PersonaScienceBench: Evaluating Scientific Agents with Diverse Simulated Researchers

1 Introduction

• Researchers may hold context needed to judge an analysis, while existing scientific benchmarks typically provide complete specifications and interactive benchmarks usually hold collaborators fixed.The benchmark gap motivates observing how agents clarify requirements, respond to feedback, and revise work as well as whether they succeed.Source: §1 · p. 1
• PersonaScienceBench adapts 200 tasks from five benchmarks into collaborative tasks while preserving source-specific deliverables and grading.Only the researcher simulator receives a complete persona profile; the actor encounters preferences through conversation, while evaluator references remain private.Source: §1 · p. 2
• Scientific outcomes and interaction metrics separately capture task performance, review rounds, revisions, clarification, message length, and language adaptation.Together, these measures can distinguish successful completion from intermediate progress such as recovering missing information in an episode that still ends unsuccessfully.Source: §1 · p. 2
• Across eighteen model configurations, the benchmark evaluates scientific performance, fact recovery, and self-repetition, with persona analyses revealing behaviors that final scores can obscure.Case studies illustrate differences in communication volume, targeted follow-up questions, and workflow revision.Source: §1 · p. 2

Language Models with Personas

• Persona conditioning and user simulation produce varied responses, while structured MatrAIx profiles span background, capability, and behavior.Prior non-collaborative simulators and persona policies show that varying user behavior can change tool-agent performance in service domains.Source: §2 · p. 3

Interactive Agent Evaluation

• Interactive benchmarks test clarification, escalation, and adaptation, but generally hold collaborators fixed or study non-scientific service domains.PersonaScienceBench varies the researcher while retaining a graded scientific objective, enabling joint analysis of collaborator-conditioned behavior and scientific outcomes.Source: §2 · p. 3

Agents for Scientific Discovery

• Scientific benchmarks assess task performance and information recovery, but generally leave collaborator behavior fixed.PersonaScienceBench varies simulated researchers on matched tasks and measures interaction patterns alongside source-specific outcomes.Source: §2 · p. 3

3.1 Persona Construction

• The Persona8B coreset combines 599,847 human-grounded and 400,000 synthetic records across a schema of 1,290 categorical dimensions.Synthetic records use dependency-graph sampling and compatibility rules; human-grounded records map source materials into the same schema, leaving unsupported fields unassigned.Source: §3.1 · p. 3
• Four profiles are selected by coverage across communication, decision-making, interaction, tool-collaboration, and workflow attributes, alongside a persona-off setting.The complete rendered populated profile goes only to the researcher simulator; the actor and evaluator never receive it, and the four profiles characterize sensitivity rather than a researcher-population average.Source: §3.1 · pp. 3–4

3.2 From Static Benchmarks to Interactive Simulations

• Required checkpoints are distinguished from voluntary clarification and revision, and longer conversations are not rewarded.The conversion aims to expose consultation and review while preserving a well-defined scientific task.Source: §3.2 · p. 4
• The benchmark adapts 200 tasks, with 40 each from five scientific benchmarks, covering four domains and computational workflows rather than image processing.The task set includes numerical workflows and diagnostic plots; image-processing tasks are outside the study’s scope.Source: §3.2 · p. 4, Table 2 · p. 4
• Task conversion separates public inputs, researcher context, and private evaluation, then runs an actor–researcher conversation under a selected protocol.The actor submits its work for evaluation after the protocol ends, while evaluator references remain private.Source: §3.2 · p. 4
• Four benchmarks use approach and result reviews, while SciConvBench retains its native specification-clarification dialogue.Source-specific deliverables and grading are retained within declared adaptations, including an EXP-Bench design-only track and researcher-held methodological guidance for ASI-Bench.Source: §3.2 · pp. 4–5

4.1 Models, Persona Settings, and Evaluation Protocol

• Eighteen model configurations use the same model for the actor and researcher simulator, and GPT-6 Astra powers model-based evaluations.Each model is tested in four persona settings and a persona-off control on the same 200 questions; most settings have one trial per question, with specified Astra settings run up to three times.Source: §4.1 · p. 5
• The 95% confidence intervals for Astra performance and self-repetition describe variation across sampled studies conditional on recorded trials, not repeated-rollout variability.Intervals use 20,000 study-cluster bootstrap resamples within benchmarks, with shared resampling draws across settings.Source: §4.1 · p. 5

