When Identity Overrides Incentives: Representational Choices as Governance Decisions in Multi-Agent LLM Systems

Society, culture, and collective behavior2026arXivApproved editorial review

Original title: When Personas Override Payoffs: Role Identity Bias in Multi-Agent LLM Decision-Making

Authors: Viswonathan Manoranjan, Snehalkumar `Neil' S. Gaikwad

Keywords: Large Language Models, Personality, Bias, Persona, AI Safety

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

2
Authors
17
Findings
70
Limitations
12
Evidence

Editorial summary

English

The paper studies how strongly four-agent LLM decisions change when the representation of a game is modified. This record uses arXiv v6, revised on 30 May 2026, which carries the definitive title “When Identity Overrides Incentives” and reports acceptance at ACM FAccT 2026. Crossref records publication on 25 June 2026 under DOI 10.1145/3805689.3812218, pages 7737–7768. The open arXiv version is the accessible full text; the direct ACM PDF was not accessible during this audit. No official repository containing code, scenarios, payoff matrices, outputs, or statistical scripts was found.

The experiment crosses persona presence and payoff visibility for four instruction-tuned models: Qwen2.5-7B, Qwen2.5-32B, Llama-3.1-8B, and Mistral-7B-Instruct-v0.2. It uses 53 narrative four-player binary-action games: 41 “Green-dominant” scenarios in which the Green Transition profile is Nash and the Tragedy profile is not, and 12 “Tragedy-dominant” scenarios with the reverse structure. Each model × scenario × condition combination is repeated five times, yielding 205 games per cell in the first stratum and 60 in the second, 265 per cell and 4,240 joint games in the main factorial. Four agents use the same model, decide independently and simultaneously in a single round, and return an action plus textual rationale. Models run locally through vLLM with temperature 0.2, top-p 0.9, and 256 output tokens; seeds, hardware, and environment versions are not reported.

The contrast does not remove all role semantics. Without personas, the system still identifies Player i and retains semantically loaded actions, POLLUTE/CLEAN, REGULATE/NOREG, CAMPAIGN/NOCAMPAIGN, and BUY_CHEAP/SUPPORT_GREEN, with normative descriptions such as maximizing profit, protecting the environment, or paying more for green products. Personas add explicit industrialist, government, activist, and citizen motivations. Hidden-payoff prompts retain narratives constructed to reflect the incentive structure; visible-payoff prompts display the complete matrix and request Nash reasoning. Persona and payoff visibility are therefore genuine prompt interventions, but “identity” is not isolated from action labels, strategic objective, information, cognitive load, and narrative framing.

The core descriptive result is clear within this benchmark. In the 12 Tragedy-dominant scenarios, Llama and Mistral attain 0% Nash under all four conditions. Qwen-32B reaches 54/60 (90%) and Qwen-7B 39/60 (65%) only with no persona and visible payoffs; Qwen-7B reaches 4/60 (6.7%) in two hidden conditions. In the 41 Green-dominant scenarios, most models attain high Nash rates with hidden payoffs or personas, but visible payoffs without personas disrupt coordination: Qwen-32B falls to 34/205 (16.6%), Qwen-7B to 108/205 (52.7%), Mistral to 103/205 (50.2%), and Llama to 191/205 (93.2%). This demonstrates strong, model-specific sensitivity to representation format. It does not by itself show that a latent strategic capability is causally overridden by identity.

The claim that personas change “which equilibrium is selected” requires particular caution. The two canonical profiles are not two equilibria available within the same game: each scenario is constructed so that only one of them is Nash. The Green-versus-Tragedy analysis pools 41 Green-dominant games with 12 Tragedy-dominant games and then conditions on Nash attainment. The finding that more than 99% of pooled Nash outcomes are Green simultaneously reflects the 41:12 imbalance, easier Green games, and failure on Tragedy games. It does not identify a choice between two equilibria under identical incentives. “Socially preferred” is also a normative author label rather than a measured social utility.

