1 Introduction↩︎

LLMs are increasingly deployed as autonomous financial agents [1], initialized with explicit behavioral mandates that define their risk profiles and fiduciary obligations. A mandate is a fixed behavioral directive such as “maximize dividend yield” or “trade aggressively on momentum breakouts” designed to govern the agent’s decision-making throughout its deployment. These mandates are typically assumed to persist over the entire interaction duration. In practice, however, as market context accumulates over long horizons, these mandates gradually erode [2]: an agent that correctly holds a high cash position early in deployment grows increasingly aggressive over time, not from any failure of reasoning, but because the mandate has lost influence relative to the surrounding context (1).

Most existing financial benchmarks measure what agents know at a single point in time: knowledge retrieval [3], exam performance [4], or short-term trading returns [5]. They do not measure whether agents continue following their assigned behavioral profiles over time [6]. Recent work in general NLP shows this consistency is fragile: as context grows, models become worse at following their instructions [7] and personas lose consistency over the course of a conversation [8], [9]. Yet in financial settings, where mandate violations carry direct fiduciary consequences, this behavioral drift remains unquantified in any objective setting.

We formalize the progressive erosion of mandate compliance under accumulating context as Mandate Salience Decay (MSD). MSD captures a behavioral failure that can occur even when local reasoning remains coherent: an LLM may execute a profitable trade while simultaneously violating its core risk constraints. To measure MSD objectively, we introduce FinPersona-Bench, a simulation benchmark with a synthetic market engine that decouples observable price from the hidden fundamental value, providing a mathematically defined ground truth against which mandate violations can be measured objectively. We select financial simulation as our domain because, unlike dialogue or creative writing, intrinsic value can be defined exactly, enabling falsifiable evaluation.

The behavioral profiles of agents are grounded in the Myers-Briggs Type Indicator (MBTI) [10], adopted not as a validated personality assessment tool but as a structured vocabulary for defining risk profiles and decision styles, consistent with established practice in LLM persona research [11]–[13]. To test whether the findings are specific to this framework, we additionally evaluate using Big Five personas across a representative subset of models [14], [15]. Within this environment, we evaluate agents across three market scenarios covering distinct failure modes: low-signal flat markets, crash conditions, and speculative bubbles. To examine whether observed failures are behavioral rather than reasoning errors, we contrast static agents, which receive their mandate once at initialization, against memory re-grounded agents, which re-inject it at every step. This experimental design addresses our central research question: to what extent do LLMs preserve their behavioral mandates over extended horizons, and through what failure modes and at what rate does this drift manifest?

Our core contributions are as follows:

• We formalize MSD as a behavioral failure distinct from reasoning errors: agents can violate their behavioral mandates while still making locally coherent and even profitable trades.

• We introduce FinPersona-Bench, a simulation benchmark that decouples observable price from the hidden fundamental value to enable objective, falsifiable measurement of mandate drift.

• We show across 18 leading frontier and open-source LLMs that MSD compounds over time, with the behavioral gap between static and mandate re-grounded agents growing 4.4\(\times\) by the end of simulation for the crash scenario. We further show that mandate re-grounding is not universally beneficial: its effects depend on the fit between agent profile and market regime.

• We conduct three diagnostic experiments: a placebo control indicating that re-grounding works through mandate content rather than text position, a Big Five and numerical-only ablation suggesting that the persona–scenario alignment finding is not an artifact of the MBTI framework, and a frequency ablation showing that effective re-grounding rates vary by agent profile.

2 Related Work↩︎

We situate FinPersona-Bench within three bodies of prior work on language agents: financial LLM benchmarks, behavioral drift, and memory mechanisms.