4.2 Evaluation Metrics

• Scientific performance normalizes benchmark scores to 0–100 and averages question scores within benchmarks, then averages across the five benchmarks.Mean Persona Performance averages the four persona settings, with persona-off reported separately.Source: §4.2 · p. 5
• Fact recovery measures the percentage of required withheld facts recovered in SciConvBench disambiguation episodes.It is defined for cases with a nonempty set of annotated required facts.Source: §4.2 · p. 5
• Self-repetition measures lexical reuse by matching word windows in an actor message against windows in earlier eligible actor messages.Messages must contain at least nine word tokens, and episodes require at least two eligible messages.Source: §4.2 · p. 5

5.1 Model-Level Scientific Collaboration

• Gemini 3.7 Flash scored 48.17 versus Gemini 3 Flash at 31.98, while Gemini 3 Flash had higher fact recovery (90.01% versus 75.24%); GPT-5 mini had the lowest self-repetition at 10.52%.Across configurations, performance and interaction measures showed contrasting strengths; GPT-5 mini’s persona mean rose from 26.37 with persona off to 31.63 across four personas, a 5.26-point difference.Source: §5.1 · p. 6, Table 3 · p. 6

5.2 Performance across Scientific Domains

• Persona 001 scored above persona 046 in mathematics and computer science (54.15 versus 50.76), but below it in social sciences and humanities (16.86 versus 18.35).Persona 003 had the highest observed mean throughout; domain-specific rankings show why an overall persona ranking can obscure relative strengths.Source: §5.2 · p. 6, Table 4 · p. 7

5.3 Semantic Diversity in Responses and Interaction Paths

• All ten persona pairs showed positive semantic separation for both simulator and actor responses, including actors who encountered profiles only through conversation.The comparisons were significant after Holm adjustment across 20 question-wise tests; selected trajectories illustrate distinct message paths, but UMAP axes and distances are not comparable across plots or effect sizes.Source: §5.3 · pp. 7–8, Figure 3 · p. 7

5.4 Communication Patterns across Researcher Profiles

• Detailed-communication personas 155 and 003 were associated with 1.12 and 0.61 thousand additional actor tokens per episode, respectively, while clarification recall and tool use showed no significant pairwise differences.All four comparisons of simulator and actor output under personas 155 and 003 were significant after Holm adjustment; clarification recall varied from −0.28 to 4.55 percentage points and tool use from −0.54 to 0.28 calls per episode.Source: §5.4 · p. 8, Figure 4 · p. 8

5.5 Interaction Behavior beyond Final Scores

• In failed SciConvBench disambiguation episodes, complete-fact recovery ranged from 20.0% to 41.7% across personas, versus 0% with persona off.Persona-conditioned episodes with complete recovery involved four to ten iterations, showing information-gathering progress that final scores alone would not report; self-repetition also varied, from 8.03% under persona 003 to 15.94% under persona 046.Source: §5.5 · p. 9, Figure 5 · p. 9

6 Discussion

• In BixBench, researcher feedback prompted replacing unavailable R/limma with Python while preserving the objective, although the matched persona-off conversation reached the same answer through a different analysis.The researcher required replacement tests to be labeled unmoderated rather than limma-trend; the collaboration difference was visible in the interaction, not the final score.Source: §6 · p. 9, Figure 6 · p. 9

7 Conclusion

• PersonaScienceBench adapts 200 tasks from five scientific benchmarks for diverse simulated researchers, keeping objectives and grading fixed while capturing interaction differences alongside task performance.Persona-conditioned interactions produced distinct responses, communication patterns, and workflows that final scores alone did not capture.Source: §7 · p. 10

8 Limitations

• The 200 tasks cover text, data analysis, and code execution, not image processing, physical experiments, or extended projects with multiple researchers.Episodes involve one actor and one simulated researcher; results describe only the tested computational workflows and interaction protocols.Source: §8 · p. 10
• The four profiles span contrasting attributes rather than representing real researchers, and their bundled traits prevent attributing differences to a particular trait or demographic category.Language, message length, and endpoint function may also contribute to semantic differences; selected cases illustrate possible behaviors, not prevalence.Source: §8 · p. 10
• Intervals are conditional on recorded trials, while coupled actor–simulator models, a shared model-based judge, and separate interaction metrics constrain interpretation.A common judge does not eliminate grading error; lexical repetition and embedding separation alone establish neither semantic collapse, collaboration quality, nor scientific validity.Source: §8 · p. 10

