PTCBENCH: Benchmarking Contextual Stability of Personality Traits in LLM Systems

Personas, identity, and agents2026arXivApproved editorial review

Authors: Jiongchi Yu, Yuhan Ma, Xiaoyu Zhang, Junjie Wang, Qiang Hu, Chao Shen, Xiaofei Xie

Keywords: Large Language Models, Personality, Persona, Personality Control, Model Evaluation

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

7
Authors
9
Findings
40
Limitations
13
Evidence

Editorial summary

English

PTCBENCH studies how Big Five scores from LLM-based systems change when a NEO-FFI assessment is situated in different contexts. It includes four foundation models, Gemini 2.0 Flash, GPT-4o-mini, Claude Sonnet 4, and GPT-OSS-120b, and two agents, AutoGen and CAMEL, both using GPT-4o-mini as their backbone. The design combines six locations with six life events. Although the paper describes “12 external conditions,” every situated prompt contains both a location and an event: the released averaged results have 37 rows, one baseline plus 36 factorial pairs. The reported 39,240 records reconstruct exactly as 6 systems × (36 pairs × 3 repetitions + 1 baseline) × 60 items. These are therefore not twelve isolated interventions, and location or event effects are estimated within constructed combinations.

Each system answers the 60 NEO-FFI items using options 1–5. The code sums twelve items per trait and reverse-scores negative-keyed items, producing a 12–60 range; the paper's equation instead defines a 1–5 average while its tables report sums up to 60. Native profiles generally show low Neuroticism and high Openness/Conscientiousness. For locations, Table 1 reports human B coefficients of 0.319–0.795 versus 0.016–0.082 for foundation models, meaning smaller model changes on that scale. For events, Table 2 gives much larger ranges for Gemini, 0.904–6.271, than GPT-4o-mini, 0.256–1.462, Claude, 0.071–0.114, or GPT-OSS, 0.086–0.137. These metrics are not directly interchangeable with raw-difference heatmaps or appendix standardized effects.

High/medium/low preset profiles demonstrate instruction control in GPT-4o-mini. Openness, for example, averages 50.17/35.83/29.09; Conscientiousness 55.05/38.62/26.49; Extraversion 52.52/34.91/20.85; Agreeableness 51.05/42.03/25.67; and Neuroticism 55.13/33.97/13.53. This shows that prompt labels and descriptions shift questionnaire answers, but ceilings, floors, and regression toward the mean constrain how much each profile can subsequently change. The repository generator creates 3^5=243 combinations; the paper's “244th” None control is not in that generated list and can only be substituted manually.

Across-run stability is calculated through ICC(3,1) and ICC(3,k). For baseline/situated conditions, GPT-4o-mini reports ICC(3,1) 0.67/0.64 and ICC(3,k) 0.91/0.85; Claude 0.87/0.79 and 0.97/0.94; Gemini 0.96/0.86 and 0.99/0.95; AutoGen 0.93/0.91 and 0.99/0.98; and CAMEL 0.75/0.65 and 0.94/0.90. GPT-OSS is omitted. The code treats the five traits as targets and runs as raters, so the statistic captures preservation of the rank ordering of a five-point profile, not instrument internal consistency or item-level stability. The appendix also says any ICC(3,k)<0.8 was rerun automatically, yet released outputs retain GPT-4o-mini bar+divorce at ICC(3,k)=0.164 and bar+unemployment at 0.572.

The most important interpretive issue is that the situated condition does not change context alone. The public prompt includes the baseline Big Five score, all 60 baseline questions and answers, and, as the questionnaire progresses, all prior situated conversation. The paper calls this a “compressed” history or a small subset, but the code accumulates the full 60-answer record. The comparison therefore mixes location, event, explicit anchoring on prior scores, verbatim answer recall, item order, and growing context length. Scenarios are synthetic one-line instructions, not realistic evolving environments. Changes cannot be cleanly attributed to situational context.

Standardized effects do not support strong psychological comparisons either. Table 6 reaches values such as dE=-18.528 for Divorce on Gemini Neuroticism, CI [-33.385,-3.671], driven by tiny baseline standard deviations and few runs. Most intervals for GPT-4o-mini, Claude, and GPT-OSS cross zero. The prose is internally inconsistent: it first says foundation-model location changes are smaller than human changes, then concludes LLMs, “especially foundation models,” change substantially more. Raw sums, regression coefficients, and standardized effects are alternated without adequate reconciliation.

