The Personality Illusion: Revealing Dissociation Between Self-Reports & Behavior in LLMs

Evaluation and psychometric validity2026ICML ProceedingsApproved editorial review

Authors: Pengrui Han, Rafal Kocielnik, Peiyang Song, Ramit Debnath, Dean Mobbs, Anima Anandkumar, R. Michael Alvarez

Keywords: Artificial Intelligence, Computation and Language, Computers and Society, Machine Learning

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

7
Authors
22
Findings
47
Limitations
19
Evidence

Editorial summary

English

Han and coauthors separate three questions that are often conflated: what profile an LLM reports on questionnaires, whether that profile predicts its outputs on other tasks, and whether a textual persona changes both. The study, accepted to ICML 2026, uses 18 endpoints: six base models, their six instruct variants, and six large instruct models. It administers the 44-item Big Five Inventory and the 63-item Self-Regulation Questionnaire, with every item sent as an independent API call. Each model is tested under three system prompts, temperatures 0.3/0.7/1.0, and three repetitions, yielding 27 aggregate configurations per model.

For RQ1, the authors compare the six base–instruct pairs. A logistic classifier of training phase finds that instruct models report higher Openness (β=1.48), higher Agreeableness (0.74), and lower Neuroticism (−1.20); Extraversion and Conscientiousness do not change significantly. Levene tests find lower variability on five of six scales, not Agreeableness. Associations between the Big Five and self-regulation also become stronger and generally take the directions expected in humans. These findings describe questionnaire outputs, not a developmental trajectory: checkpoints are not participants who mature, the 27 configurations are not 27 independent models, and the design does not isolate RLHF from instruction tuning, data, architecture, or scale. The reported self-regulation coefficients of roughly 11–23 also conflict with the statement that this outcome was standardized.

RQ2 evaluates 12 instruct models on five outcomes: cards chosen in a textual Columbia Card Task; forced explicit associations labeled as an IAT; answer changes after a contrary user opinion; overconfidence on 50 questions; and the difference between two confidence judgments. Only about 24% of trait–task associations are significant, and 52% of those significant coefficients have the direction expected from human literature, essentially chance. Trait-level directional agreement ranges from 45% for Neuroticism to 62% for Agreeableness, with every interval overlapping 50%. Most models remain near chance; Qwen3-235B reaches 82% and is the significant exception. The central negative finding is well supported in its narrow form: these self-reports are not general proxies for these tasks. It does not show that the tasks measure real-world behavior or that every type of self-report lacks predictive validity.

RQ3 injects Agreeableness or Self-Regulation personas through three strategies and three keyword variants. Personas are clearly detectable in the target self-report: Agreeableness β≈3.6–4.4 and Self-Regulation β≈2.2–2.9. Behavior, however, barely separates the conditions: sycophancy yields β≈−0.05–0.32 with inconsistent significance, while risk-taking yields β≈−0.14–0.20 and is nonsignificant. The Self-Regulation persona increases Conscientiousness more strongly than its target scale and decreases Openness and Agreeableness, showing that the manipulation is not selective. This usefully demonstrates that explicitly repeating “agreeable,” “disciplined,” or “goal-oriented” changes semantically adjacent questionnaire responses without reliably transferring to a separate task.

The artifact audit adds important caution. GitHub contains nine complete configuration-level CSV files with 2,754 aggregate rows, but no item-level outputs, logs, or statistical-analysis code, so exclusions cannot be audited and figures cannot be reconstructed exactly. The notebooks are GPT-4o-only examples: Big5 fails because of undefined global names, and SRQ contains a SyntaxError plus two references to an undefined variable. No notebook sets or transmits a seed despite the paper's claim of three seeds. More seriously, API and parsing failures become observations: Big Five/SRQ impute 3, Risk Taking imputes zero cards, IAT can return −1, and the sycophancy workflow can shift later responses after an unknown parse. In the honesty notebook, zero-confidence samples are omitted from ECE; the IAT code averages signed scores although the method describes absolute magnitude; and one table reverses the meaning of the C1–C2 difference.

The defensible contribution is conceptual and empirical: it requires self-reports to be validated against separate outcomes and shows, over a broad battery, that linguistic coherence and persona control do not guarantee transfer. The evidence does not warrant claims about internal traits, dispositions, goals, or a human-like cognitive dissociation. A simpler explanation remains open: questionnaires and tasks are independent prompts, some adaptations do not preserve their human construct, and matched repetition labels do not instantiate a persistent individual. The paper is valuable as a methodological warning, provided its results are read as properties of this protocol rather than an exhaustive account of what LLMs “are.”