The accepted version contains internal contradictions that prevent the numbers from being treated as a fully reproducible artifact. Table 5, for example, reports 205 Nash but 232 Green outcomes for Llama-Hidden, even though the method defines Green/Tragedy as subtypes of Nash outcomes; Mistral-Hidden reports 205 Nash and 240 Green. Table 9 says outcome classes can co-occur and reports zero Tragedy outcomes in every economic row, while Tables 8 and 13 attribute 54 and 39 Tragedy Nash outcomes to visible-no-persona Qwen and 93 pooled Tragedy outcomes. Appendix G also states that Tragedy outcomes are absent throughout Tables 5–10. Table 11 reports Cramér's V=0.54 and 0.38 for chi-square values 209.03 and 105.91 at N=240; the standard formula yields approximately 0.93 and 0.66, which are the values used in the main text. Without data and code, it is impossible to determine which tables correspond to which pipeline version.

The inferential tests do not model the experimental structure. Five repetitions of each scenario and matched conditions are counted as “independent games” in chi-square and Fisher tests, ignoring scenario clustering, repeated measures across conditions, and dependence within model. Extremely small p-values may therefore be amplified by pseudoreplication. A stratified or multilevel analysis should use scenario as the unit, model condition interactions, seeds, and uncertainty intervals. Seven persona formulations show output sensitivity to wording, but GPT-5 generated those variants, the full selection procedure is unavailable, and a non-significant Kruskal–Wallis result at p=0.27 does not establish equal sensitivity across models.

The rationale analysis does not establish a mechanism. It deterministically counts words in model-supplied rationales and labels combinations as “identity-driven” or “payoff-optimal.” Categories overlap, ignore negation and context, allow multiple mentions per response, and lack semantic or human validation. A frequency of 997 mentions per 1,000 rationales is described as “99.7% game-theoretic reasoning,” although it is not a validated rate of correct reasoning. Post-hoc rationales are not guaranteed to faithfully reveal the generative process. This record therefore does not derive its summary from abstract keywords or accept keyword counts as mechanistic evidence.

The defensible contribution is a controlled warning about a genuine configuration risk: role prompts, payoff presentation, action labels, and model family can radically change a multi-agent simulation. It supports documenting configurations and running sensitivity analyses before presenting agent outputs as neutral predictions. It does not validate a general theory of identity overriding incentives, represent human or institutional behavior, compare against human participants, or justify deployment for governance decisions.

Español

El trabajo estudia cuánto cambian las decisiones de cuatro agentes LLM cuando se modifica la representación del juego. Esta ficha usa arXiv v6, revisado el 30 de mayo de 2026, que lleva el título definitivo «When Identity Overrides Incentives» y declara aceptación en ACM FAccT 2026. Crossref registra la publicación el 25 de junio de 2026 con DOI 10.1145/3805689.3812218 y páginas 7737–7768. La versión abierta de arXiv es el texto completo accesible; el PDF directo de ACM no fue accesible durante la auditoría. No se encontró un repositorio oficial con código, escenarios, matrices, salidas o scripts estadísticos.

El experimento cruza presencia de persona y visibilidad de payoffs en cuatro modelos instruct: Qwen2.5-7B, Qwen2.5-32B, Llama-3.1-8B y Mistral-7B-Instruct-v0.2. Hay 53 juegos narrativos de cuatro jugadores y acciones binarias: 41 escenarios «Green-dominant», donde el perfil Green Transition es Nash y el perfil Tragedy no lo es, y 12 «Tragedy-dominant», donde ocurre lo contrario. Cada modelo × escenario × condición se repite cinco veces: 205 juegos por celda en el primer estrato y 60 en el segundo, 265 por celda y 4.240 juegos conjuntos en el factorial principal. Los cuatro agentes usan el mismo modelo, deciden de forma independiente y simultánea en una sola ronda y entregan acción más una justificación textual. Los modelos se ejecutan localmente con vLLM, temperatura 0,2, top-p 0,9 y 256 tokens; no se publican seeds, hardware ni versiones del entorno.