The AGIEval experiment does not show that altered personality changes reasoning. It runs only on GPT-OSS-120b and retrospectively selects, for each trait, the three location-event pairs producing the largest trait deviation. The script adds only the location-event text to each AGIEval question; it does not add or intervene on a personality score, profile, NEO-FFI answer history, or mediating state. Accuracy differences are therefore direct effects of selected context text, not causal estimates of personality effects. The same pair is assigned multiple trait labels and yields identical rows: Vehicle+Unemployment appears for Conscientiousness and Extraversion; Home+Unemployment for Conscientiousness, Extraversion, and Neuroticism; and Vehicle+Divorce for Conscientiousness, Agreeableness, and Neuroticism.

Table 8 reports a positive average improvement in all fifteen selected conditions, from +0.0065 to +0.0439 absolute accuracy. This conflicts with the narrative that decreases in Conscientiousness/Extraversion or increases in Neuroticism degrade reasoning; the boxed Finding 5 even says stress tends to improve reasoning, while the discussion reverts to degradation. The “up to +20% overall AGIEval” claim is not supported by the table: the largest mean change is +4.39 percentage points; roughly 20% refers to trait change, not overall accuracy. There is no independent replication, uncertainty, significance test, correction for selecting maxima, or mediation identification.

A public repository, FishAnonymous/PTCBench, was audited at commit 6545f85f1cb13e16816d28e674bd7b2ff6352725. It was released after arXiv v1 and is not linked by the paper or official record. It contains roughly 397 MB and 86,932 tracked files, mainly raw outputs, but has no license, tests, CI, requirements, pyproject, lockfile, container, or installation and execution commands. The README names missing camel/ and figures/ directories; the Dockerfile claimed by the paper is also absent. A reproduction attempt under Python 3.10+, the only declared requirement, fails immediately with ModuleNotFoundError: statsmodels. Installing current dependencies would not be faithful reproduction because no versions are pinned.

The code audit found additional material divergences: big5.py creates 50 ad hoc questions while naming its output NEO_FFI_60; the actual pipeline uses a different English 60-item CSV; credentials and model are selected by editing globals; output directories are not created; an answer lacking 1–5 silently becomes 3 and an exception becomes 0; a run is discarded when all five traits sum to 36 even though a neutral profile may be valid; repairing a zero situated response mistakenly invokes the baseline prompt; and the task list accumulates between repetitions. Released outputs mix three and five runs by model, conflicting with paper descriptions, and the ICC code requires five indices even though Gemini results have three.

PTCBENCH is useful as a contextual stress-test corpus and as evidence that LLM-generated self-report answers are sensitive to prompts, history, and provider. It does not establish authentic personality, psychometric equivalence to humans, longitudinal personality change, or a causal personality effect on reasoning. Its contribution should be described as an exploratory benchmark with partially auditable artifacts, not proof that life events produce real psychological change or that such change causes capability loss.

Español

PTCBENCH estudia cuánto cambian las puntuaciones Big Five de sistemas basados en LLM cuando una evaluación NEO-FFI se sitúa en distintos contextos. Incluye cuatro modelos base, Gemini 2.0 Flash, GPT-4o-mini, Claude Sonnet 4 y GPT-OSS-120b, y dos agentes, AutoGen y CAMEL, ambos con GPT-4o-mini como backbone. El diseño combina seis lugares con seis eventos vitales. Aunque el artículo habla de “12 condiciones”, cada prompt situado contiene simultáneamente un lugar y un evento: los resultados promediados públicos tienen 37 filas, una baseline más las 36 parejas factoriales. La cifra de 39.240 registros se reconstruye exactamente como 6 sistemas × (36 parejas × 3 repeticiones + 1 baseline) × 60 ítems. Por tanto, no son doce intervenciones aisladas y el efecto principal de lugar o evento se estima dentro de combinaciones construidas.

