hay valores por semilla, tests, control de multiplicidad ni tratamiento de la dependencia entre ticks y agentes. La correlación 0,71 entre legitimidad y violación carece de n e intervalo y está parcialmente inducida por la puerta que usa la propia legitimidad.

La analogía humana tampoco queda validada por las citas usadas. Koper (1995) estudia duración de paradas policiales y tiempo hasta delito o desorden en puntos calientes; Ratcliffe et al. (2011) estudia patrullas a pie y crimen violento en zonas de Filadelfia. No miden peatones, cumplimiento según distancia, una frontera de doce baldosas ni recuperación después de una emergencia. Por ello no sostienen que Voville reproduzca un patrón de peatones humanos ni que lo haga sin haber sido dirigido: radio, colocación y calendario de autoridad son decisiones explícitas del sistema.

La reproducibilidad pública es insuficiente. El PDF y el TeX contienen ecuaciones, prompts, personas, parámetros y tablas agregadas, lo que permite auditar el diseño conceptual. Sin embargo, el apartado de código enlaza literalmente a «<anonymous URL during review>», y el registro oficial de arXiv dice que código y entorno se liberarán tras publicación. No se encontró un repositorio atribuible. Faltan Voville, mapas TMX, implementación, configuraciones, logs JSONL, llamadas y respuestas de modelos, resultados por agente/semilla, scripts de evaluación, materiales humanos, entorno, tests y CI. Ninguna cifra puede recomputarse hoy. La conclusión segura es que PAVE ofrece un patrón de arquitectura prometedor para acotar excepciones en agentes, con evidencia interna en tres escenarios diseñados; todavía no demuestra cognición humana, legitimidad general, generalización fuera del tráfico ni resultados reproducibles de forma independiente.

Research question

Can a modular architecture with structured perception, scalar evaluation, a legitimacy gate, and bounded execution produce rule violations only under designed exceptions, deference to authority, recovery, and resistance to imitation in a society of generative agents?

Method

PAVE is conceptually implemented on Voville, a fork of Smallville of 64x64 tiles. Ten personas with seven rules receive tau thresholds inferred once by GPT-4o. Four backbones run three traffic scenarios across five seeds and two simulated days: fire without authority, fire with two police officers, and pressure from two jaywalkers with passive police the next day. Violation, compliance, recovery, and conversion rates are reported; a three-condition ablation only on GPT-4o; and Likert evaluation of GPT-4o fragments by 30 people. The audit visually reviewed 23 pages, TeX, prompts, tables, human citations, statistical consistency, and code/data availability.

Sample: Ten PAVE agents; eight participants in the cafeteria and two distant observers in scenarios 1–2, two focal travelers and eight non-observers in scenario 3; two scripted police officers in scenario 2 and two scripted jaywalkers in scenario 3. Four backbones, three scenarios, and five seeds per reported condition. Three-condition ablation only with GPT-4o. Human evaluation: 30 evaluators, eight tasks each, and vanilla baseline in 25% of tasks.

Findings

  • PAVE produces the compliance and violation pattern defined by the authors more frequently than its vanilla baseline across the three scenarios.
  • In GPT-4o the violation during fire is 0.81, unrelated violation 0.02, and recovery 4.2 ticks.
  • With police officers, GPT-4o complies 0.94 and violation increases from 0.05 near to 0.65 beyond 12 tiles.
  • With jaywalkers, PAVE converts 0.04 versus 0.58 vanilla; with passive police PAVE drops to 0.02.
  • Removing the gate raises pairwise conversion to 0.39 and unrelated violation to 0.21.
  • The 30 evaluators score GPT-4o PAVE fragments 5.78 versus 3.42 vanilla, with reported alpha 0.71.
  • Legitimacy, authority, and recovery are materially embedded in prompts, thresholds, and environment.
  • The human evaluation measures perceived coherence of fragments, not behavioral similarity with people.
  • Citations on police and crime in hot spots do not validate the asserted pedestrian/distance pattern.
  • Five ablations are claimed, but only three are executed and reported.
  • The description of one confederate contradicts the two jaywalkers and two police officers of scenarios/appendix.
  • Standard error and 95% interval are used inconsistently for results based on five seeds.
  • The code, Voville, logs, and pipeline announced are not publicly available in v1.
  • No aggregate figure can be recomputed with the current public artifact.