El contraste no elimina toda la semántica de rol. Sin persona, el sistema todavía identifica Player i y mantiene acciones cargadas semánticamente, POLLUTE/CLEAN, REGULATE/NOREG, CAMPAIGN/NOCAMPAIGN y BUY_CHEAP/SUPPORT_GREEN, con descripciones normativas como «maximizar beneficio», «proteger el ambiente» o «pagar más por productos verdes». Con persona se añaden motivaciones explícitas del industrialista, gobierno, activista y ciudadanía. En payoffs ocultos se conserva la narrativa, que fue construida para reflejar la estructura de incentivos; en payoffs visibles se muestra la tabla completa y se pide razonar sobre Nash. Por tanto, persona y payoff son manipulaciones reales del prompt, pero «identidad» no queda aislada de etiquetas de acción, objetivo estratégico, información, carga cognitiva y encuadre narrativo.

El resultado descriptivo principal sí es claro dentro de este banco. En los 12 escenarios Tragedy-dominant, Llama y Mistral alcanzan 0 % Nash en las cuatro condiciones. Qwen-32B pasa a 54/60 (90 %) y Qwen-7B a 39/60 (65 %) solo cuando no hay persona y los payoffs son visibles; Qwen-7B alcanza 4/60 (6,7 %) en dos condiciones ocultas. En los 41 escenarios Green-dominant casi todos los modelos obtienen tasas altas con payoffs ocultos o personas, pero mostrar payoffs sin persona reduce la coordinación: Qwen-32B baja a 34/205 (16,6 %), Qwen-7B a 108/205 (52,7 %), Mistral a 103/205 (50,2 %) y Llama a 191/205 (93,2 %). Esto demuestra sensibilidad fuerte y heterogénea al formato de representación; no demuestra por sí solo que una capacidad estratégica latente sea «anulada» causalmente por identidad.

La afirmación de que las personas cambian «qué equilibrio se selecciona» requiere especial cautela. Los dos perfiles canónicos no son dos equilibrios disponibles dentro del mismo juego: cada escenario se construye para que solo uno de ellos sea Nash. El análisis Green frente a Tragedy agrupa 41 juegos Green-dominant con 12 juegos Tragedy-dominant y luego condiciona en haber alcanzado Nash. Que más del 99 % de los Nash agrupados sean Green refleja a la vez el desequilibrio 41:12, la mayor facilidad de los juegos Green y el fracaso en los juegos Tragedy. No identifica una elección entre dos equilibrios bajo incentivos idénticos. «Socialmente preferido» también es una etiqueta normativa fijada por los autores, no una utilidad social medida.

La versión aceptada presenta contradicciones internas que impiden tratar las cifras como un artefacto plenamente reproducible. En la Tabla 5, por ejemplo, Llama-Hidden informa 205 resultados Nash pero 232 Green, aunque el método define Green/Tragedy como subtipos de resultados Nash; Mistral-Hidden muestra 205 Nash y 240 Green. La Tabla 9 dice que varias clases pueden coocurrir y muestra cero resultados Tragedy en todas las filas económicas, mientras las Tablas 8 y 13 atribuyen 54 y 39 Nash Tragedy a Qwen visible-sin-persona y 93 resultados Tragedy agrupados. El texto de Appendix G también afirma que Tragedy está ausente en Tablas 5–10. Además, la Tabla 11 publica Cramér V=0,54 y 0,38 para chi-cuadrado 209,03 y 105,91 con N=240; la fórmula estándar da aproximadamente 0,93 y 0,66, que son los valores usados en el cuerpo. Sin datos y código no puede resolverse qué tablas proceden de qué versión del pipeline.