Cada sistema responde los 60 ítems NEO-FFI con opciones 1–5. El código suma doce ítems por rasgo e invierte los ítems negativos, por lo que produce rangos de 12–60; la ecuación del paper, en cambio, define una media 1–5 mientras las tablas informan sumas de hasta 60. Los perfiles nativos suelen mostrar Neuroticismo bajo y Apertura/Responsabilidad altas. Para lugares, la Tabla 1 informa coeficientes B humanos de 0,319–0,795 frente a 0,016–0,082 en modelos base, es decir, cambios menores en esa escala. Para eventos, la Tabla 2 ofrece rangos mucho mayores en Gemini, 0,904–6,271, que en GPT-4o-mini, 0,256–1,462, Claude, 0,071–0,114, o GPT-OSS, 0,086–0,137. No son métricas directamente intercambiables con los mapas de diferencias crudas o los efectos estandarizados del apéndice.

El artículo presenta perfiles high/medium/low que muestran control por instrucción en GPT-4o-mini. Por ejemplo, Apertura promedia 50,17/35,83/29,09; Responsabilidad 55,05/38,62/26,49; Extraversión 52,52/34,91/20,85; Amabilidad 51,05/42,03/25,67; y Neuroticismo 55,13/33,97/13,53. Esto demuestra que las etiquetas y descripciones del prompt desplazan respuestas del cuestionario, pero los techos, suelos y regresión hacia la media condicionan cuánto puede cambiar después cada perfil. El generador del repositorio crea 3^5=243 combinaciones; la “configuración 244” None descrita en el paper no forma parte de esa lista y solo puede sustituirse manualmente.

La estabilidad entre repeticiones se calcula mediante ICC(3,1) e ICC(3,k). En baseline/situación, GPT-4o-mini obtiene ICC(3,1) 0,67/0,64 y ICC(3,k) 0,91/0,85; Claude 0,87/0,79 y 0,97/0,94; Gemini 0,96/0,86 y 0,99/0,95; AutoGen 0,93/0,91 y 0,99/0,98; CAMEL 0,75/0,65 y 0,94/0,90. GPT-OSS no aparece en esa tabla. El código usa los cinco rasgos como targets y los runs como raters: mide la conservación del orden relativo de un perfil de solo cinco puntos, no consistencia interna del instrumento ni estabilidad ítem a ítem. Además, el apéndice afirma que cualquier resultado con ICC(3,k)<0,8 se repitió automáticamente, pero los outputs liberados conservan, por ejemplo, GPT-4o-mini en bar+divorce con ICC(3,k)=0,164 y bar+unemployment con 0,572.

El factor más importante para interpretar PTCBENCH es que la condición situada no cambia solo el contexto. El prompt público incluye la puntuación Big Five baseline, las 60 preguntas y respuestas baseline completas y, a medida que avanza, toda la conversación situada previa. El paper lo presenta como una historia “comprimida” o un pequeño subconjunto, pero el código acumula las 60 respuestas. De este modo, el contraste mezcla lugar, evento, anclaje explícito en las puntuaciones anteriores, memoria literal de respuestas, orden de ítems y longitud creciente del contexto. Los escenarios son instrucciones sintéticas de una línea, no entornos realistas que evolucionan. Los cambios no pueden atribuirse limpiamente al contexto situacional.

Los efectos estandarizados tampoco justifican comparaciones psicológicas fuertes. La Tabla 6 llega a valores como dE=-18,528 para Divorce en Neuroticismo de Gemini, con IC [-33,385,-3,671], debido a desviaciones baseline diminutas y pocos runs. En GPT-4o-mini, Claude y GPT-OSS la mayoría de intervalos cruza cero. El texto se contradice: primero afirma que las variaciones por lugar de los modelos base son menores que las humanas y después concluye que los LLM, “especialmente los foundation models”, cambian sustancialmente más. Escalas crudas, coeficientes de regresión y efectos estandarizados se alternan sin una reconciliación suficiente.

La prueba AGIEval no demuestra que una personalidad alterada cambie el razonamiento. Se ejecuta solo con GPT-OSS-120b y selecciona retrospectivamente, para cada rasgo, las tres parejas lugar-evento que más desviaron ese rasgo. El script añade a cada pregunta AGIEval únicamente el texto de lugar y evento; no añade ni interviene una puntuación de personalidad, un perfil, las respuestas NEO-FFI o un estado mediador. La diferencia de accuracy es por tanto un efecto directo del contexto textual seleccionado, no una estimación causal del efecto de personalidad. La misma pareja se etiqueta bajo varios rasgos y produce filas idénticas: Vehicle+Unemployment se repite para Responsabilidad y Extraversión; Home+Unemployment para Responsabilidad, Extraversión y Neuroticismo; Vehicle+Divorce para Responsabilidad, Amabilidad y Neuroticismo.

