Agent-ValueBench: A Comprehensive Benchmark for Evaluating Agent Values

Evaluation challenges

• Agent-value evaluation is non-trivial because agent tasks require executable environments, trajectory-level behavioral adjudication, and controlled comparison across harnesses.These dataset-, evaluation-, and system-level requirements extend beyond standard text-only LLM value evaluation.Source: §1 · p. 2

Problem identification

• Figure 1 shows that models sharing the same underlying LLM can exhibit strikingly divergent value tendencies in LLM versus agentic modalities.The paper also formalizes this agent–LLM value divergence as Theorem 1.1.Source: §1 · pp. 1–2, Figure 1 · p. 2

Benchmark contribution

• Agent-ValueBench evaluates agent values across 394 executable environments, 16 domains, 4,335 tasks, 28 value systems, and 332 value dimensions.It is presented as the first benchmark dedicated to evaluating autonomous-agent values.Source: §1 · pp. 2–3

Benchmark design

• The benchmark addresses these challenges through psychologist-curated synthesis, two pole-aligned golden trajectories with checkpoints, and an audited LLM-as-Judge meta-rubric.Its pipeline jointly synthesizes environments and value-conflict tasks, while experts verify, repair, or rewrite each artifact.Source: §1 · p. 2

Experimental evaluation

• Across 14 frontier models and 4 mainstream harnesses, the study identifies a shared Value Tide, non-additive harness effects, and stronger deliberate steering through embedded skills.The authors consequently position harness alignment and skill steering as increasingly important agent-value alignment levers.Source: §1 · pp. 2–3

2. Dataset Construction

• The dataset pipeline jointly synthesizes executable environments, value-conflict tasks, and trajectory-level rubrics, with per-instance expert-in-the-loop refinement.Each stage is capped by expert review to improve the benchmark’s validity and usability.Source: §2 · p. 3

2.1. Environment Construction

• Psychologist curation reduces 450 candidate environments to 394 executable environments spanning 16 domains, auditing fidelity, authenticity, and self-consistency.The automated process includes source discovery, program synthesis, behavioral validation, and iterative diagnosis–repair before final curation.Source: §2.1 · pp. 3–4

2.2. Task Construction

• The task pipeline combines 28 value systems and 332 value dimensions with paired golden trajectories whose checkpoints make opposing value alignments behaviorally observable, yielding 4,335 tasks.Tasks are synthesized from sampled environments and non-overlapping value pairs, then psychologist-vetted for authenticity, conflict design, consistency, discriminability, and executability.Source: §2.2 · pp. 4–5

2.3. Rubric-based Evaluation

• Trajectory evaluation uses a symmetric, behaviorally anchored rubric that analyzes Attention, Construal, and Execution through task-specific three-level scoring items applied by an LLM-as-Judge.A hand-designed meta-rubric drives task-specific rubric synthesis, with each rubric item tied to a checkpoint, weight, scoring question, and observable anchors.Source: §2.3 · p. 5

3.1. Experimental Setup

• The study benchmarks 14 proprietary and open-weights frontier models from 10 families across four mainstream harnesses, measuring Value Adherence and Value Priority.Value Adherence captures conformity to a value dimension across trajectories, while Value Priority captures preference between conflicting values.Source: §3.1 · p. 6

3.2. RQ1: Agent Values Exhibit a Value Tide

• Across 14 models, adherence remains tightly clustered while priority converges by system, forming a shared Value Tide with selective crests, shallows, and interpretable counter-currents.Average adherence spans [5.94, 6.27] on MFT08, [5.16, 6.17] on PVQ40, [5.39, 6.64] on HEXACO, and [5.06, 5.92] on LVI; Qwen3 30B A3B is lowest on three systems and elevates Hedonism to first in PVQ40.Source: §3.2 · pp. 6–8, Figure 3 · p. 7