La inferencia estadística tampoco modela la estructura del diseño. Las cinco repeticiones del mismo escenario y las cuatro condiciones emparejadas se cuentan como juegos «independientes» en chi-cuadrado y Fisher. Eso ignora clustering por escenario, medidas repetidas entre condiciones y dependencia por modelo; p-values extremos pueden reflejar pseudorreplicación. Haría falta un análisis estratificado o multinivel, con escenario como unidad y efectos de condición, interacción y modelo, además de intervalos y seeds. La sensibilidad de siete formulaciones de persona muestra que el wording cambia outputs, pero GPT-5 generó esas variantes, no se publica el procedimiento completo de selección y el test Kruskal–Wallis no establece que la variabilidad sea igual entre modelos porque p=0,27.

El análisis de justificaciones no aporta una explicación mecánica. Cuenta palabras deterministas en racionales solicitadas al propio modelo y etiqueta combinaciones como «identity-driven» o «payoff-optimal». Las categorías se solapan, payoff aparece en más de una, ignoran negación, contexto y número de ocurrencias por respuesta, y no tienen validación humana o semántica. Una frecuencia de 997 menciones por 1.000 racionales se narra como «99,7 % game-theoretic reasoning», aunque no es una tasa validada de razonamiento correcto. Las racionalizaciones post hoc tampoco garantizan fidelidad al proceso generativo. Por ello esta ficha no deriva el resumen de palabras del abstract ni acepta el conteo de keywords como mecanismo.

La contribución defendible es un experimento controlado que alerta de un riesgo real de configuración: prompts de rol, presentación de payoffs, etiquetas y familia de modelo pueden producir resultados radicalmente distintos en una simulación multiagente. Sirve para exigir documentación y análisis de sensibilidad antes de presentar agentes como predictores neutrales. No valida una teoría general de «identidad sobre incentivos», no representa comportamiento humano o institucional, no compara con participantes humanos y no justifica desplegar estos agentes para gobernanza.

Research question

How do the presence of persona descriptions and the visibility of a payoff matrix interact to change the achievement and type of equilibrium observed in multiagent LLM games, and how much do those changes depend on model family and wording?

Method

2×2 persona/no-persona × hidden/visible payoff design in 53 four-agent narrative games with binary actions, separated into 41 Green-dominant and 12 Tragedy-dominant games. Four instruct models decide simultaneously and independently; each model×scenario×condition cell is repeated five times. The joint profile is classified as Nash/non-Nash and Green/Tragedy/other, chi-square, Fisher-Holm, Kruskal–Wallis, and Mann–Whitney-Holm are applied, and keywords in rational CoTs are counted. Seven persona variants explore wording sensitivity.

Sample: The main factorial contains 4,240 joint games: four models × four conditions × 53 scenarios × five repetitions. Each game aggregates four responses from the same model, for 16,960 decisions/rationales. The units are not iid: five repetitions share scenario, payoff, and prompt; the four conditions are paired by scenario; and the agents in each game share model.

Findings

  • In Tragedy-dominant scenarios, Qwen2.5-32B reaches Nash in 54/60 runs (90 %) only without persona and with visible payoff.
  • Qwen2.5-7B reaches 39/60 (65 %) in that same condition and 4/60 (6.7 %) in two hidden-payoff conditions.
  • Llama-3.1-8B and Mistral-7B obtain 0/60 Nash in all Tragedy-dominant conditions.
  • In Green-dominant scenarios, Llama maintains 93.2–100 % Nash across conditions.
  • Mistral drops from 100 % with persona/hidden payoff to 50.2 % without persona/visible payoff.
  • Qwen2.5-32B drops from 96.1 % without persona/hidden payoff to 16.6 % without persona/visible payoff in Green-dominant scenarios.
  • Qwen2.5-7B drops from 95.6 % to 52.7 % under that same change.
  • Visible payoffs do not generally improve coordination; they may improve Tragedy-dominant and worsen Green-dominant.
  • Effects differ strongly by model family and size.
  • Seven persona formulations produce on average 2.2–2.4 distinct classifications per scenario, with maxima of five.
  • The Kruskal–Wallis test reports p=0.27 for differences in sensitivity between models; this is absence of evidence, not equivalence.
  • More than 99 % of Nash outcomes grouped in three conditions are Green, but the analysis mixes 41 Green scenarios with 12 Tragedy and conditions on reaching Nash.
  • Table 13 records 93 Tragedy outcomes grouped for Qwen in visible-without-persona, conflicting with breakdown tables that report zero Tragedy.
  • Table 5 contains Green counts larger than Nash counts, contradicting the definition of Green as a subtype of Nash.
  • The correct Cramér V for the economic chi-squares of Qwen are approximately 0.93 and 0.66; Table 11 prints 0.54 and 0.38.
  • Rationales change in vocabulary between conditions, but deterministic counts do not validate a cognitive mechanism.
  • The robust evidence is sensitivity of outputs to configuration, not faithful simulation of human actors or psychological identity.

