InCharacter: Evaluating Personality Fidelity in Role-Playing Agents through Psychological Interviews

Personas, identity, and agents2024ACL AnthologyApproved editorial review

Authors: Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, Jiangjie Chen, Cheng Li, Yanghua Xiao

Keywords: role-playing agents, personality assessment, psychological scales, character fidelity, LLM evaluation, Big Five personality traits, psychometric testing, 16Personalities, LLM-as-a-judge

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

13
Authors
38
Findings
60
Limitations
35
Evidence

Editorial summary

English

InCharacter evaluates whether role-playing agents reproduce personality labels attributed to their target characters. Instead of asking an agent to complete a questionnaire directly, it rewrites items from 14 scales as open-ended questions, asks each in an isolated context, and uses another LLM either to convert the answer back to an option or to score a dimension as a simulated expert. A preliminary validation compares GPT-3.5, GPT-4, and Gemini with human judgments on 100 BFI cases. GPT-4 reaches 89% accuracy in expert rating and Pearson, Spearman, and Kendall correlations of 0.925, 0.927, and 0.837, while still differing by more than one point in four cases. The main benchmark covers 32 popular characters: 16 ChatHaruhi and 16 RoleLLM configurations, normally run on gpt-3.5-turbo-1106. For BFI and 16Personalities, positive or negative types come from Personality Database vote percentages, whereas continuous distances use scores from two or three invited annotators per character. With batched expert rating and GPT-4, Table 2 reports 76.6% dimension accuracy, 31.2% all-dimensions accuracy, and 18.2% normalized absolute error for BFI; corresponding 16Personalities results are 80.7%, 44.8%, and 20.5%. Thus, the abstract's 80.7% does not mean that 80.7% of characters have their entire personality reproduced correctly. Across 14 scales using GPT-3.5 as evaluator, published averages are 78.9% per dimension, 58.5% full accuracy, and 8.1% error, but mean inter-annotator reliability is κ = 0.609 and falls to 0.438 for EPQ-R, 0.473 for ECR-R, and 0.463 for CABIN. Interview methods generally outperform self-report on the selected metrics and yield more dispersed character measurements, although the comparison is asymmetric: refusals or non-compliant self-reports are mapped to neutral, while evaluator failures are regenerated and Gemini may be replaced by GPT-4. Name masking also leaves narrative clues that may reveal the character to the evaluator. The study contains no real-person assessment, longitudinal behavior, or test of whether the scales retain construct validity when applied to agents. The official repository publishes compilable code, 14 questionnaires, and aggregated labels under MIT, but it does not contain the announced 18,304 dialogues, result files, or individual annotations; dependencies are unpinned, and reproduction requires retired model snapshots, external services, and an unofficial 16Personalities API. The evidence supports partial agreement with human labels and discrimination among configurations within this benchmark; it does not establish internal personality or general character fidelity.

Español

InCharacter evalúa si agentes que interpretan personajes reproducen etiquetas de personalidad atribuidas a esos personajes. En lugar de pedir al agente que marque directamente un cuestionario, transforma los ítems de 14 escalas en preguntas abiertas, formula cada una en un contexto aislado y usa otro LLM para convertir las respuestas a opciones o puntuar una dimensión como supuesto experto. La validación previa compara GPT-3.5, GPT-4 y Gemini con juicios humanos en 100 casos BFI. GPT-4 obtiene 89 % de aciertos en expert rating y correlaciones de 0,925, 0,927 y 0,837 para Pearson, Spearman y Kendall; todavía discrepa en cuatro casos por más de un punto. El benchmark principal contiene 32 personajes populares: 16 configuraciones de ChatHaruhi y 16 de RoleLLM, normalmente ejecutadas sobre gpt-3.5-turbo-1106. Para BFI y 16Personalities, el tipo positivo o negativo procede de porcentajes de Personality Database, mientras que la distancia continua usa puntuaciones de 2–3 anotadores invitados por personaje. Con expert rating por lotes y GPT-4, la Tabla 2 reporta para BFI 76,6 % de precisión por dimensión, 31,2 % de personajes con todas las dimensiones correctas y 18,2 % de error absoluto normalizado; para 16Personalities, 80,7 %, 44,8 % y 20,5 %. Por tanto, el 80,7 % del abstract no significa que el 80,7 % de los personajes tenga su personalidad completa correctamente reproducida. En 14 escalas, usando GPT-3.5 como evaluador, la media publicada es 78,9 % por dimensión, 58,5 % de acierto completo y 8,1 % de error, pero la fiabilidad entre anotadores promedia κ = 0,609 y cae a 0,438 en EPQ-R, 0,473 en ECR-R y 0,463 en CABIN. La entrevista supera generalmente al autoinforme en las métricas elegidas y produce mediciones más dispersas entre personajes, aunque la comparación no es completamente simétrica: las negativas o respuestas no conformes del autoinforme se convierten en neutral, mientras los fallos de evaluación se regeneran y Gemini puede ser sustituido por GPT-4. El anonimizado oculta nombres pero no referencias narrativas que pueden revelar el personaje al evaluador. No hay evaluación de personas reales, conducta longitudinal ni una prueba de que estas escalas conserven validez de constructo al aplicarse a agentes. El repositorio oficial publica código, 14 cuestionarios y etiquetas agregadas bajo MIT, y el código compila; sin embargo, no incluye los 18.304 diálogos, resultados ni anotaciones individuales anunciados, las dependencias no están fijadas y la reproducción requiere modelos retirados, servicios externos y una API no oficial de 16Personalities. La evidencia respalda que el procedimiento discrimina configuraciones y se acerca parcialmente a etiquetas humanas bajo este benchmark; no demuestra una personalidad interna ni fidelidad general del personaje.