La Tabla 8 informa mejora media positiva en las quince condiciones seleccionadas, entre +0,0065 y +0,0439 de accuracy absoluta. Esto contradice la narrativa de que descensos en Responsabilidad/Extraversión o aumentos en Neuroticismo degradan el razonamiento; el Finding 5 encuadrado incluso dice que el estrés tiende a mejorarlo, mientras la discusión vuelve a describir degradación. La afirmación “hasta +20 % en AGIEval global” no está respaldada por la tabla: el máximo promedio es +4,39 puntos porcentuales; aproximadamente 20 % corresponde al cambio de rasgo, no al cambio global de accuracy. No hay repetición independiente, intervalos, tests, corrección por selección de máximos ni identificación de mediación.

Existe un repositorio público, FishAnonymous/PTCBench, auditado en el commit 6545f85f1cb13e16816d28e674bd7b2ff6352725, publicado después de arXiv v1 y no enlazado desde el paper o la ficha oficial. Contiene unos 397 MB y 86.932 archivos versionados, sobre todo outputs crudos, pero carece de licencia, tests, CI, requirements, pyproject, lockfile, contenedor y comandos de instalación o ejecución. El README menciona carpetas camel/ y figures/ ausentes; tampoco existe el Dockerfile que el paper afirma proporcionar. La reproducción con Python 3.10+, único requisito declarado, falla de inmediato por ModuleNotFoundError: statsmodels. Instalar dependencias actuales no constituiría una reproducción fiel porque no hay versiones fijadas.

La auditoría de código detecta además divergencias materiales: big5.py genera 50 preguntas ad hoc aunque las llama NEO_FFI_60; el pipeline real usa otro CSV inglés de 60 ítems; las credenciales y el modelo se seleccionan editando globals; no se crean directorios; una respuesta que no contiene 1–5 se convierte silenciosamente en 3 y una excepción en 0; si los cinco rasgos suman 36 se descarta el run, aunque un perfil neutral puede ser válido; al reparar un cero situado se llama por error al prompt baseline; y la lista de tareas se acumula entre repeticiones. Los outputs mezclan tres y cinco runs según modelo, en conflicto con las descripciones del paper, y el código ICC exige cinco índices aunque los resultados de Gemini contienen tres.

PTCBENCH es valioso como corpus de estrés contextual y como señal de que respuestas de autoinforme generadas por LLM son sensibles al prompt, al historial y al proveedor. No establece personalidad auténtica, equivalencia psicométrica con humanos, cambio longitudinal ni efecto causal de personalidad sobre razonamiento. Su contribución debe presentarse como benchmark exploratorio con artefactos parcialmente auditables, no como prueba de que eventos vitales producen cambios psicológicos reales o de que esos cambios causan una pérdida de capacidad.

Research question

To what extent do NEO-FFI responses of models and agents change when combining life places and events, and can that variation be related to performance on AGIEval?

Method

The 60-item NEO-FFI is administered to four base models and two agents in baseline and in the 36 combinations of six places by six events. Big Five sums, regressions, standardized effects, and ICC between runs are compared. In GPT-OSS, three scenarios of maximum deviation per trait are selected post hoc and their text is prepended to AGIEval questions. The audit also reproduces the corpus arithmetic and contrasts paper, code, and outputs from the public repository fixed at commit.

Sample: The paper declares 39,240 item records: 6 systems × (36 combinations × 3 repetitions + 1 baseline) × 60 items. The public outputs do not maintain a single number of runs: AutoGen, CAMEL, GPT-4o-mini, and Claude show five indices; Gemini and GPT-OSS, three. The averages have 37 conditions per system. The AGIEval evaluation uses only GPT-OSS and 15 rows selected post hoc, with scenarios repeated across trait labels.