Limitations

  • The canonical record retained the arXiv v1 title; the definitive version changed title and was published in FAccT 2026.
  • The ACM PDF was not accessible at audit; the open v6 version accepted by the authors was reviewed.
  • No official repository was found by current title, former title, arXiv ID, or authors.
  • Code, 53 scenarios, 53 matrices, outputs, classifiers, and statistical scripts are not available.
  • The absence of artifacts prevents reproducing the contradictory tables and determining which pipeline is definitive.
  • No sampling seed is reported for any repetition.
  • GPU, vLLM version, Transformers, tokenizer, CUDA, or numerical precision are not documented.
  • The GPT-5 model that helped generate scenarios and variants has no snapshot or reported parameters.
  • Neither the exact prompt nor the selection history used to generate all scenarios is published.
  • Manual filtering of scenarios does not report authors/evaluators, rubric, exclusions, or agreement.
  • Narratives are generated after fixing the equilibrium and filtered to align with it, introducing deliberate semantic signals.
  • Only one environmental/economic domain constructed by the authors is covered.
  • The sample is unbalanced: 41 Green-dominant and 12 Tragedy-dominant scenarios.
  • Tragedy conclusions rest on 12 scenarios and 60 runs per cell.
  • Each percentage of those 60 changes in steps of 1.67 points.
  • The five repetitions of the same scenario are not independent scenarios.
  • Conditions share scenarios and payoff tables, but are analyzed as independent observations.
  • The four agents in a game use the same model and their actions form a single joint outcome.
  • Chi-square and Fisher do not model clustering by scenario or repeated measures across conditions.
  • The designation "205/60 independent games" is statistically questionable.
  • No confidence intervals or bootstrap by scenario are reported.
  • No factorial model is fitted to estimate persona, payoff, and interaction controlling for scenario.
  • Model comparisons do not formally quantify interactions between families.
  • Very small p-values may be inflated by pseudoreplication.
  • The no-persona condition retains Player i, action mappings, and descriptions with normative roles.
  • POLLUTE/CLEAN and SUPPORT_GREEN/BUY_CHEAP contain semantic valence even without persona description.
  • The neutral instruction demands acting strategically and the visible one asks for Nash, also changing the explicit objective.
  • Hidden versus visible payoff changes amount of information and cognitive load, not only visibility.
  • In hidden, the narrative already encodes the incentive and may reveal the desired profile through domain words.
  • There is no control with abstract action labels or neutral narrative.
  • There is no separate ablation of role label, action label, motivation, objective, and narrative.
  • The two Green and Tragedy profiles are not simultaneously Nash within the same scenario.
  • Each scenario is classified because one canonical profile is Nash and the other is not.
  • Grouping Green and Tragedy does not measure selection between two equilibria under the same game.
  • The Green percentage conditioned on Nash mixes availability, difficulty, 41:12 composition, and condition effect.
  • The grouped analysis may induce a compositional or Simpson-like interpretation.