Research question

Can the personality of a role-playing agent be measured with greater fidelity through open-ended questions derived from psychological scales and subsequent evaluation by another LLM than through direct self-report, and to what extent do the resulting measurements coincide with human labels of the target characters?

Method

InCharacter consists of interview and evaluation. A LLM rewrites and the authors manually verify each item as an open-ended question; the RPA answers each question in an independent context. Then, an evaluator LLM applies option conversion, dimension-specific conversion, or expert rating over batches of 3-4 responses or all responses of a dimension. The evaluator's capability is contrasted with 100 human BFI judgments. The study runs three repetitions of several methods over 32 RPAs from ChatHaruhi and RoleLLM, compares scores with Personality Database types and invited annotations, and calculates accuracy per dimension, full hit, normalized MAE, and three measures of variation. It extends the analysis to 14 scales, base models, character data sources, languages, question adaptation, and self-report with in-context examples.

Sample: The main benchmark uses 32 characters selected for availability of RPA and at least ten annotations in Personality Database, diversity of types, and manual identity verification; 16 come from ChatHaruhi and 16 from RoleLLM. Six ChatHaruhi characters use data and interview in Chinese and the rest in English. Each BFI/16P method is repeated three times. For the own labels, between two and three students familiar with each character contribute 93 complete annotations in total, each with 73 dimensions across 14 scales. The interviewer LLM validation uses 100 manually judged BFI cases. No humans participate as evaluated subjects nor are users observed interacting with the RPAs.