9 Ethical Considerations

• Simulated profiles are not faithful representations of specific people or demographic groups, and profile-condition differences should not rank or guide decisions about real researchers.The benchmark is intended to assess agent robustness and adaptation across collaborator conditions.Source: §9 · p. 10
• Simulators may give incorrect or inappropriate scientific guidance, and benchmark traces require independent expert review before use as scientific, medical, or operational guidance.Separating grounded context from optional advice and keeping evaluator references private do not guarantee scientific validity.Source: §9 · p. 10
• Use is limited to released structured persona representations; releases must preserve source attribution, licensing requirements, and privacy protections.The paper does not claim that generated simulators reconstruct any particular person’s identity, beliefs, or behavior.Source: §9 · p. 10

A.1 General Principles and Conversion Rubric

• Review checkpoints let researchers request clarification or revision, while required handoffs are protocol events rather than evidence of spontaneous initiative.The protocol does not require extra questions or reward longer conversations; SciConvBench keeps its native clarification dialogue.Source: §A.1 · pp. 14–15
• Conversions retain each task’s identity, objective, deliverables, and evaluation contract while separating public inputs, researcher context, and evaluator-only references.Facts are provenance-linked and typed as context, fixed requirements, or optional advice; persona identity is separated from task context.Source: §A.1 · p. 14
• Conversions preserve benchmark-specific information asymmetries and deliverable types, and acceptance requires inspectable evidence rather than a score for the actor.All attempts and failure statuses remain recorded; a schema-valid package does not establish scientific adequacy or runtime readiness.Source: §A.1 · pp. 15–16, Table 5 · p. 16

A.2 DiscoveryBench

• DiscoveryBench retains a data-supported natural-language hypothesis as its deliverable, scored by agreement with the reference hypothesis rather than execution correctness.The score combines context recall with the mean product of variable F1 and relation agreement; accompanying analysis artifacts are recorded separately.Source: §A.2 · p. 16

A.3 BixBench

• BixBench retains its open-ended question and data capsule, excludes answer-bearing materials, and requires an executed analysis declaration plus a script or notebook for submission.Its retained verifier scores answer correctness; the analysis artifact and interaction trace separately support examination of how the answer was obtained.Source: §A.3 · pp. 16–17

A.4 EXP-Bench

• EXP-Bench is design-only: actors submit experimental designs for reference-coverage evaluation, without executing experiments or producing empirical conclusions.The coverage score and monitor-gated score are recorded separately; these measure reference coverage, not successful execution or reproduction.Source: §A.4 · p. 17

A.5 SciConvBench

• SciConvBench retains native clarification tasks, with the simulator holding the complete or corrected specification and the actor resolving request gaps or inconsistencies.Evaluation separately records recovery, intent-gated outcomes, clarification precision and recall, and memory consistency; no scientific solver runs.Source: §A.5 · p. 17

A.6 ASI-Bench

• ASI-Bench is a declared collaborative variant: the actor receives B3 while the researcher may provide optional methodological guidance drawn from B2.It therefore measures B3 with researcher-held B2 consultation, not autonomous performance under the original B3 condition.Source: §A.6 · pp. 17–18

B Analysis Coverage and Uncertainty

• 18 model configurations were evaluated on the same 200 questions across four profiles and persona off, with source-specific execution checks and GPT-6 Astra model-based judges.Each question has one evaluation per setting, except specified repeated GPT-6 Astra trials.Source: B Analysis Coverage and Uncertainty · p. 18
• Table 3 weights five benchmarks equally, whereas Table 4 weights four scientific domains equally, so their overall means need not match.GPT-6 Astra’s reported intervals characterize study-level variation conditional on recorded trials, not variability across independent rollouts.Source: B Analysis Coverage and Uncertainty · p. 18
• Interaction metrics have metric-specific eligibility: fact recovery is limited to annotated SciConvBench disambiguation cases, and self-repetition requires at least two eligible actor messages.Semantic comparisons require nonempty role messages for the same questions across settings.Source: B Analysis Coverage and Uncertainty · p. 18
• Self-repetition uses normalized, tokenized actor messages and a selected six-token matching window; its pointwise intervals do not account for exploratory selection.Necessary terminology, repeated facts, language, and message length can all contribute to measured overlap.Source: B Analysis Coverage and Uncertainty · p. 18
• Cross-model comparisons are descriptive because reported metrics use heterogeneous estimation procedures, including collaborator-supplied summaries and differing task or episode samples.Some configurations exclude unresolved execution errors, and reported intervals and aggregation procedures are not harmonized.Source: B Analysis Coverage and Uncertainty · pp. 18–19