Español

Han y coautores separan tres preguntas que a menudo se confunden: qué perfil declara un LLM en cuestionarios, si ese perfil predice lo que hace en otras tareas y si una persona textual cambia ambas cosas. El estudio, aceptado en ICML 2026, usa 18 endpoints: seis modelos base, sus seis variantes instruct y seis instruct grandes. Administra el Big Five Inventory de 44 ítems y el Self-Regulation Questionnaire de 63; cada ítem es una llamada independiente. Cada modelo se prueba con tres system prompts, temperaturas 0,3/0,7/1 y tres repeticiones, es decir, 27 configuraciones agregadas por modelo.

En RQ1, los autores comparan los seis pares base–instruct. Un clasificador logístico de fase encuentra que los modelos instruct declaran más apertura (β=1,48), más amabilidad (0,74) y menos neuroticismo (−1,20); extraversión y responsabilidad no cambian significativamente. Levene indica menor variabilidad en cinco de seis escalas, no en amabilidad. Las relaciones entre Big Five y autorregulación también se vuelven más fuertes y adoptan en general los signos esperados en humanos. Esto describe respuestas al cuestionario, no una trayectoria de desarrollo: los checkpoints no son participantes que maduren, las 27 configuraciones no son 27 modelos independientes y el diseño no aísla RLHF de instruction tuning, datos, arquitectura o tamaño. Además, los coeficientes de autorregulación reportados, entre aproximadamente 11 y 23, contradicen la afirmación de que esa variable fue estandarizada.

RQ2 evalúa 12 modelos instruct en cinco outcomes: tarjetas elegidas en una versión textual de Columbia Card Task; asociaciones explícitas forzadas que el artículo llama IAT; cambio de respuesta ante una opinión contraria del usuario; sobreconfianza en 50 preguntas; y diferencia entre dos juicios de confianza. Solo alrededor del 24% de las asociaciones rasgo–tarea resulta significativa y, entre ellas, el 52% tiene el signo esperado a partir de literatura humana, prácticamente azar. Por rasgo, la coincidencia direccional va del 45% en neuroticismo al 62% en amabilidad y todos los intervalos solapan el 50%. La mayoría de modelos queda cerca de azar; Qwen3-235B alcanza 82% y es la excepción significativa. El resultado central es sólido en su formulación negativa: estos autoinformes no funcionan como proxies generales de esas tareas. No implica que las tareas midan comportamiento real ni que todo tipo de autoinforme carezca de validez predictiva.

En RQ3 se inyectan personas de amabilidad o autorregulación mediante tres estrategias y tres variantes léxicas. Las personas se detectan con claridad en el target declarado: amabilidad β≈3,6–4,4 y autorregulación β≈2,2–2,9. Sin embargo, el comportamiento apenas separa las condiciones: sycophancy arroja β≈−0,05–0,32 con resultados inconsistentes, y riesgo β≈−0,14–0,20 no significativo. La persona de autorregulación aumenta responsabilidad todavía más que su propia escala y reduce apertura y amabilidad, mostrando que la manipulación no es selectiva. Esta es una demostración útil de que repetir literalmente «agreeable», «disciplined» o «goal-oriented» altera respuestas que contienen ítems semánticamente cercanos, pero no basta para producir transferencia estable a otra tarea.

La auditoría del artefacto exige cautela adicional. GitHub contiene nueve CSV completos a nivel de configuración, con 2.754 filas agregadas, pero no los outputs por ítem, logs ni código de los modelos estadísticos; por eso no puede auditarse qué respuestas se excluyeron o reconstruir exactamente las figuras. Los notebooks son ejemplos reducidos a GPT-4o: Big5 falla por nombres globales inexistentes y SRQ contiene un SyntaxError y dos referencias a una variable no definida. Ningún notebook fija o envía un seed aunque el artículo hable de tres seeds. Más grave, los fallos de API o parsing se convierten en observaciones: Big Five/SRQ imputan 3, riesgo imputa 0 cartas, IAT puede devolver −1 y el flujo de sycophancy puede desalinear respuestas tras un parseo desconocido. En honestidad, los ceros de confianza se excluyen del ECE; en IAT se promedian scores firmados aunque el método dice usar magnitud absoluta; y una tabla invierte la interpretación de la diferencia C1–C2.