Findings

  • InCharacter measures the agreement between outputs of conditioned agents and labels attributed to characters; it does not administer a test to real people.
  • The interview converts closed items into open-ended questions and keeps each question in an isolated context to avoid interference between items.
  • The evaluation phase offers option conversion, conversion with descriptive options per dimension, and direct expert rating.
  • Batch expert rating normally groups three or four question-answer pairs from one dimension.
  • Anonymization replaces character and experimenter names before evaluation, but preserves the rest of the narrative content.
  • In 100 BFI cases, option conversion with GPT-4 reaches 71.0% accuracy and Pearson 0.600.
  • Dimension-specific conversion with GPT-4 improves to 82.0% and Pearson 0.847.
  • Batch expert rating with GPT-4 obtains 89.0% and Pearson correlations 0.925, Spearman 0.927, and Kendall 0.837.
  • In expert rating, 82 of 100 GPT-4 judgments are considered correct, 14 differ by exactly one point, and 4 differ by more than one point according to the appendix breakdown.
  • The main benchmark uses PDb types for accuracy and invited scores for continuous error in BFI and 16P.
  • With batch ER and GPT-4, BFI obtains 76.6% accuracy per dimension, 31.2% full, and MAE 18.2%.
  • With the same configuration, 16P obtains 80.7% per dimension, 44.8% full, and MAE 20.5%.
  • The full accuracy of 16P is 36 points lower than the per-dimension accuracy highlighted in the abstract.
  • In BFI, the best self-report in the table is around 63.7% per dimension and 7.3% full, below ER.
  • In 16P, self-report reaches at most 66.1% per dimension and 22.9% full in the main table.
  • d-OC with GPT-4 achieves 80.2% per dimension and 45.8% full in 16P, showing that the best full metric does not come from ER.
  • ER measurements show greater dispersion among characters in the PCA than self-report measurements.
  • The mean standardized variance of BFI profiles is 1.03 for batch ER, compared to 0.71 for SR and 0.68 for SR-CoT.
  • Across the 14 scales with batch ER and GPT-3.5, the published mean is 78.9% per dimension, 58.5% full, and MAE 8.1%.
  • The average of 78.9% combines scales with very different numbers of dimensions and difficulties.
  • Accuracy per dimension ranges from 68.2% in ECR-R to 93.1% in GSE within Table 12.
  • CABIN obtains 75.5% per dimension, but only 6.3% full hit due to its 41 dimensions.
  • The mean agreement among annotators is kappa = 0.609 across the 14 scales.
  • BFI and 16P show kappas of 0.759 and 0.770, while EPQ-R, ECR-R, and CABIN fall below 0.48.
  • With PDb labels, higher accuracy is obtained; with invited labels, lower MSE is obtained, so the two sources are not interchangeable.
  • Character descriptions alone achieve results close to combining descriptions and memories in several conditions.
  • RPAs based on GPT-3.5 and GPT-4 offer the highest personality fidelity among the compared families.
  • GPT-4 does not consistently outperform GPT-3.5 across all metrics or character groups.
  • Open models show important differences between English and Chinese characters; Llama-2-Chat 13B is competitive in English and weak in Chinese.
  • The incremental adjustment of RP-Mistral-2 reduces some role problems, but provides limited fidelity improvements against already strong bases.
  • The character.ai RPAs fall below comparable GPT-3.5 configurations in this benchmark.
  • In a sample of character.ai responses converted by GPT-4, 65.8% are agree or strongly agree and 17.1% disagree or strongly disagree.
  • Adapting seven potentially anachronistic 16P questions per character improves dimensional accuracy from 80.7 to 81.8 and full accuracy from 44.8 to 46.9.
  • General in-context examples worsen SR and SR-CoT, something the authors attribute to interference from non-personalized personalities.
  • The authors declare informed consent, compensation above the local minimum wage, and privacy protection for the annotators.
  • The current repository has an MIT license, includes the 14 questionnaires and aggregated labels of the 32 characters, and its Python files compile.
  • The repository does not contain results, interview dialogues, or individual annotations that allow recalculating the tables.
  • The article was published in ACL 2024, pages 1840-1873, DOI 10.18653/v1/2024.acl-long.102.