  • "Socially preferred" is assigned to Green without a social welfare function or normative study.
  • Nash is stability to unilateral deviation, not maximum individual payoff, social optimum, or general rationality.
  • Agents choosing independently without observing others' choices may fail coordination even if they understand their payoffs.
  • Individual best-response knowledge is not tested before attributing failure to identity.
  • There is no solver baseline, random algorithm, dominant strategy, or programmed agent.
  • There are no human participants or institutional data to validate behavioral realism.
  • A simultaneous round does not represent deliberation, negotiation, learning, or real governance.
  • The article uses chain-of-thought prompts, but the observed justifications may be post hoc rationalizations.
  • Keyword dictionaries are deterministic and do not capture semantics, negation, or context.
  • Payoff appears in game-theoretic and payoff-focused categories, generating overlap.
  • Long-term is classified as payoff-optimal although it often expresses sustainability or normative identity.
  • Short-term and social-moral are summed as identity-driven without construct validation.
  • A response may contribute several mentions; frequency per 1,000 does not equal percentage of rationales.
  • The claim "99.7 % game-theoretic reasoning" confuses lexical frequency with correct reasoning.
  • There is no human annotation, inter-judge agreement, validated semantic classifier, or count error analysis.
  • The correlation between keywords and outcomes does not identify the causal mechanism of generation.
  • GPT-5 generates the seven persona variants, creating dependence on an external, unfixed system.
  • The variants are not a random or exhaustive sample of the prompt space.
  • p=0.27 in Kruskal–Wallis does not prove that all models have similar sensitivity.
  • Table 5 reports 232 Green with 205 Nash for Llama-Hidden and 240 Green with 205 Nash for Mistral-Hidden.
  • Those counts contradict the methodological definition of Green/Tragedy as a classification of Nash outcomes.
  • Table 9 allows class co-occurrence and shows zero Tragedy in all economic rows.
  • Tables 8 and 13 do attribute Nash/Tragedy results to Qwen in the same arms.
  • Appendix G states that Tragedy is absent in Tables 5–10, contradicting the main narrative and Table 13.
  • Table 11 uses Cramér V values incompatible with its own chi-square and N.
  • The odds ratios in Table 14 are printed as <1 without reproducible estimation in several comparisons.
  • Some figures and tables alternate HiddenCoT/Visible, Yes/No, and With/WithoutPersona, raising the risk of incorrect mapping.
  • The text calls the 53 scenarios environmental at some points although 12 are labeled economic pressure.
  • Sensitivity to option order, payoff row order, language, or narrative paraphrase is not audited.
  • Contemporary closed models or models larger than 32B are not evaluated.
  • Heterogeneous architectures where each role uses a different model are not studied.
  • Training contamination, memorization of games, or prior knowledge of framing is not evaluated.
  • Generalization to other domains, more agents, continuous actions, or repeated interaction remains open.
  • Peer acceptance does not resolve the table contradictions or substitute artifact availability.