3.3. RQ2: The Tide Bends Under Harness Pull

• Harness substitution changes adherence and priority substantially: 74% of 93 within-model ranges exceed 1.0, while the same harness can drive opposite value drifts across models.GPT-5.4’s NFCC2000 Discomfort with Ambiguity rises from 4.08 to 8.48 under OpenClaw. Codex’s Closed-Mindedness decreases on GPT-5.4 but increases on Kimi K2.5.Source: §3.3 · pp. 9–10

3.4. RQ3: The Tide Bends to Deliberate Steering

• Skill steering places 3.7 of 5 values at exact target ranks on average, versus 1.3 under prompt steering, and once recovers GPT-5.4’s full target ordering.Embedded skills also produce larger adherence shifts, including Sonnet Purity Δ=+5.67 and GPT-5.4 Fairness Δ=−1.89, whereas prompts yield Δ=0 and Δ=−0.27.Source: §3.4 · pp. 10–11

4. Conclusion

• Agent-ValueBench is the first benchmark dedicated to agent values, revealing a Value Tide that bends under harness pull and deliberate steering.These findings motivate shifting alignment attention from model alignment and prompt steering toward harness alignment and skill steering.Source: §4 · p. 11
• The benchmark’s causal mechanisms remain unexplored, so future work should investigate the mechanistic underpinnings of the Value Tide.Source: §4 · p. 11

(Appendix)

• The appendix contains ethics, related work, notation, implementation details, a value-system inventory, comprehensive experimental results, prompt templates, proofs, and validation, executability, and stability studies.Source: (Appendix) · p. 25

A. Ethics Statement

• Agent-ValueBench is a descriptive audit rather than a normative authority, universal moral taxonomy, or deployment certification, with results scoped to comparative profiles under its benchmark distribution.It uses synthesized executable environments, psychologist-curated tasks, golden trajectories, and task-specific rubrics without private logs, surveillance, biometric, or personally identifiable behavioral data.Source: A. Ethics Statement · pp. 25–26
• Because value evaluations could be misused to game evaluations or steer agents toward harmful orientations, the authors recommend human supervision, safety review, sandboxing, and restrictions on harmful applications.They release artifacts for reproducibility and auditing while emphasizing additional domain-specific, legal, and human-subject review for deployment.Source: §4 · p. 11, A. Ethics Statement · p. 26

B. Related Work

• Related work treats values as structured motivational priorities revealed through behavioral trade-offs, while prior evaluations largely target LLMs, leaving autonomous-agent values underexplored.Agent-ValueBench operationalizes value theory through pairwise conflicts in executable environments and addresses this gap as the first benchmark explicitly dedicated to agent-value evaluation.Source: B. Related Work · pp. 26–27

C. Notation

• Table 2 consolidates notation for environment auditing, task and value-system indices, conflicting value pairs, executable task components, and rubric synthesis.Key symbols include NegPAcc, ℓ, 𝑒ℓ, 𝑠ℓ, P𝑠, {𝑣𝑎, 𝑣𝑏}, 𝑐ℓ, M, and Φrubric.Source: Table 2 · p. 27, C. Notation · p. 27

Hyperparameter Settings

• Experiments use fixed model assignments, deterministic decoding, bounded trajectories, OpenRouter access, and a four-GPU server.Evaluation uses temperature 0.0, a 50-step and 12,000-token cap, and synthesis acceptance thresholds of 1.0 for NFAcc, PPAcc, and NegPAcc.Source: D. Implementation Details · p. 28

Value Priority Computation Details

• Value priorities aggregate trajectory-level adherence comparisons into pairwise wins, awarding ties as 0.5 wins to both dimensions.The resulting win counts are computed for every observed value pair within each agent–value-system configuration.Source: D. Implementation Details · p. 28
• Bradley–Terry models independently estimate each value’s positive strength within every agent–value-system configuration.A weak symmetric smoothing term α=0.01 is applied only to existing comparison edges to prevent separation without adding unobserved pairs.Source: D. Implementation Details · p. 28
• Final value-priority rankings sort normalized log-strengths in descending order after Minorize–Maximization estimation.Scale normalization zero-centers log-strengths within each value system after every iteration to ensure identifiability.Source: D. Implementation Details · pp. 28–29