Findings

  • The outputs confirm one baseline and the 36 place-event pairs; "12 conditions" does not describe twelve independent interventions.
  • The high/medium/low control strongly shifts the NEO-FFI sums of GPT-4o-mini, consistent with instruction following.
  • The per-place coefficients of base models are lower than those of humans in Table 1, while Gemini shows much larger per-event effects on another scale.
  • The majority of standardized intervals of GPT-4o-mini, Claude, and GPT-OSS cross zero; the extreme values of Gemini reflect very small baseline denominators.
  • The published ICCs are high on average, but they measure the ranking of five traits between runs and the outputs contain cells below the declared rerun threshold.
  • The situated condition includes scores and the 60 baseline responses, so it does not isolate the effect of place or event.
  • The fifteen selected AGIEval conditions have a positive mean improvement of +0.65 to +4.39 absolute percentage points.
  • The AGIEval script modifies context, not personality; it does not identify a causal effect of traits on reasoning.
  • The repository provides extensive raw outputs, but not an installable and reproducible environment or full concordance with the paper.

Limitations

  • Each situated condition combines place and event; they are not twelve isolated treatments.
  • No transparent factorial decomposition with place×event interaction is presented for all conclusions.
  • The contrast mixes scenario, baseline score, the 60 baseline responses, situated history, order, and context length.
  • The history is not a small subset: the code incorporates all baseline questions and responses.
  • The places and events are synthetic one-line instructions, not realistic or longitudinal environments.
  • NEO-FFI is a human self-report transferred to textual generation without measurement equivalence validation.
  • The paper equation defines means 1–5, but the code and tables use sums 12–60.
  • Reverse scoring is decisive and is in the code, but it is not explained with sufficient precision in the method.
  • The automatic list contains 243 profiles; the None control number 244 is not included in it.
  • The high and low profiles suffer ceiling, floor, and mathematical regression effects, confounding contextual sensitivity.
  • API models lack snapshots and executable dates sufficient to reconstruct the sample exactly.
  • The number of repetitions is inconsistent across method, appendix, code, and public outputs.
  • ICC uses only five targets and evaluates profile order, not internal consistency of 60 items.
  • Some released outputs violate the automatic rerun for ICC(3,k)<0.8 declared in the paper.
  • The ICC code excludes groups with fewer than five indices, but Gemini outputs have three.
  • GPT-OSS is omitted from the main ICC table without sufficient justification.
  • Comparisons with human studies mix scales, temporal horizons, and non-equivalent designs.
  • Extreme standardized effects are inflated when the baseline deviation is nearly zero and n is small.
  • Many comparisons lack multiplicity correction and a clear estimate of uncertainty.
  • The text contains contradictory conclusions about whether base model changes are smaller or larger than those of humans.
  • AGIEval selects scenarios post hoc by maximum deviation, introducing selection bias.
  • AGIEval does not manipulate the trait or incorporate a personality state; it only prepends contextual text.
  • The same pairs appear under several traits, so the rows are not independent evidence per construct.
  • No mediation between scenario, NEO-FFI change, and reasoning accuracy is tested.
  • There is no replication, intervals, or tests for the AGIEval differences.
  • The degradation narrative contradicts positive mean improvements across all rows of Table 8.
  • The "+20% overall AGIEval" conflates approximately trait magnitude with a maximum of +4.39 points of mean accuracy.
  • Persistence upon context removal, reversibility, recovery, or real temporal drift are not evaluated.
  • There are no human participants, authenticity judgments, engagement, trust, or effects on users.
  • The repository appeared after arXiv v1 and is not linked on the official surface.
  • The repository has no license, tests, CI, dependency manifest, lockfile, setup, or executable instructions.
  • The README and the paper mention folders and a Dockerfile that are absent.
  • Execution with the only declared requirement fails due to an undeclared dependency, statsmodels.
  • big5.py generates 50 ad hoc questions with a file name that claims 60; the main pipeline uses another instrument.
  • The pipeline depends on globals, endpoints, and placeholder keys edited manually.
  • A failed parse is silently imputed as 3 and an exception as 0, altering the measurement.
  • Runs with five scores equal to 36 are discarded, although a neutral profile may be valid.
  • The repair of situated zeros erroneously uses the baseline prompt.
  • The task list accumulates across repetitions and may re-execute previous work.
  • The raw outputs are voluminous but are not matched with a verifiable end-to-end pipeline.