Limitations

  • The variable called personality is an agreement with external labels of the character, not a validation of the agent's internal personality.
  • The characters are fictional and their personalities do not have a single objective truth.
  • Personality Database is an open platform; the study does not control who votes, their knowledge of the character, duplicates, or fandom biases.
  • The selection requires availability and popularity in PDb, which favors well-known characters with stereotyped identities.
  • The 32 characters are chosen to cover PDb types and do not constitute a random sample of characters or RPAs.
  • The benchmark mixes franchises, narrative versions, and temporal snapshots; a static label may not represent the entire evolution of the character.
  • For BFI and 16P, PDb types are used for accuracy and invited scores for MAE, so one row is judged against two different references.
  • PDb percentages are converted to high above 60%, low below 40%, and marginal in between; the thresholds are a study decision.
  • Marginal dimensions are excluded from accuracy and may remove precisely the most ambiguous cases.
  • Only two or three invited annotators evaluate each character; 93 annotations cover 32 characters.
  • Annotators score 73 dimensions, an extensive workload that may introduce fatigue and superficial responses.
  • Individual annotations, disagreements, times, and rationales are not published; only aggregated averages.
  • The mean agreement kappa = 0.609 is moderate and three scales fall below 0.48.
  • ECR-R evaluates romantic relationships and CABIN professional interests that are rarely represented in the works, as the authors acknowledge.
  • The EPQ-R lie scale is ambiguous for characters and contributes to low agreement.
  • 16Personalities is a proprietary tool inspired by MBTI, not a validated equivalence of MBTI or a reproducible open scale.
  • The code sends the 60 responses to undocumented 16Personalities web endpoints as a public API, which may change or impose external conditions.
  • Transforming closed items into open-ended questions changes the administration and may alter the construct the scale measures.
  • Although the transformed questions are manually reviewed, no study of item-by-item semantic or psychometric equivalence is reported.
  • Each question is conducted in an isolated context; the procedure is not an adaptive interview with follow-up, clarifications, or shared history.
  • Expert rating simulates an evaluator using a LLM, not a clinical professional or psychometrician.
  • The evaluator validation uses only 100 BFI cases and is not repeated on the other thirteen scales.
  • An independent set of characters to validate the evaluator and then measure the main benchmark is not specified.
  • The correlations in Table 1 lack confidence intervals, hypothesis tests, and analysis by character or dimension.
  • The hit criterion of less than one point and half hit at exactly one point is a discrete rule whose sensitivity is not analyzed.
  • GPT-4 still produces four errors greater than one point in 100 expert rating cases.
  • Anonymization replaces names with text, but responses such as Liyue, Geo Archon, Hogwarts, or character quotes may reveal the identity to the evaluator.
  • A pretrained evaluator may know both the character and popular stereotypes and labels, producing agreement by memory rather than by evidence from the interview.
  • The textual substitution of aliases is naive, sensitive to capitalization and context, and may incorrectly delete or preserve fragments.
  • The code adds detected aliases to shared global lists, which may accumulate substitutions across runs.
  • Noncompliant self-report responses are mapped to neutral instead of being treated as missing data or method failures.
  • Interview results are regenerated when JSON fails and use temperature 0.2 in regenerations, a more favorable treatment than the forced neutral of self-report.
  • When Gemini fails repeatedly or blocks a response, the code falls back to GPT-4; a figure labeled as Gemini may include decisions from another model.
  • GPT-3.5 splits batches that exceed the token limit, so the evaluator context varies by model and scale.
  • The three repetitions are not accompanied by confidence intervals or statistical tests of differences between methods or models.
  • The claim that GPT-4 does not significantly outperform GPT-3.5 is not accompanied by a described significance test.
  • AccDim weights each evaluable dimension, AccFull requires hitting all of them, and MAE uses continuous scores; highlighting only one may change the conclusion.
  • The highlighted 80.7% is dimensional accuracy of a single 16P configuration and not full character fidelity.
  • The average of 78.9% across 14 scales assigns the same weight to instruments with one, two, five, or 41 dimensions.
  • Some one-dimensional scales have AccFull identical to AccDim, while CABIN suffers a combinatorial penalty, which limits the comparability of the full average.
  • The PCA and greater dispersion show differentiation between outputs, but a more dispersed personality does not imply greater validity.
  • The main RPAs depend on descriptions and memories curated by ChatHaruhi or RoleLLM; the quality of the data, the model, and the evaluation method are not completely separated.
  • Six characters use Chinese and the rest English; the sizes per language and franchise are small to generalize linguistic effects.
  • The API models and snapshots are from late 2023 and may have been withdrawn or modified.
  • The character.ai experiment depends on a closed system, private descriptions, and non-controllable conditions.
  • Question adaptation uses GPT-4 and knowledge of the character, introducing another layer of automatic judgment and possible information leakage.
  • The improvement from adaptation of 1.1 dimensional points and 2.1 full points does not include uncertainty.
  • It is not studied whether agents maintain the profile over long sessions, new events, changing memory, or continued relationships.
  • There is no human evaluation of the quality, naturalness, or fidelity of the open-ended responses generated in the main benchmark.
  • Scale scores are not linked to decisions or behaviors of the character in external tasks.
  • The analysis does not cover RPA safety, harm from stereotypes, impersonation, parasociality, or effects on users.
  • DTDD may provoke harmful content and the article acknowledges the need for safeguards, but does not implement or evaluate controls.
  • The ethics section does not report institutional approval, exclusion procedure, exact amount, or demographic characteristics of annotators.
  • The repository does not contain the 18,304 dialogues whose release the article and the project page announce.
  • It also does not contain per-repetition outputs or scripts that directly reconstruct all published tables and figures.
  • requirements.txt does not pin versions of most dependencies and there is no lockfile, containerized environment, or CI.
  • The file called test_rpa_methods.py runs experiments with external services; it is not an automated test suite.
  • Reproduction requires keys, external downloads, proprietary services, and endpoints that cannot be frozen in the artifact.
  • The code uses pickle caches and relative paths dependent on running from code, without integrity validation or provenance of results.
  • Various error paths enter interactive pdb or replace results with neutral values, which hinders unsupervised and auditable runs.