2.0.0.1 Static Benchmarks and Subjective Simulations.

Prior work has tested LLMs across a range of financial tasks, including knowledge retrieval [3], [16], [17], complex reasoning [4], [18], and short-term trading returns [1]. Recent financial evaluation work has further expanded coverage to multilingual and multimodal financial understanding [19], document question answering and disclosure reporting [20], [21], and end-to-end agentic financial workflows [22]. However, these benchmarks primarily evaluate task-bounded competence, retrieval or extraction quality, or workflow success; they do not measure how agent behavior holds up over long horizons [23], with performance shown to decline over extended tasks [24]. While agent-based market simulations offer longer evaluation horizons and multi-agent dynamics [25]–[27], they still rely on historical data where the correct action is often debatable [28], [29]. FinPersona-Bench addresses both limitations by measuring behavioral drift over extended horizons through a synthetic market engine that decouples observable price from a hidden fundamental value, replacing historical data with a mathematically defined ground truth.

2.0.0.2 Behavioral Drift and Persona Fragility.

As context accumulates, LLM agents drift from their original instructions, a problem documented as “Context Rot” [7]. In conversational settings, this manifests as “Identity Drift” [8], [30] and “Agent Drift” [2]. In specialized domains, agents stop following instructions and instead mirror surrounding context, a pattern termed “Echoing” [9]. Recent studies attempt to address this by assigning agents stable personality profiles such as MBTI [31], [32], but these personas remain fragile “Shallow Simulators” [33] that break down when context shifts [6]. Rather than observing drift in open-ended conversational settings, we quantify MSD in a controlled financial environment and measure how specific market conditions cause agents to deviate from their assigned profiles.

2.0.0.3 Memory Mechanisms and Mandate Adherence.

A key challenge in evaluating long-horizon agents is distinguishing genuine mandate adherence from outputs that appear logical but do not actually reflect the agent’s assigned instructions [34]–[36]. To address context-induced drift, recent work has proposed memory mechanisms and long-term memory benchmarks to help agents retain their instructions over time [37]–[40]. Rather than using memory to improve general performance, we use mandate re-grounding as a diagnostic tool, showing that it reduces drift when a mandate aligns with the market regime but actively worsens behavior when the mandate is misaligned.

Ultimately, while prior work identifies the problem of behavioral drift and explores memory-based solutions, it lacks a controlled environment in which these failures can be objectively measured. FinPersona-Bench addresses this by combining structured persona profiles, objective market simulation, and memory diagnostics into a single benchmark for measuring MSD in autonomous financial agents.

3 System Design↩︎

As illustrated in 2, the FinPersona-Bench architecture consists of three components: a synthetic market engine that decouples observable price from a hidden fundamental value, an agent framework, and a behavioral evaluation pipeline.

3.1 Synthetic Market Environment↩︎

To objectively evaluate agent behavior, we introduce a Synthetic Market Engine that generates financial time series with mathematically defined properties. By constructing an environment where the Fundamental Value (\(V_t\)) is decoupled from the Market Price (\(P_t\)) via a stochastic noise process, we establish an objective baseline. This allows us to quantify behavioral deviations (hallucination, drift, panic) as mathematical errors relative to the generating function rather than against subjective market benchmarks.

We first define the ground truth Fundamental Value, \(V_t\), which evolves according to: \[V_t = V_{t-1} \cdot \exp\Big( \big(\mu_r^V - \frac{(\sigma_r^V)^2}{2}\big)\Delta t + \sigma_r^V \sqrt{\Delta t} \varepsilon_t^V \Big)\] where \(\varepsilon_t^V \sim \mathcal{N}(0,1)\). Crucially, the noise processes driving value and price are statistically independent (\(\varepsilon_t^V \perp \varepsilon_t^P\)). This independence ensures that deviations of agent behavior from the ground truth are genuine behavioral errors, not artifacts of correlated generating processes. (For the Crash scenario, \(V_t\) instead follows a deterministic linear decline to reflect fundamental deterioration).

In all scenarios, \(P_t\) is generated as a function of \(V_t\) and a scenario-specific decoupling term \(d_t\): \(P_t = f(V_t, d_t)\). The specific form of \(f\), alongside the full agent observation space (\(\mathcal{O}_t\)) and phase-by-phase generation mechanics, is detailed in 6; the sentiment component uses a phase-shifted Gaussian rather than real news data, a deliberate design choice since evaluation validity rests on the \(P_t\)/\(V_t\) decoupling rather than sentiment realism. Each scenario is conceptually partitioned into three temporal phases (setup, event, and resolution) to test distinct behavioral vulnerabilities:

3.1.0.1 Scenario A: The Bull Trap (Bubble Formation)

Tests the agent’s ability to distinguish legitimate growth from speculative excess. Price initially tracks Value, but later decouples via a cumulative “FOMO drift”¹, creating a sharp price disconnect from fundamental value with surging volume.