E. Value-System Inventory

• The inventory covers 28 value systems and 332 value dimensions.It includes the complete benchmark inventory.Source: E. Value-System Inventory · p. 29
• Its frameworks include broad values such as Tradition, Benevolence, Universalism, Self-Direction, Stimulation, and Hedonism.These capture commitments to tradition, others’ welfare, autonomy, novelty, and pleasure.Source: E. Value-System Inventory · p. 30
• The inventory also distinguishes interpersonal and social values, including Dependability, Caring, Tolerance, Concern, conformity, security, power, and social face.These dimensions address in-group reliability and welfare, acceptance of difference, equality, rule compliance, safety, dominance, resources, and public image.Source: E. Value-System Inventory · p. 30
• It includes cognitive and worldview dimensions contrasting analytic with holistic thinking, field with parts attention, interactionist with dispositionist causality, and cyclic with linear change.The inventory further covers contradiction resolution, perspective taking, empathy, sincerity, fairness, and related social orientations.Source: E. Value-System Inventory · pp. 30–31
• Personality-oriented systems cover emotionality, extraversion, agreeableness, conscientiousness, openness to experience, and altruism through more specific facets.These facets include fearfulness, anxiety, dependence, sociability, patience, diligence, prudence, creativity, and sympathy toward others.Source: E. Value-System Inventory · pp. 31–33
• Additional worldview constructs address social cynicism, reward for application, social complexity, fate determinism, and fate alterability.These capture beliefs about social outcomes, effort, behavioral variability, multiple solutions, and whether personal action can change fate.Source: E. Value-System Inventory · p. 33
• The value-system inventory covers beliefs about fate, religiosity, social exchange, and moral foundations including care, fairness, loyalty, authority, and purity.It also distinguishes intuitions about harm avoidance, equality, proportionality, ingroup cooperation, legitimate authority, and contamination.Source: E. Value-System Inventory · p. 33
• The inventory includes emotion-regulation strategies, broad personality traits, life satisfaction, positive coping, perceived control, and narrower interpersonal, achievement, organizational, openness, and expressivity dimensions.Cognitive reappraisal changes emotional impact through reinterpretation, whereas expressive suppression inhibits ongoing emotion-expressive behavior.Source: E. Value-System Inventory · p. 34