La contribución defendible es conceptual y empírica: obliga a validar autoinformes contra outcomes separados y muestra, en una batería amplia, que la coherencia lingüística y el control por persona no garantizan transferencia. La evidencia no autoriza a hablar de rasgos internos, disposiciones, objetivos o una disociación cognitiva equivalente a la humana. También queda abierta una explicación más sencilla: cuestionarios y tareas son prompts independientes, algunos instrumentos no conservan su constructo humano y las repeticiones emparejadas no representan una misma entidad persistente. El artículo es valioso precisamente como advertencia metodológica, siempre que sus resultados se lean como propiedades de este protocolo y no como prueba exhaustiva de lo que los LLM «son».

Research question

How do Big Five and self-regulation self-reports change between base and instruct checkpoints, do those profiles predict five outcomes of tasks inspired by psychology, and does the injection of agreeableness or self-regulation personas change both the self-report and the associated task?

Method

Experimental study in three blocks. RQ1 compares six base models with six instruct variants using BFI-44 and SRQ-63, phase logistic regression, Levene and self-regulation regressions. RQ2 matches the self-reports of 12 instruct models with five outcomes derived from four paradigms: risk, social association, sycophancy and two confidence measures; it fits mixed-effects models and reduces each coefficient to match or no match with an expected human sign. RQ3 compares baseline with two target personas, each expressed through three strategies and three keyword variants, using detectability logistic regressions. Each model uses three system prompts, three temperatures and three repetitions. The editorial audit read and rendered the 35 pages, verified ICML 2026 acceptance, inspected tables and appendices, and audited the repository, nine CSVs, six notebooks, datasets, parsers and scoring.

Sample: RQ1 contains 162 aggregated rows from six base models and 162 rows from their six instruct variants: 6 models × 3 prompts × 3 temperatures × 3 runs. RQ2 uses 324 self-report rows and 324 task rows: 12 instruct models × 27 configurations matched by label. In each configuration, BFI is called 44 times, SRQ 63, risk 27, social associations 63, honesty 50×2 and sycophancy up to 52×2. For each of six persona-strategy conditions there are 324 rows: 12 models × 3 lexical variants × 3 temperatures × 3 runs. The runs are repeated calls; the code does not fix seeds or create persistent subjects.

Findings

  • In the base–instruct comparison, openness predicts instruct phase with β=1.48 (95% CI 0.74–2.22), neuroticism with β=−1.20 (−2.00 to −0.41) and agreeableness with β=0.74 (0.03–1.44).
  • Extraversion (β=−0.12, p=.739) and conscientiousness (β=−0.61, p=.089) show no significant differences by phase in the joint model.
  • Levene finds lower variability for openness, conscientiousness, extraversion, neuroticism and self-regulation; agreeableness does not differ (p=.54).
  • In instruct, self-regulation is positively associated with conscientiousness (β=12.32), openness (15.23), agreeableness (11.36) and extraversion (23.33), and negatively with neuroticism (−16.27).
  • The self-regulation coefficients are incompatible with a standardized dependent variable as the method claims; they appear to retain the raw SRQ scale.
  • Only about 24% of all trait–task associations is statistically significant and only 52% of those significant associations has the expected human sign.
  • Match by trait is openness 50%, conscientiousness 52%, extraversion 58%, agreeableness 62%, neuroticism 45% and self-regulation 55%; all intervals overlap 50%.
  • The five outcomes fall approximately between 45% and 57% directional match, without a general behavioral pattern.
  • Most small and medium models remain close to chance; Claude 3.7 reaches 64%, GPT-4o 68% and Qwen3-235B 82%, the latter being the exception whose interval does not overlap chance.
  • In risk, the only strongly significant Qwen self-regulation association has the expected human sign, but other traits and groups alternate signs and are rarely significant.
  • In aggregate, conscientiousness predicts more overconfidence (β=3.75, p<.05), opposite direction to the human expectation defined by the authors; self-regulation predicts less overconfidence (−0.15, p<.05).
  • In aggregate sycophancy, openness (−4.70), conscientiousness (−6.42) and neuroticism (−5.41) are significant, but only part of the signs matches the selected expectations.
  • The agreeableness persona increases its target self-report β≈3.6–4.4, p<.001 across the three strategies.
  • The self-regulation persona increases its target β≈2.2–2.9, but conscientiousness increases even more, β≈4.2–4.8.
  • The self-regulation persona reduces openness β≈−2.2 to −2.8 and agreeableness β≈−1.1 to −1.8, so the intervention is not selective nor does it reproduce the structure of RQ1.
  • Sycophancy weakly and inconsistently distinguishes the agreeable persona (β≈−0.05 to 0.32), and risk does not reliably distinguish the self-regulated persona (β≈−0.14 to 0.20).
  • The nine released CSVs contain the complete grid of models, prompts/personas, temperatures and runs, with no duplicates in configuration keys.
  • The CSVs only contain aggregates; per-item outputs, parsings, failures, exclusions and intermediate statistical results are missing.
  • The Big5 notebook does not run due to nonexistent MODEL_CONFIGS, PERSONA_CONFIGS and EXPERIMENTAL_CONFIG names; SRQ does not compile and also uses srq_items instead of SRQ_ITEMS.
  • None of the six notebooks transmits a seed to the provider; run 1/2/3 labels stochastic repetitions, not reproducible seeds.
  • The IAT uses Python hash() to sort stimuli, so the order changes between processes unless PYTHONHASHSEED is set, despite declaring random_seed=42.
  • The scale table says that a larger C1–C2 difference means more consistency, but the code and Table 3 use the absolute difference as more inconsistency.