What the study does not establish

  • It does not demonstrate that RPAs possess personality, identity, emotions, cognition, or subjective experience.
  • It does not demonstrate that psychological scales maintain construct validity when their subjects are generative models.
  • It does not establish a true and unique personality for each fictional character.
  • It does not demonstrate that Personality Database labels are a bias-free gold standard.
  • It does not prove that 80.7% of characters are completely well represented.
  • It does not prove that 80.7% is general to other scales, evaluators, prompts, or models.
  • It does not demonstrate that expert rating is equivalent to an interview conducted by psychologists.
  • It does not eliminate dependence on a LLM judge or its biases, prior knowledge, and errors.
  • It does not demonstrate that anonymization prevents identifying the character from the response.
  • It does not separate agreement based on the interviewed content from agreement based on memorized knowledge of the character.
  • It does not causally establish that open-ended interviews are superior under equal conditions to self-report, given the asymmetric treatment of failures.
  • It does not demonstrate that greater PCA dispersion corresponds to greater psychological fidelity.
  • It does not demonstrate temporal stability or multi-turn consistency of the measured profile.
  • It does not demonstrate transfer from questionnaire scores to actions, decisions, prolonged dialogue, or relationships.
  • It does not demonstrate validity for less popular, original, historical characters, or real people.
  • It does not demonstrate robust generalization across English, Chinese, and other languages or cultures.
  • It does not establish that GPT-4 is always better than GPT-3.5 as an agent or evaluator.
  • It does not validate 16Personalities as a substitute for MBTI or its labels as a diagnosis.
  • It does not validate DTDD, EPQ-R, ECR-R, or CABIN for clinical inferences about agents or users.
  • It does not demonstrate that character.ai RPAs are globally worse; it evaluates a sample and a closed snapshot under this protocol.
  • It does not demonstrate that adapting questions materially improves all characters or scales.
  • It does not allow exactly reproducing the tables without the absent dialogues, results, individual annotations, and historical services.
  • It does not justify using the system for mental health, evaluation of people, education, employment, or high-impact decisions.
  • It does not demonstrate that the measured fidelity is safe, desirable, or beneficial for users.

Traceability

Scope: Full text

Version: ACL 2024 final proceedings paper, pp. 1840-1873, DOI 10.18653/v1/2024.acl-long.102, 34 pages; official code at paper-era commit cb2cfff96d6dd39c8b6fab2a470371c1f92ba25c and current commit f554202a94d4a83dc5407245bb18981899e872e6

Consulted source: https://aclanthology.org/2024.acl-long.102.pdf

Review: Codex full-text, bilingual-fidelity, visual, bibliographic, psychometric, LLM-judge, annotation-quality, experimental-design, metric-interpretation, code, reproducibility, privacy and ethics audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • gpt-3.5-turbo-1106 como modelo base de los RPA principales y como evaluador
  • gpt-4-1106-preview como evaluador y modelo base
  • Gemini Pro como evaluador
  • Qwen1.5-110B-Chat como evaluador
  • Qwen 7B
  • OpenChat 3.5 7B
  • Mistral-2 7B
  • Llama-2-Chat 13B
  • Mixtral 8x7B
  • CharacterGLM 6B
  • RP-Qwen 7B
  • RP-Mistral-2 7B
  • character.ai RPAs
  • ChatHaruhi
  • RoleLLM

Instruments and metrics

  • Big Five Inventory (BFI)
  • 16Personalities / NERIS Type Explorer
  • Bem Sex Role Inventory (BSRI)
  • Dark Triad Dirty Dozen (DTDD)
  • Experiences in Close Relationships-Revised (ECR-R)
  • Emotional Intelligence Scale (EIS)
  • Empathy Scale
  • Eysenck Personality Questionnaire-Revised (EPQ-R)
  • General Self-Efficacy Scale (GSE)
  • Implicit Culture Belief (ICB)
  • Love of Money Scale (LMS)
  • Life Orientation Test-Revised (LOT-R)
  • Wong and Law Emotional Intelligence Scale (WLEIS)
  • Comprehensive Assessment of Basic Interests (CABIN)
  • Cohen's quadratic-weighted kappa
  • Pearson, Spearman and Kendall agreement
  • Dimension accuracy, full accuracy and normalized MAE
  • Item-, dimension- and score-level standard deviation

Data used

  • 32 selected fictional-character configurations: 16 from ChatHaruhi and 16 from RoleLLM
  • Personality Database crowd labels for BFI and 16Personalities
  • 93 invited annotation sets covering 73 dimensions and 14 scales for 32 characters
  • 100 manually judged BFI cases for interviewer-LLM validation
  • ChatHaruhi character descriptions and memories
  • RoleLLM character descriptions and memories
  • ChatHaruhi-English-62K used for RP-Mistral-2 fine-tuning
  • Official repository questionnaires and aggregated characters_labels.json
  • Announced collection of 18,304 interview dialogues, not found in the audited official release