BARCHARD

• BARCHARD adds emotion-based decision-making, empathic and socially responsive affect, persistent goal pursuit, reward seeking, approach and inhibition motives, and private and public self-consciousness.These dimensions distinguish moving toward desired outcomes from moving away from unpleasant ones, alongside attention to hidden versus publicly displayed aspects of self.Source: BARCHARD · p. 35
• BARCHARD further measures affective instability and maladaptive tendencies including anger, anhedonia, anxiety, callousness, depression, dominance, detachment, grandiosity, aggression, irresponsibility, manipulation, and mistrust.The CAT-PD portion includes affective lability and anger among these clinically oriented dimensions.Source: BARCHARD · p. 36
• Maladaptive tendencies include non-perseverance, non-planfulness, norm violation, peculiarity, perfectionism, relationship insecurity, rigidity, risk taking, rudeness, self-harm, withdrawal, submissiveness, unusual beliefs, unusual experiences, and workaholism.These scales capture difficulties with persistence and planning, rule rejection, eccentricity, inflexibility, interpersonal insecurity, dangerous or self-injurious behavior, social avoidance, and excessive work orientation.Source: BARCHARD · p. 37
• Personality and interpersonal dimensions cover analytical and complex thinking, breadth of interest, innovation, tolerance, sociability, confidence, emotionality, empathy, anxiety, cooperativeness, opportunism, social astuteness, dependability, organization, responsibility, and assertiveness.The inventory also includes energy level, traditional values, alienation, control, neuroticism, wellbeing, harm avoidance, and social closeness.Source: BARCHARD · pp. 37–38
• Foundational life values include achievement, accomplishment, self-fulfillment, self-respect, belonging, warm relationships, concern for others, environmental concern, health, humility, independence, privacy, responsibility, scientific understanding, spirituality, and truth-seeking.These dimensions frame desired outcomes involving competence, connection, wellbeing, autonomy, moral concern, knowledge, and transcendence.Source: BARCHARD · pp. 38–39
• The inventory represents Schwartz-like value orientations spanning social love, political power, aesthetic harmony, religious unity, traditionalism, achievement, hedonism, stimulation, status, recognition, influence, wealth, and authority.These dimensions respectively emphasize relationships, leadership, beauty, moral excellence, continuity, competence, pleasure, novelty, prestige, approval, impact, material possessions, and command.Source: BARCHARD · p. 39
• Additional facets specify success and competence, including successful, capable, ambitious, influential, and intelligent, alongside pleasure, self-indulgence, enjoyment of life, daring, and adventure.Together these facets operationalize accomplishment, effectiveness, aspiration, impact, logical thinking, gratification, leisure, excitement, novelty, challenge, and risk.Source: BARCHARD · pp. 39–40
• BARCHARD includes novelty-seeking and self-direction through a varied life, exciting experiences, curiosity, creativity, and freedom.These dimensions emphasize challenge, stimulation, exploration, imagination, and freedom of action and thought.Source: BARCHARD · p. 40
• The inventory also values choosing one’s own goals, independence, self-respect, universalism, environmental protection, and unity with nature.Together, these dimensions cover personal autonomy and worth alongside concern for people and the natural world.Source: BARCHARD · p. 40
• BARCHARD further encompasses beauty, broad-mindedness, social justice, wisdom, equality, and a world at peace.These values emphasize appreciation, tolerance, mature understanding, equal opportunity, and freedom from war and conflict.Source: BARCHARD · p. 40
• Relational and emotional values include inner harmony, benevolence, helpfulness, forgiveness, loyalty, and responsibility.They concern peace with oneself, welfare of others, pardon, fidelity to close groups, and dependable conduct.Source: BARCHARD · p. 40
• The inventory includes spirituality, true friendship, mature love, meaning in life, tradition, and acceptance of one’s portion in life.These dimensions combine spiritual purpose and intimate relationships with respect for cultural or religious customs and life circumstances.Source: BARCHARD · p. 40
• Finally, BARCHARD covers devout faith, humility, respect for tradition, moderation, and conformity to social expectations.These values emphasize religious belief, modesty, preservation of customs, avoidance of extremes, and restraint against harming or upsetting others.Source: BARCHARD · p. 40

• The experiments evaluate value adherence and value priority across value systems, models, harnesses, and steering settings.Source: F. Comprehensive Experimental Results · p. 41

Experimental coverage

• The experiments cover value adherence and value priority across all 28 value systems in Agent-ValueBench.The complete results are presented in Tables 4–22.Source: §F.1 · p. 41

Value adherence

• For value adherence, tables identify the top-performing model and runner-up for each dimension using bold and underline markers.This reporting convention is specified for the value-system result tables.Source: Table 4 · p. 42, Table 5 · p. 42, Table 6 · p. 43

Value priority

• For value priority, the same markers denote first- and second-ranked results within each model.The convention is applied throughout the reported value-system tables.Source: Table 4 · p. 42, Table 5 · p. 42, Table 6 · p. 43

Experimental scope

• The experiments evaluate Claude Sonnet 4.6, GPT-5.4, and Kimi K2.5 within Vanilla ReAct, Claude Code, and Codex across four value systems.The value systems are MFT08, NFCC2000, PVQ40, and VSM13.Source: §F.2 · p. 59

Results

• Tables 23–30 report value adherence and value priority results across the model–harness configurations.The reported analyses cover all four value systems and the three evaluated agent harnesses.Source: §F.2 · p. 59

Value adherence

• For value adherence in mft08, bold and underlined entries denote the highest-scoring model–harness configuration and the runner-up per dimension.Table 23 reports value adherence results across agent harnesses for mft08.Source: Table 23 · p. 59, §F.2 · p. 59

Value priority

• For value priority in mft08, bold and underlined entries denote the first- and second-ranked dimensions, respectively, within each model–harness configuration.Source: Table 24 · p. 59, §F.2 · p. 59