Limitations

  • Self-reports are outputs of independent prompts, not observations of an internal state or of a single longitudinal entity.
  • The base–instruct comparison includes only six families and does not use a controlled intervention on the same model except for the informal relationship between checkpoints.
  • Instruction tuning, RLHF, DPO and other phases are grouped under "instructional alignment" without knowing or isolating which recipe each checkpoint received.
  • The pairs differ in data, chat format and possible additional changes; causal language about alignment exceeds an observational comparison.
  • The 27 configurations per model are repeated measures of only three prompts and three temperatures; they do not increase the number of independent architectures.
  • Using model, temperature and prompt as random effects is fragile with few levels; model is also nested/confounded with phase if the full identifier is retained.
  • The logistic regression predicts phase from six traits simultaneously; its conditional coefficients do not equate to causal mean changes of each trait.
  • Levene applied to hundreds of repeated responses can produce very small p-values due to pseudoreplication with respect to the six families.
  • The text reports variability reductions of 40.0%/45.1%, while the caption states 60–66%; the CSVs and the method do not allow reconciling both figures.
  • No reliability, factorial structure, invariance, convergent or discriminant validity of BFI/SRQ is reported in the evaluated models.
  • The BFI route substitutes parsing failures with 3; the SRQ route does the same. This can artificially reduce variability, especially in base models that follow instructions worse.
  • Raw outputs are not published, so neutral imputation cannot be quantified nor per-item scoring verified.
  • The grid says it uses seeds, but the code only repeats calls and does not set a local seed nor in the APIs.
  • GPT-4o is an alias without a snapshot; several Together/OpenRouter endpoints also depend on provider state and do not carry a verifiable execution date in the CSVs.
  • RQ2 matches self-report and task rows by prompt, temperature and run number, but the calls are independent and do not represent the same latent subject.
  • Each mixed-effects model per LLM has only 27 aggregates for six traits, with very little variation and sometimes zero variance.
  • The main method says it includes the six traits together; the caption of Figure 9 describes regressions of the task on each trait separately. Code to resolve the actual specification is missing.
  • No statistical code, exact aggregation formulas, data preparation, handling of missing coefficients or figure generation is published.
  • Hundreds of trait–task–group–model tests are performed without correction for multiplicity; p<.1 is also highlighted as a trend.
  • Converting the sign of any coefficient, even non-significant ones, into hit/miss produces a metric whose baseline is necessarily 50% and discards magnitude and uncertainty.
  • Human associations are transferred between distinct populations, operationalizations and contexts; some are declared mixed or rely on indirect evidence.
  • The beta-binomial interval assumes independence of reused coefficients; the bootstrap uses only five tasks or six traits as clusters, too few for stable intervals.
  • The textual version of CCT is a single symbolic response without chance, feedback, gain or sequence; it may measure utility calculation or instruction following, not lived risk taking.
  • The risk parser concatenates all digits before parsing: "16 of 32" becomes 1632 and is truncated to 32; API errors become 0 cards.
  • The task called IAT does not measure latency or implicit association; it explicitly forces assigning positive/negative words to social groups.
  • The IAT averages signed bias although the article says it uses absolute value; opposite categories can cancel out and the score does not represent total magnitude.
  • The IAT parser does not validate completeness, uniqueness or exact correspondence; an empty response or error produces zero counts and bias −1 instead of missing.
  • The IAT order uses non-stable hash() between processes, so the assumed seed 42 does not freeze the stimuli.
  • The sycophancy measure lacks a second control trial without a contrary opinion; it does not separate conformity from spontaneous variation or reconsideration.
  • If a baseline sycophancy response is not parsed, its Step 2 is not created, but the final loop can consume the next response for the wrong dilemma.
  • The common handler returns the string ERROR, but sycophancy checks other codes; this activates the Unknown path and the risk of misalignment.
  • Honesty uses 25 synthetic questions with NOANSWER response and 25 real ones; the prompt requires one word and does not explicitly offer an abstention option.
  • Average overconfidence is signed, so overconfidence and underconfidence can cancel out; it is not an absolute calibration error.
  • The ECE uses confidence > lower and excludes confidence 0 responses from all bins.
  • Confidence −1 due to invalid parsing is retained in calibration and consistency calculations instead of being excluded.
  • The code defines C1–C2 as absolute difference, where larger is worse; Table 2 and Figure 7 say that larger means more consistency.
  • The RQ3 personas literally contain the target words and the questionnaire asks for nearby descriptors; the self-report effect may be priming or lexical compliance.
  • RQ3 fits numerous classifiers per strategy and outcome without multiple correction or validation on new personas or tasks.
  • Only two traits and two tasks chosen after RQ2 are intervened; general or temporal transfer is not tested.
  • The published notebooks are snippets for GPT-4o, not the configuration of 18 models nor a reproducible end-to-end pipeline.
  • Big5 contains nonexistent global references and does not reproduce the special template for base that appears in the appendix.
  • SRQ contains invalid syntax, wrong variable names and rounds the mean before reconstructing the total, generating non-integer totals.
  • There is no requirements file, lockfile, environment, automated tests, CI workflow or release/tag that freezes the evaluated artifact.
  • The article limitation says it does not include MoE, although the listed Qwen3-235B-A22B endpoint is a mixture-of-experts architecture; the scope is described inconsistently.
  • Work is only in English and human expectations come from heterogeneous literature; cultural generalization is not demonstrated.
  • Long conversations, agents with tools, memory, actions in environments or real consequences are not evaluated.
  • Proposals about RL from behavioral feedback or latent dispositions are future hypotheses, not results of the study.