C Semantic Analysis Details

• Semantic comparisons retain questions with nonempty final messages for both roles under all five settings, from the shared 200-question evaluation set.An endpoint is each role’s last nonempty generated message, which may be a closing message, handoff, question, or specification rather than a successful submission.Source: C Semantic Analysis Details · p. 19
• The statistic subtracts average within-persona cosine distance from mean cross-persona distance, equally weighting domains and measuring excess separation rather than omnibus distributional differences.Messages are separately embedded with frozen, 1,024-dimensional Qwen3-Embedding-0.6B representations, retaining generated language and style.Source: C Semantic Analysis Details · p. 19
• Persona-pair tests swap labels within each question using 10,000 permutations and Holm-adjust across 20 comparisons; bootstrap intervals are pointwise, not multiplicity-adjusted.Whole-question bootstrap resamples are stratified by benchmark and domain; shared-study dependence beyond question is not modeled.Source: C Semantic Analysis Details · p. 19
• Because generated language and style remain in the embeddings, this analysis does not isolate language-independent behavioral differences.The four profiles follow the selection strategy used in performance analyses.Source: C Semantic Analysis Details · p. 19
• BixBench and SciConvBench trajectories were manually selected from a gallery enriched for longer interactions, and each UMAP map has its own coordinates.Axes, distances, and path lengths cannot be compared across maps or interpreted as effect sizes; examples are illustrations, not independent statistical evidence or measures of task quality or simulator realism.Source: C Semantic Analysis Details · p. 19

D Prompts

• The prompt templates summarize implemented instructions rather than full runtime prompts; researcher personas go only to the simulator, and examples paraphrase shared messages through termination.Actor handoffs for four benchmarks include stage, answer, evidence, assumptions, and uncertainties; SciConvBench keeps its native question-and-answer interface.Source: D Prompts · pp. 19–20

D.1 Example Persona Profile: Persona 155

• Persona 155 is one of four profiles in the current analysis, supplied only to the simulator as synthetic attributes rather than verified biography or task requirements.The reproduced profile includes an agronomy specialty and high detail orientation.Source: §D.1 · pp. 20–21

D.2 Case-Study Profile: Persona 127

• Persona 127 is an English-language synthetic profile used for case studies, not one of the five evaluation settings.Its identity block was injected in the GPT-6 Astra run and is not a set of verified biographical facts or task requirements.Source: §D.2 · p. 21

D.3 DiscoveryBench

• The DiscoveryBench example proposes standardized PCA for 39 time slices and 12 capital indicators, reporting loadings while retaining uncertainty about the original specification.Reported PC1 loadings were 0.429, 0.507, and 0.570 for house size, daggers, and hatchets/swords; the original PCA specification remained undocumented.Source: §D.3 · p. 23

D.4 BixBench

• Dentate gyrus had 194 qualifying genes, versus 88 in baseline blood and 14 in final blood, using signed log2FoldChange > 0.5, baseMean > 10, and adjusted p < 0.05.The comparison labels and contrast direction were verified; the example retained the executed analysis script as evidence.Source: §D.4 · p. 24

D.5 EXP-Bench

• For MAPPO on VMAS balance and sampling, the design recommends matched interaction budgets, repeated runs, held-out evaluation, and task-specific performance and uncertainty reporting.Only a task description was available, so versions and numerical settings remained contingent; the interaction concluded without claiming an executed experiment or outcome.Source: §D.5 · p. 24