PVQ40

• PVQ40 steering results likewise pair prompt- or skill-guided conditions with unguided harness baselines in Tables 33 and 34.The reported target orderings include GPT-5.4: Stimulation≻Hedonism≻Tradition≻Power≻Benevolence≻Self-Direction≻Conformity≻Achievement≻Security≻Universalism; Kimi K2.5: Benevolence≻Power≻Tradition≻Conformity≻Stimulation≻Achievement≻Hedonism≻Self-Direction≻Universalism≻Security.Source: §F.3 · pp. 59, 62

MFT08

• MFT08 steering results pair every prompt- or skill-guided condition with its unguided harness baseline in Tables 31 and 32, using model-specific target orderings.The targets are Sonnet 4.6: Purity≻Authority≻Fairness≻Care≻Loyalty; GPT-5.4: Care≻Authority≻Purity≻Loyalty≻Fairness; Kimi K2.5: Authority≻Fairness≻Purity≻Care≻Loyalty.Source: §F.3 · p. 59, Table 31 · p. 62

G. Prompt Templates

• The prompt-template section supplies materials for case construction, rubric synthesis, and trajectory-level evaluation.Source: G. Prompt Templates · p. 62

F.4. Agent–LLM Priority Comparison

• Identical backbone models are evaluated as standalone LLMs and agents using multiple-choice adaptations of Agent-ValueBench to compare their value-priority profiles.The comparison covers Claude Sonnet 4.6, DeepSeek V3.2, Gemini 3.1 Pro Preview, and GPT-5.4; Table 35 reports rankings on mft08 and nfcc2000.Source: §F.4 · p. 62, Table 35 · p. 64

G.1. Task Construction Prompts

• Task-construction prompts require a realistic, concrete operational task whose conflict is implicit in structural constraints, uses authoritative value definitions, and maintains causal coherence among the task, state keys, and checkpoint plans.Source: §G.1 · pp. 65–66
• Task-construction prompts require executable, causally coherent cases with non-leaking descriptions, non-mirrored value-consistent checkpoint lists, and at least two checkpoints per track.Source: §G.1 · p. 69

G.2. Rubric Synthesis and Online Judging Prompts

• Rubric prompts map each checkpoint to observable, non-redundant items scored independently across full trajectories, prioritizing state changes and tool evidence over reasoning or keywords.The rubric uses attention, interpretation, and enactment lenses, preserves checkpoint order, assigns integer weights, and returns UNSCORABLE when a checkpoint cannot yield an observable item.Source: §G.2 · pp. 70, 72

G.3. Value-Priority Skill Example

• The example MFT08 skill imposes the priority order Care, Authority, Purity, Loyalty, Fairness, applying higher-priority values to resolve conflicts during planning and tool use.Care prioritizes harm reduction, Authority governs legitimate procedures when Care is not at stake, and lower values cannot override higher-priority ones.Source: §G.3 · pp. 73–74
• The skill directs agents to choose the available action that satisfies the highest-priority value involved in the conflict.The priority comparison covers Authority, Purity, Loyalty, and the resulting action choice.Source: §G.3 · p. 75
• Care overrides Authority when following a legitimate rule or supervisor instruction would create meaningful hardship, safety risk, privacy harm, or other serious harm.The preferred action reduces meaningful harm while using the task’s provided tools and evidence.Source: §G.3 · p. 75
• Authority overrides Purity, Purity overrides Loyalty, and Loyalty overrides Fairness when the higher-priority value is implicated and no still-higher value blocks the choice.The examples require following valid workflows, preserving clean and respectful handling, and honoring justified commitments unless Care, Authority, or Purity takes precedence.Source: §G.3 · pp. 75–76
• Fairness guides action only when alternatives are equally safe, authorized, clean, and loyal, favoring the option that applies criteria consistently across similarly situated parties.Fairness remains the final value and should not displace a different outcome required by Care, Authority, Purity, or Loyalty.Source: §G.3 · p. 76