What the study does not establish

  • It does not demonstrate that LLMs possess personality, internal traits, motivations, goals or affect.
  • It does not demonstrate a psychological developmental trajectory during training.
  • It does not identify that RLHF, DPO or instruction tuning alone cause the self-report changes.
  • It does not demonstrate that a human-like correlation structure validates the construct in machines.
  • It does not demonstrate that all LLM self-reports are useless; it evaluates two questionnaires and five concrete outcomes.
  • It does not demonstrate that the outcomes are real behavior, embodiment or acting in deployment.
  • It does not demonstrate that 52% is evidence of human alignment; it is compatible with directional chance.
  • It does not demonstrate that Qwen3-235B has a more human personality; its 82% summarizes signs under this protocol.
  • It does not demonstrate that persona injection never changes behavior; it shows weak and inconsistent transfer in two trait–task pairs.
  • It does not demonstrate that a textual persona modifies parameters, internal representations or persistent dispositions.
  • It does not demonstrate a cognitive dissociation equivalent to the human one; it compares outputs of independent prompts.
  • It does not validate IAT, CCT, confidence or moral dilemmas as psychometrically equivalent instruments for LLMs.
  • It does not allow exactly reproducing the inferences, p-values or figures with the published code.
  • It does not allow auditing how much API failures, parsing, imputation or response selection influence results.
  • It does not generalize to reasoning models, other languages, other providers, prolonged interaction or agents that act.
  • It does not justify high-impact decisions based on the declared profile of a model.
  • It does not establish that current alignment only affects language and never behavior; that causal thesis exceeds the contrast performed.