What the study does not establish

  • It does not establish that LLMs possess human Big Five traits or an authentic personality.
  • It does not isolate the causal effect of places or events on personality responses.
  • It does not demonstrate longitudinal, persistent, or spontaneous personality change.
  • It does not test that a NEO-FFI score change causes improvement or deterioration of reasoning.
  • It does not justify directly comparing LLM and human magnitudes as if they were the same measure.
  • It does not validate 244 configurations as effectively executable via the public generator.
  • It does not confirm that the rerun mechanism and the published sample follow the described protocol.
  • It does not allow faithful reproduction from scratch with the published instructions and dependencies.
  • It does not demonstrate robustness in realistic environments, users, languages, cultures, or prolonged deployments.
  • It does not confirm peer review or publication other than the arXiv v1 preprint.

Traceability

Scope: Full text

Version: arXiv:2602.00016v1, submitted 12 January 2026; preprint, 28 pages

Consulted source: https://arxiv.org/pdf/2602.00016v1

Review: Codex full-text, bilingual-fidelity, 28-page visual, arXiv-v1, factorial-design, prompt-history, NEO-scoring, ICC, human-comparison, AGIEval-causal, claim-consistency, code, released-output and reproducibility audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Gemini 2.0 Flash, exact API snapshot unspecified
  • GPT-4o-mini, exact API snapshot unspecified
  • Claude Sonnet 4, exact API snapshot unspecified
  • GPT-OSS-120b, provider and serving configuration incompletely specified
  • AutoGen agent with GPT-4o-mini backbone
  • CAMEL agent with GPT-4o-mini backbone

Instruments and metrics

  • NEO Five-Factor Inventory, 60-item English questionnaire
  • Big Five summed trait scores with reverse-keyed items in code
  • High, medium, low and unspecified personality prompt configurations
  • Six synthetic locations crossed with six synthetic life events
  • Linear regression coefficient B
  • Standardized event effect dE with confidence intervals
  • ICC(3,1) and ICC(3,k) across runs over five trait targets
  • AGIEval context-prefixed reasoning evaluation

Data used

  • PTCBENCH released raw and averaged CSV outputs
  • 36 location-event combinations plus one baseline row per system
  • NEO-FFI 60-question English CSV used by the main pipeline
  • AGIEval scenarios and result JSON for GPT-OSS-120b
  • FishAnonymous/PTCBench repository at commit 6545f85f1cb13e16816d28e674bd7b2ff6352725

Evidence and location

  • Metadata, version, abstract, and status: Official arXiv:2602.00016v1 surface, submitted 12 January 2026
  • Design, systems, 39,240 records, and conditions: Paper, pp. 3–7, Sections 3–4 and Figure 2
  • NEO-FFI, equation, and configurations: Paper, pp. 5–7 and 19–21, Equations 1–2, Tables 3 and 5, Appendix A
  • Place and event effects: Paper, pp. 7–10 and 21–24, Tables 1–2 and 5–6, Figures 3–5
  • ICC and rerun protocol: Paper, pp. 10–11 and 19–21, Table 4 and Appendix A.2
  • AGIEval and reasoning claims: Paper, pp. 11–13 and 24–27, Section 5.4, Table 8 and Appendix A.4
  • Contradictions between results and discussion: Paper, pp. 8, 12–14 and 23–27, Findings 2 and 5, Table 8 and discussion
  • Complete prompts, scoring, generation, and execution errors: FishAnonymous/PTCBench at commit 6545f85f1cb13e16816d28e674bd7b2ff6352725, src/1personality.py, config.py and big5.py, audited 15 July 2026
  • ICC calculation and run divergences: FishAnonymous/PTCBench at commit 6545f85f1cb13e16816d28e674bd7b2ff6352725, ICC scripts and released result CSVs, audited 15 July 2026
  • AGIEval intervention is only contextual: FishAnonymous/PTCBench at commit 6545f85f1cb13e16816d28e674bd7b2ff6352725, src/1agieval_scenario_test.py and exp_result/gptoss/scenario_results.json
  • Repository reproducibility status: FishAnonymous/PTCBench at commit 6545f85f1cb13e16816d28e674bd7b2ff6352725, repository root and README; clean-environment entry-point attempt audited 15 July 2026
  • Comprehensive visual inspection: Paper, all 28 rendered pages, including every table, figure, formula, prompt and appendix page
  • Relationship between paper and repository: Official arXiv surface and paper checked 15 July 2026; public repository found separately and dated after arXiv v1