H. Proof of Agent–LLM Value Non-Identity

• Agent and LLM value profiles coincide for a fixed model-side law exactly when text-only and agentic channels induce the same value-evidence distribution.If the induced evidence laws differ, a bounded evidence-scoring rule separates the profiles; uniform coincidence over all laws requires pointwise equality of the two evidence kernels.Source: H. Proof of Agent–LLM Value Non-Identity · pp. 77–78
• Agent values therefore are not reducible to raw LLM text values by default, becoming identical only when agentic realization preserves the relevant value-evidence law.A minimal binary-evidence example shows that sharing the same model-side source does not guarantee strict identity.Source: H. Proof of Agent–LLM Value Non-Identity · pp. 77–78

I. Human Validation, Executability, and Stability Studies

• Human validation used psychologists with calibration and blinding procedures, distinguishing construction-stage full-pool curation from confirmatory retained-pool validation.For environment and task studies, ICC(2,3) estimated reliability of the three-expert panel mean.Source: §I · p. 78
• The validation package combined seven checks spanning executability, artifact design, golden trajectories, rubrics, judge agreement, rollout stochasticity, and judge replay.Population-level evidence used stratified retained-pool validation, with curation and estimates summarized in Tables 36 and 37.Source: §I.1 · p. 78, §I · p. 78

I.1. Validation Overview and Population-Level Evidence

• More than 94% of retained environments and tasks met the pre-specified mean-score ≥4 expert-quality criterion, while golden trajectories recovered their intended A/B value poles under blinded assignment.The rubric and judge studies further supported scoring-layer content validity and alignment with human application.Source: §I.1 · p. 79, Table 37 · p. 79

I.2. Automated Executability Audit

• The automated audit checked imports, tool schemas, state dependencies, trajectory replay, terminal validity, flakiness, lightweight-agent completion, and formal rollout quality before freeze.Construction-time failures were repaired and regression-tested, with final outcomes reported in Table 38.Source: §I.2 · p. 79, Table 38 · p. 79
• 4 / 60,480 rollouts = 0.007% execution-attributable invalidity after deterministic retry, with final unscored rollouts below 0.1%.The smoke test reached terminal states on 4,333 of 4,335 tasks and satisfied finite-state completion on 4,331; one benchmark-attributable edge case was repaired before freeze.Source: §I.2 · p. 79, Table 38 · p. 79

I.3. Environment and Task Design Reliability

• Environment validation used structured expert review of source summaries and synthesized specifications, rating fidelity, authenticity or realism, and tool/resource self-consistency.Task validation similarly reviewed task descriptions, environments, functions, value definitions, and A/B checkpoint lists.Source: §I.3 · p. 80, Figure 6 · p. 80, Figure 7 · p. 81
• Environment ratings were high for fidelity, realism, and tool/resource consistency, while task ratings supported realism, implicit conflict design, value consistency, and behavioral discriminability.Task executability was assessed at the specification level; code-level execution was established separately by the automated audit in Sec. I.2.Source: §I.3 · p. 82, Table 39 · p. 82, Table 40 · p. 82

I.4. Golden-Trajectory Value Anchors

• The blind semantic protocol tested whether golden trajectories expressed intended values without cross-value leakage, while rollout coverage tested whether anchors remained operationally active in non-failed rollouts.Every retained task and trajectory was reviewed during construction, with confirmatory semantic validity estimated on a stratified retained-pool sample.Source: §I.4 · p. 82

I.4.1. Value-Item Instantiation and Blind Discriminability

• Six psychologists evaluated 120 retained tasks and both trajectories per task under blinded conditions, hiding intended labels, rubric items, judge scores, model identity, and generation metadata.A prior three-rater pilot supported protocol debugging only and was excluded from headline estimates.Source: §I.4.1 · p. 82
• Across 1,440 expert trajectory ratings, 97.1% cited behavioral evidence spans, while primary-basis confounding was 2.9% for both value-item instantiation and blind assignment.These results support treating golden trajectories as behaviorally recoverable value anchors rather than stylistic or success-only artifacts.Source: §I.4.1 · p. 82, Table 41 · p. 82, Table 42 · p. 82
• Acceptance required high intended-value ratings, low opposing-value ratings, large discriminant margins, reliable panel estimates, blind A/B recoverability, behavioral evidence, and low non-value-confound rates.Tables 41 and 42 report results satisfying these criteria.Source: §I.4.1 · p. 82, Table 41 · p. 82, Table 42 · p. 82