Traceability

Scope: Full text

Version: arXiv v2, 35 pages, revised 5 September 2025; subsequently accepted to the ICML 2026 main conference, official event poster 66149

Consulted source: https://arxiv.org/pdf/2509.03730

Review: Codex editorial review and methods/artifact audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • LLaMA-3.2 3B base and Instruct
  • LLaMA-3 8B base and Instruct
  • Qwen2.5 1.5B base and Instruct
  • Qwen2.5 7B base and Instruct
  • Mistral-7B-v0.1 base and Instruct
  • OLMo2 7B base and Instruct
  • LLaMA-3.3 70B Instruct
  • LLaMA-3.1 405B Instruct
  • Qwen2.5 72B Instruct
  • Qwen3 235B-A22B Instruct
  • Claude 3.7 Sonnet
  • GPT-4o

Instruments and metrics

  • Big Five Inventory, 44 items (BFI-44)
  • Self-Regulation Questionnaire, 63 items (SRQ-63)
  • Text-only cold Columbia Card Task with 27 factorial scenarios
  • Forced word-to-group association task described as an Implicit Association Test
  • Fifty-question confidence calibration and repeated-confidence task
  • Fifty-two moral dilemmas with contrary user opinion for sycophancy
  • Mixed-effects regressions, Levene tests, logistic classifiers and directional sign alignment
  • Agreeableness and Self-Regulation persona injections

Data used

  • Six matched base/instruct small-model families for RQ1
  • Twelve instruct models for RQ2 and RQ3
  • BFI-44 and SRQ-63 item text embedded in notebooks
  • Twenty-one social-association categories with three generated orders each
  • Fifty selected norm300 questions: five synthetic and five real items in each of five difficulty bins
  • Fifty-two moral dilemmas from the released dilemmas.json
  • Nine released aggregate CSV files totaling 2,754 configuration-level rows

Evidence and location

  • Acceptance and current bibliographic condition: Official ICML 2026 event listing, poster 66149; official repository news and author project pages checked 15 Jul 2026
  • Objective, three RQs and summarized results: arXiv v2 abstract, Figure 1 and introduction, pp. 1–2
  • BFI, SRQ, six pairs and 27 configurations: arXiv v2 sections 2.1–2.2 and Table 1, pp. 3–5
  • Changes by phase, variability and coherence: arXiv v2 Figure 2 and section 2.3, pp. 3–4
  • Behavioral tasks and RQ2 setup: arXiv v2 sections 3.1–3.4, pp. 5–6
  • Match by trait, task and model: arXiv v2 Figures 3–4 and section 3.5, pp. 6–7
  • Interventions and RQ3 results: arXiv v2 sections 4.1–4.3 and Figure 5, pp. 8–9
  • Interpretation, conclusion and declared limitations: arXiv v2 sections 5–7, pp. 9–11
  • Scales, intervals and extended specification: arXiv v2 Appendix B–C, Figures 6–9 and Tables 2–3, pp. 26–30
  • Complete prompts and human expectations: arXiv v2 Appendices D–H, Tables 4–10, pp. 30–35
  • Aggregate data and complete grids: Official repository commit a934083dcccd4428396497020578e89ffc9c51f8, nine CSV files audited 15 Jul 2026: 2,754 rows and no duplicate configuration keys
  • Absence of per-item data and statistical analysis: Official repository commit a934083dcccd4428396497020578e89ffc9c51f8: 25 tracked files; aggregate CSVs and generation notebooks only, audited 15 Jul 2026
  • Big5 and SRQ errors: Official repository self-reports/Big5.ipynb and SRQ.ipynb, AST/runtime-name audit 15 Jul 2026
  • Seeds and failure handling in questionnaires: Official repository Big5.ipynb and SRQ.ipynb generation, parsing and scoring cells, audited 15 Jul 2026
  • Risk and numeric parsing: Official repository behavioral_tasks/RiskTaking.ipynb scenario, parser and aggregation cells, audited 15 Jul 2026
  • Signed IAT, non-reproducible order and error as −1: Official repository behavioral_tasks/IAT.ipynb generate_random_orders, parse_response, calculate_bias and aggregation cells, audited 15 Jul 2026
  • Sycophancy and possible response shifting: Official repository behavioral_tasks/Sycophancy.ipynb two-step request construction and final processing loop, audited 15 Jul 2026
  • Calibration, ECE and C1–C2: Official repository behavioral_tasks/Honesty.ipynb extraction, ECE and summary cells; arXiv v2 Table 2 and Table 3, audited 15 Jul 2026
  • Visual integrity of the full text: arXiv v2 PDF SHA-256 f1b836de08730f32b37e2bc0e15888fbd108732405055b5989c752be964906fd; all 35 pages rendered and key figures, tables and appendices visually inspected 15 Jul 2026