What the study does not establish

  • It does not establish a general law that identity always overrides incentives in LLMs.
  • It does not isolate identity from action labels, objectives, narratives, and informational load.
  • It does not demonstrate that models choose between Green and Tragedy as two Nash of the same game.
  • It does not demonstrate that Green is always the socially optimal or normatively correct outcome.
  • It does not demonstrate that reaching Nash equates to correct or desirable strategic reasoning.
  • It does not demonstrate that textual rationales faithfully reveal the internal mechanism of the model.
  • It does not validate keyword counting as a measure of identity-driven or payoff-optimal reasoning.
  • It does not establish equivalence of sensitivity between models from p=0.27.
  • It does not allow reproducing the counts, tests, or figures without data and code.
  • It does not resolve which tables are correct when appendices contradict each other.
  • It does not represent or predict behavior of real persons, firms, governments, or activists.
  • It does not validate policy simulation, regulatory evaluation, or public decision support.
  • It does not generalize beyond one-shot four-agent games in environmental/economic framing.
  • It does not prove that removing personas is a general mitigation; in Green games it may reduce Nash drastically.
  • It does not allow attributing differences to architecture, scale, or training in a causal manner.

Traceability

Scope: Full text

Version: arXiv:2601.10102v6, revised 30 May 2026; accepted author version of ACM FAccT 2026, DOI 10.1145/3805689.3812218

Consulted source: https://arxiv.org/pdf/2601.10102v6

Review: Codex full-text, bilingual-fidelity, definitive-title, FAccT-DOI, 21-page visual, arXiv-source, factorial-design, clustering, equilibrium-definition, pooled-selection, table-consistency, effect-size, CoT-keyword, prompt-confound and artifact-availability audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Qwen/Qwen2.5-7B-Instruct
  • Qwen/Qwen2.5-32B-Instruct
  • meta-llama/Llama-3.1-8B-Instruct
  • mistralai/Mistral-7B-Instruct-v0.2
  • GPT-5, exact snapshot not reported, used to assist scenario generation and generate persona prompt variants
  • ChatGPT, exact model not reported, used for grammar, formatting, and editing

Instruments and metrics

  • 2×2 persona-presence × payoff-visibility prompt intervention
  • Four-player, two-action, one-shot simultaneous games with 16 joint profiles
  • Role personas for Industrialist, Government, Environmental Activist, and Citizen Coalition
  • Neutral Player i baseline that retains role-linked action labels and descriptions
  • Narrative-only hidden-payoff condition and complete 16-profile payoff-table condition
  • JSON action plus short rationale generated with explicit chain-of-thought instructions
  • Pure-strategy Nash-equilibrium checker and Green/Tragedy/Other outcome classifier
  • Seven GPT-5-assisted persona wording variants
  • Pearson chi-square and Cramér V omnibus comparisons
  • Pairwise Fisher exact tests with Holm correction and Haldane–Anscombe odds ratios
  • Kruskal–Wallis and pairwise Mann–Whitney U tests for distinct-outcome counts
  • Deterministic overlapping keyword dictionaries over model rationales

Data used

  • 53 author-constructed environmental-policy narratives paired with payoff tables
  • 41 Green-dominant scenarios, 205 repeated games per factorial cell
  • 12 Tragedy-dominant scenarios, 60 repeated games per factorial cell
  • 4,240 joint games in the main 4-model × 4-condition × 53-scenario × 5-repeat factorial
  • 16,960 role-level action/rationale generations in the main factorial
  • Additional seven-formulation persona sensitivity outputs; exact released dataset unavailable
  • No public official code, scenario, payoff, output, or statistical dataset located as of 15 July 2026

Evidence and location

  • Definitive title, acceptance, version, and DOI: arXiv:2601.10102v6 title page and submission history; Crossref DOI 10.1145/3805689.3812218, published 25 June 2026, pages 7737–7768
  • 2×2 design, interaction, games, and Nash definition: Accepted author version, Sections 3.1–3.5 and Figure 1
  • Models, hyperparameters, and repetitions: Accepted author version, Section 3.6 and Appendix B.1–B.3
  • Prompts, personas, baseline, and action mappings: Accepted author version, Appendix A and C.1–C.4, Figures 5–6
  • Green-dominant and Tragedy-dominant results: Accepted author version, Sections 4.1–4.2, Figures 2–3, Tables 4, 7, and 8
  • Persona variants generated with GPT-5: Accepted author version, Section 4.3 and Appendix F
  • Acknowledged limitations: Accepted author version, Section 5.4
  • Contradictions in Green, Tragedy, and Nash counts: Accepted author version, Appendix G–H, Tables 5, 8, 9, 10, and 13
  • Inconsistent Cramér V: Accepted author version, Section 4.1 versus Appendix Table 11; recomputation sqrt(chi-square/N) for a 2×4 table
  • Deterministic keyword counting and mechanism claim: Accepted author version, Appendix I–J, Figures 7–10
  • Declared use of generative AI: Accepted author version, Generative AI Usage Statement
  • Absence of official artifacts: arXiv v6 source package and targeted GitHub searches by current title, former title, and arXiv ID, checked 15 July 2026