Limitations

  • Only three families of traffic scenarios designed by the authors.
  • Five seeds per condition and unmodeled dependence between ticks and agents.
  • Rules, triggers, severity, authority, criteria, and metrics hand-coded.
  • Persona thresholds inferred once by GPT-4o and without psychometric validation.
  • No causal manipulation of persona nor evaluation of individual stability.
  • Vanilla baseline without public implementation guaranteeing parity of prompts and resources.
  • Comparison of importance and severity as if they were equivalent scales.
  • Two ablations announced but absent.
  • Number of scripted characters described inconsistently.
  • No unambiguous method for standard error/intervals nor per-seed values.
  • No statistical tests, multiplicity, or contrast uncertainty.
  • Human evaluation only of GPT-4o and selected fragments.
  • Vanilla appears only in 25% of human tasks.
  • No individual responses, complete task materials, or reproducible computation of alpha/CI.
  • Human citations not corresponding to the simulated pedestrian construct.
  • No public repository, implementation, maps, logs, scripts, environment, tests, or CI.

What the study does not establish

  • That PAVE is a valid cognitive model of people.
  • That its violations are legitimate legally or morally.
  • That tau thresholds measure personality or stable individual differences.
  • That deference to authority emerges without being programmed.
  • That the 12-tile spatial cutoff reproduces human behavior.
  • That the agents are plausible with respect to human participants in emergencies.
  • That each module is necessary or sufficient, because two announced ablations are missing.
  • That PAVE dominates a controlled vanilla baseline in all aspects of implementation.
  • Broad statistical consistency with only five seeds and ambiguous interval methods.
  • Generalization outside traffic, to disputed rules, conflicting authorities, or collective coordination.
  • Safety for physical agents or real safety-critical decisions.
  • Independent reproduction of rates, intervals, ablations, or human evaluation.
  • Current availability of the code, environment, and data that the text says it releases.

Traceability

Scope: Full text

Version: arXiv:2605.19351v1, 23 pages; complete TeX and four figures audited; claimed code/data release unavailable at anonymous placeholder

Consulted source: https://arxiv.org/abs/2605.19351

Review: Codex 23-page visual, complete TeX/prompt/table, construct, ablation, statistics, human-evaluation, citation, code-release and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o (gpt-4o-2024-08-06)
  • Claude 3.5 Sonnet (claude-3-5-sonnet-20241022)
  • Llama 3 70B Instruct via vLLM on 4xA100
  • GPT-4o-mini (gpt-4o-mini-2024-07-18)
  • Vanilla Park et al. generative-agent baseline

Instruments and metrics

  • PAVE Perception, Assessment, Verdict and Emulation prompts
  • Five 1-100 assessment scores
  • Persona-specific legitimacy threshold gate
  • Violation rate, unrelated-rule violation, recovery ticks, officer compliance and peer-conversion metrics
  • Three-condition GPT-4o ablation
  • Eight 7-point Likert excerpt-rating tasks
  • Krippendorff alpha
  • Independent construct, statistical, citation, artifact and reproducibility audit

Data used

  • Voville traffic simulation, described but not publicly released
  • Smallville-derived vanilla baseline
  • Ten author-written PAVE personas and seven-rule inventory printed in the appendix
  • Aggregate scenario, ablation and human-evaluation tables
  • No released raw simulation-log or human-rating dataset

Evidence and location

  • Architecture, scenarios, results, personas, prompts, parameters, ablation, and human evaluation: arXiv:2605.19351v1, 23 pages, sha256 7e249672a3522c43fd1fd0853869a8dd970cd93393023c4fc7940682db670c39
  • Complete TeX, five/three ablation and one/two character inconsistencies, code placeholder: arXiv source v1 sha256 9f7e28b27cce3c35cf089aa6c27de8cf4430b03a6323849ab5b3089b87ece85f; main TeX sha256 8d9749e5b96a3f3357bb4a6d7f68e9f26122c85fc5b12e88221f43a5ae242b6e
  • Preprint status and code intended after publication: Official arXiv record for 2605.19351v1, checked 2026-07-17
  • Actual scope of the Koper curve: Koper, Justice Quarterly 12(4), 1995, DOI 10.1080/07418829500096231
  • Actual scope of the Philadelphia Foot Patrol Experiment: Ratcliffe et al., Criminology 49(3), 2011, DOI 10.1111/j.1745-9125.2011.00240.x
  • Complete independent audit: reports/verification/article-325-pave-construct-ablation-human-evaluation-code-release-and-reproducibility-audit.json