I.4.2. Anchor Coverage in Real Rollouts

• 89.3% ± 8.4% mean anchor-opportunity coverage and 92.7% human–anchor agreement show that intended anchors are present and behaviorally aligned in real rollouts.The audit covers 1,200 task-completing rollouts spanning all 28 value systems, 16 domains, 14 models, and 4 harnesses; anchor margin correlates with rubric priority margin at Spearman ρ = 0.82.Source: §I.4.2 · p. 83, Table 43 · p. 83
• Anchor coverage is robust: 96.2% of rollouts cover at least 75% of opportunities, only 2.8% fall below 70%, and swapped or random-anchor controls collapse the intended alignment.These placebo and ablation results are reported alongside the anchor-coverage summary in Tables 43 and 44.Source: §I.4.2 · p. 83, Table 44 · p. 83

I.5. Rubric Content Validity and Traceability

• Post-revision rubrics satisfy standard and stricter content-validity criteria, chance-adjusted agreement criteria, and complete source-checkpoint coverage after repair.The study evaluates 60 frozen rubrics sampled across value systems, value-pair types, domains, and rubric lengths; Tables 45 and 46 report the results.Source: §I.5 · pp. 83–84, Table 45 · p. 84, Table 46 · p. 84
• Expert semantic audits confirm rubric side fidelity at 99.7%, behavioral observability at 99.4%, item atomicity at 98.4%, non-redundancy at 97.3%, and score-anchor clarity at 98.1%.Human psychologists assess coverage, relevance, clarity, evidence-groundedness, and non-redundancy using the structured interface shown in Figure 8.Source: §I.5 · pp. 83–84, Figure 8 · p. 84

I.6. LLM-as-Judge Validation under Rubrics

• LLM-as-Judge validation uses three blinded psychologists scoring 100 stratified task–trajectory pairs with frozen, task-specific rubrics and observable-evidence rules.Scores use 0–2 criterion anchors, record evidence locations, and classify value-track leaning with a pre-specified tolerance of ε = 0.50.Source: §I.6 · pp. 84–85
• 91.8% exact agreement across 1,080 rubric items and mean absolute error of 0.09 show high human–LLM scoring agreement, with occasional over-crediting of partial satisfaction.The three-way matrix correctly identifies 41 A-leaning, 12 near-tie/mixed, and 39 B-leaning cases out of 100.Source: §I.6 · p. 85, Table 47 · p. 85
• The human–LLM value-track classifications remain robust to the near-tie threshold, with agreement and threshold-sensitivity analyses reported in Tables 47 and 48.The scoring protocol requires applying the rubric exactly, using observable behavior only, avoiding hidden-intent inference, and assigning the lower defensible score under ambiguity.Source: §I.6 · p. 85, Table 48 · p. 85

Computational stability

• Aggregate value profiles, priority rankings, harness effects, and steering effects remain directionally stable across repeated rollouts and judge replays, despite nonzero rollout variability.Rollout subsets were repeated across the main model set, harness comparisons, and steering settings; fixed trajectories were judged five times with fixed rubrics and prompts.Source: §I.7 · p. 85, Table 49 · p. 86, Table 50 · p. 86

Validation argument

• Validation across Sections I.2–I.7 supports executable artifact construction, expert-rated design quality, behaviorally grounded anchors, and human-aligned rubric application.The studies use panel-mean artifact scores, ICC(2,k) reliability, intended/cross-value ratings, anchor coverage, CVI, checkpoint-to-rubric traceability, and human–LLM rubric-score agreement with priority from Δ = A−B under a pre-specified near-tie tolerance.Source: §I.7 · p. 86