D.6 SciConvBench

• SciConvBench’s actor prompt asks for one focused clarification at a time, then a final specification marked [COMPLETE] once ready.Source: §D.6 · p. 24
• For SciConvBench, the simulator answers clarification questions using intended requirements and configured knowledge policy, while persona affects communication but not requirements.Unspecified details must not be presented as source facts.Source: §D.6 · p. 24
• In the self-weight example, one clarification establishes a uniform bar fixed at the top and free at the bottom; the actor then submits the stress, displacement, and elongation specification.The submitted elongation is δ=ρgL^2/(2E), using consistent units.Source: §D.6 · p. 25
• The SciConvBench actor terminated after one clarification, but the example does not establish that all intended requirements were recovered or that a solver was executed.The recorded dialogue is the actor’s submitted specification, not a solver result.Source: §D.6 · p. 25

D.7 ASI-Bench

• ASI-Bench actors inspect permitted inputs, propose an approach and await approval, then execute approved work and hand off outputs, diagnostics, limitations, and artifact paths.Simulators review the proposed method and reported outputs against the original contract, treating guidance as advice and actor reports as unverified.Source: §D.7 · p. 25
• The deblurring example reports λ=10^-4, 1,335 iterations, and residual 2.16684884 against an approximate target of 2.358912, with five artifacts passing reported format checks.No ground-truth error is claimed.Source: §D.7 · p. 25

E Full Interactions for the Discussion Cases

• The full-interaction section pairs persona 127 conversations with matched persona-off conversations for the same tasks; both discussion cases use GPT-6 Astra at low reasoning effort and score 100/100.C4 documents workflow revision and C5 incremental clarification; persona 127’s profile is in Appendix D.2.Source: E Full Interactions for the Discussion… · p. 25
• The export retains original wording and source notation, marks editorial highlights and omissions, and excludes private tool logs, system prompts, and evaluator material; the cited transcript references limma, Benjamini–Hochberg, and Holm.Source paths, event-log hashes, and byte-identical conversation snapshots are retained in the review/discussion_candidates archive.Source: E Full Interactions for the Discussion… · pp. 25–26

• Gene dependence and unknown normalization limit inference, and tissue effects should be read as balanced averages across response groups rather than necessarily uniform effects.Response-by-tissue inspection found 3,075 nominal and 298 BH-significant genes; the analysis lacked empirical-Bayes moderation and does not establish equivalence.Source: §E.1 · pp. 28, 31
• Inspection found 90 expression columns matched to metadata by Code and 30 complete inferred subject triplets, but raw counts and documented subject-ID definitions were unavailable.The workbook contained 25,402 unique gene identifiers and 10 samples per tissue-response cell; the pairing was inferred from consistent triplets.Source: §E.1 · pp. 28–29
• The authorized analysis used unmoderated paired OLS on log2(NormCount + 1), with inferred subject and tissue effects, a common expression filter, and BH correction within contrasts.The 90×32 design had rank 32 and residual df=58; the filter retained 17,049 of 25,402 genes, all testable. This was explicitly not limma-trend.Source: §E.1 · pp. 30–31
• Final vs Baseline Blood had the smallest effects, but all three contrasts rejected the Holm-adjusted binomial no-enrichment null at 5%; relative weakness is not equivalence.The section reports no evidence that any comparison meets a no-enrichment criterion; “negligible” describes relative weakness only.Source: §E.1 · p. 31
• K/N was 10,689/17,049 (62.70%) for Final vs Baseline Blood, 16,059/17,049 (94.19%) for Dentate gyrus vs Baseline blood, and 16,066/17,049 (94.23%) for Dentate gyrus vs Final blood.BH discoveries were 10,105, 16,048, and 16,049, respectively; median absolute log2-expression differences were 0.248, 1.636, and 1.691, respectively.Source: §E.1 · p. 31
• Holm-adjusted log10 exact-binomial upper-tail p-values were -9159.634, -19275.953, and -19293.209 for the three contrasts, respectively.The working test used testable genes as trials and unadjusted two-sided t-test p<0.05 as success, under an independent, calibrated no-DE null; BH discoveries were separate.Source: §E.1 · p. 31