Evidence and location

  • Official bibliographic record: ACL Anthology 2024.acl-long.102: 13 authors, ACL 2024, pp. 1840-1873, DOI 10.18653/v1/2024.acl-long.102
  • Complete audited source: .cache/editorial-sources/article-094/source.pdf; official ACL PDF; 34 pages; sha256 85ba41b1a78e6b1f7e99b32757d63d96e9cf3b76eea33d821e3dd9d92b8ae098
  • Objective and contributions: Full text pp. 1840-1841, Abstract and Introduction
  • Two phases of InCharacter: Full text pp. 1841-1843, sections 3.1-3.2 and Figure 2
  • Open-ended questions and isolated contexts: Full text p. 1842, Constructing Question List and Interviewing RPAs
  • Option conversion, d-OC, and expert rating: Full text pp. 1842-1843, Assessment
  • Declared anonymization: Full text p. 1843, paragraph before Experimental Setup
  • Human validation of the evaluator: Full text p. 1843, section 4.1 and Table 1; 100 BFI cases
  • GPT-4 evaluator results: Full text p. 1843, Table 1; Appendix p. 1859, Table 8
  • Selection of 32 characters: Full text p. 1844, RPAs and Characters; Appendix p. 1855, Character Selection
  • Models and snapshots: Full text p. 1844, Interviewer LLMs and footnote 8
  • Fourteen scales: Full text p. 1844, Psychological Scales; Appendix pp. 1852-1854
  • Label sources: Full text p. 1844, Personality Labels; Appendix pp. 1857-1858, Human Annotations
  • Ninety-three annotations: Full text p. 1844 and Appendix p. 1857: 2-3 annotators per character, 93 annotation sets, 73 dimensions
  • Annotator agreement: Appendix p. 1857, Table 6: mean quadratic-weighted kappa 60.9%
  • Alignment and consistency metrics: Full text pp. 1844-1845, Metrics
  • Main comparison BFI and 16P: Full text pp. 1845-1846, Table 2
  • Difference between AccDim and AccFull: Full text p. 1846, Table 2: ERbatch GPT-4 16P 80.7 versus 44.8
  • Differentiation versus self-report: Full text pp. 1845-1846, Figure 3; Appendix p. 1859, section F.2
  • Results across fourteen scales: Full text pp. 1846-1847, Figure 4; Appendix p. 1861, Table 12
  • Data and base model comparison: Full text p. 1847, Table 3; Appendix pp. 1859-1860, Tables 9-11
  • Language sensitivity: Appendix p. 1860, Table 10
  • Compliant behavior of character.ai: Full text p. 1847; Appendix p. 1866, section H.2 and Table 20
  • Specific question adaptation: Appendix p. 1863, section F.7 and Table 16
  • ICL worsens self-report: Appendix pp. 1863-1864, section F.8 and Table 17
  • Treatment of rejections and retries: Appendix p. 1858, Implementation Details; official code personality_tests.py and utils.py
  • Recognized limitations: Full text p. 1848, Limitations
  • Annotation, consent, and risks: Full text pp. 1848-1849, Ethical Statement
  • Evaluation prompts: Appendix p. 1865, Table 19
  • Official code and audited commits: GitHub Neph0s/InCharacter, paper-era cb2cfff96d6dd39c8b6fab2a470371c1f92ba25c and current f554202a94d4a83dc5407245bb18981899e872e6
  • Scope of published artifact: Official repository tree contains 14 questionnaire JSON files, characters metadata and aggregated labels; no results directory, interview dialogues or individual annotations
  • Environment reproducibility: Official repository requirements.txt is largely unpinned; no lockfile, CI or automated unit tests; compileall passes
  • 16Personalities dependency: Official code api_16personality.py posts responses to 16personalities.com test-results and reads its session endpoint
  • Asymmetric error treatment: Official code personality_tests.py maps noncompliant self-reports to neutral; utils.py regenerates evaluator JSON and can fall back from Gemini to GPT-4
  • Comprehensive reading and visual inspection: All 34 pages rendered and inspected, including Figures 1-7, Tables 1-28, ethical statement, prompts and example responses; checked 15 Jul 2026