E.1 Workflow Revision: BixBench with Persona 127

• Four paired sensitivity runs and an unpaired Response-adjusted model did not change the conclusion or the rejection of the binomial nulls.Paired-run blood median absolute differences remained 0.231–0.262 versus 1.426–1.731 for dentate-gyrus contrasts; the unpaired model retained 10,151/16,058/16,068 BH discoveries and unchanged effect estimates.Source: §E.1 · pp. 31–32
• Interpretation is limited by the lack of a biological equivalence margin, unmoderated gene-wise OLS, correlated genes, and unknown cross-tissue normalization and covariates.Response-by-tissue heterogeneity was present for a subset of genes.Source: §E.1 · p. 32
• 0.248 versus 1.636 and 1.691 median absolute expression differences showed the blood comparison was weaker; all three exact-binomial enrichment tests rejected after Holm correction.The conclusion was that none of the comparisons qualified as negligible under the specified no-enrichment test at 5%, which is not evidence of biological equivalence.Source: §E.1 · p. 32
• The agent executed analysis/tissue_comparisons.py and reported the approved unmoderated Python fallback, assumptions, limitations, and execution artifacts.The report also recorded a full-rank paired design with 58 residual degrees of freedom and 17,049 testable genes.Source: §E.1 · p. 32

E.2 Matched BixBench Conversation without a Persona

• The workbook contained 25,402 genes and 90 matched expression samples, but pairing, raw counts, normalization provenance, and numbered DESeq contrast definitions were undocumented.The proposed contrasts used the normalized expression columns rather than the workbook’s unassigned DESeq blocks.Source: §E.2 · pp. 33–35
• The approved analysis modeled log2(normalized count + 1) with categorical Response adjustment and HC3 tests, then tested nominal p<0.05 gene-level successes against p0=0.05 with Holm correction.The binomial tail is only a heuristic global summary because gene-level trials may be dependent; normalized fractional values, not raw counts, were analyzed.Source: §E.2 · pp. 34–35
• Final–Baseline Blood had the smallest median absolute effect (0.217 versus 1.179 for Dentate–Baseline and 1.208 for Dentate–Final), but all three contrasts showed binomial enrichment, not biological equivalence.The blood comparison is the best-supported relative answer; all three enrichment tests rejected their nulls, and no formal equivalence margin was specified.Source: §E.2 · pp. 36–37
• In the tissue-by-response sensitivity analysis, balanced sampling made average effects identical to additive-model estimates, while standard errors differed and response-specific blood effects varied.Final–Baseline median absolute effects were 0.202 for Control, 0.302 for good_responder, and 0.298 for bad_responder; the analysis found no evidence that averaging concealed an absent blood effect.Source: §E.2 · p. 37

E.3 Incremental Clarification: SciConvBench with Persona 127

• Both SciConvBench conversations received full credit despite taking different clarification paths.Figure 7 presents incremental clarification with explanatory replies.Source: §E.3 · p. 38, Figure 7 · p. 38
• Clarifications set the persona conversation’s full shaft length to 1500 mm, allowable twist to 0.75°, and shear modulus to 25,939.8 N/mm², with torsional shear only.The researcher specifies no additional transverse shear force and directs sizing solely from the twist limit.Source: §E.3 · p. 39
• The persona conversation’s final specification assumes linear-elastic, homogeneous isotropic material and constant cross-section and torque, and gives d_min = [32TL/(πGφ_allow)]^(1/4) for twist-based sizing.It specifies 6061-T6 aluminum, T = 1,000,000 N·mm, L = 1500 mm, and φ_allow = 0.75° = π/240 radians.Source: §E.3 · p. 40
• The persona specification requests the minimum diameter without rounding down, twist-limit verification, and maximum surface torsional shear stress, but no strength-adequacy claim.No allowable stress is specified, and sizing is solely from the twist limit.Source: §E.3 · p. 40

E.4 Matched SciConvBench Conversation without a Persona

• In the persona-off conversation, clarification requests a numerical N–mm–MPa calculation for 6061-T6 aluminum with a 0.75° full-length twist limit using φ = TL/(JG) and J = πd^4/32.The conversation also specifies L = 1500 mm, G = 25939.8 MPa, and T = 1,000,000 N·mm.Source: §E.4 · pp. 40–41
• The persona-off final specification assumes a homogeneous, isotropic shaft with constant circular cross-section and torque, requests the minimum diameter without rounding down, and requires twist-limit verification.It sets the relative twist limit over the full shaft length to 0.75°.Source: §E.4 · p. 41
• The persona-off conversation treats sizing as torsional-stiffness-only and excludes strength, bending, transverse shear, stress concentrations, and connection effects.It states that no separate transverse shear load is specified.Source: §E.4 · p. 41