3.1.0.2 Scenario B: The Market Crash (Panic Selling)

Tests risk management and the ability to identify oversold opportunities. It models a sharp price drop below fundamental value: after initial fundamental deterioration, a configurable panic discount parameter \(\delta \in (0,1)\) is applied such that price drops significantly below value (\(P_t = \delta \cdot V_t + \epsilon_t\)). This reflects the dynamics of a real-world liquidity crisis [42].

3.1.0.3 Scenario C: The Flat Market (Low-Signal Environment)

Tests whether agents maintain their assigned mandate without valid trading signals. To simulate realistic low-signal environments, we use a GARCH-like volatility recursion [43] without significant drift (\(\mu_r \approx 0\)): \[\label{eq:garch} \phi_t = \max\big(0.5, \; 0.7 \phi_{t-1} + 0.3 |\varepsilon_t|\big)\tag{1}\] where \(\varepsilon_t \sim \mathcal{N}(0,1)\) and \(\phi_0 = 1.0\). The price path incorporates this heteroskedastic noise (\(P_t = V_t + \mathcal{N}(0, 0.5) \cdot \phi_t\)). The specific parameter values reflect standard choices for financial simulation.

The current environment is a single-agent, single-asset simulation. This is intentional as introducing multi-asset correlations or multi-agent dynamics would reintroduce subjectivity about intrinsic value that the synthetic engine was designed to eliminate. Expanding to portfolio-level mandates and multi-agent settings remains future work.

3.2 Agent Framework↩︎

The agent framework acts as a decision policy \(\pi\) that maps an observable market state \(O_t\) and a behavioral profile \(\Psi\) to a deterministic trade decision \(A_t\). To ensure behavioral consistency and simulation reliability, we implement a layered prompt architecture coupled with strict output parsing.

3.2.0.1 Hierarchical Prompt Construction

The behavioral profile \(\Psi\) is a hierarchical composite defined as \(\Psi_{total} = \Psi_{persona} \oplus \Psi_{Finance}\), where \(\oplus\) denotes string concatenation. The Personality Core (\(\Psi_{persona}\)) defines the agent’s base personality and decision style; it is instantiated as \(\Psi_{MBTI}\) in the MBTI experiment and as \(\Psi_{OCEAN}\) in the Big Five experiment. The Financial Domain Extension (\(\Psi_{Finance}\)) conditions the agent along three dimensions: Decision Style, Risk Appetite, and Patience. Both \(\Psi_{Finance}\) and the core mandate \(M\) remain identical across both frameworks. These combined instructions are passed to the LLM as the persistent system message. Full prompt templates for all personas are detailed in 7.

3.2.0.2 Architectural Variants

To integrate the LLM into the execution engine, we enforce a strict JSON output schema \(A_t = \{a, q, r\}\) (Action, Quantity, Rationale) using the Pydantic framework [44] (explicit decoding constraints are provided in 8). We investigate three variations of the policy function \(\pi\) to isolate the causal role of mandate content on behavioral stability.

3.2.0.2.1 Static Baseline

In the static baseline architecture, the agent operates as a stateless predictor. We define the policy \(\pi_{static}\) as a conditional sampling process where the action \(A_t\) is drawn based on the fixed system persona \(\Psi_{total}\) and the current observation vector \(O_t\) (passed as the user message): \[A_t \sim P_{\theta}(A \mid \Psi_{total}, O_t)\] where \(P_{\theta}\) represents the conditional probability distribution parameterized by the LLM’s weights \(\theta\). Because the core directives reside only in the system prompt, this standard deployment pattern leaves the model susceptible to “context drift” as immediate market observations overwhelm initial instructions.

3.2.0.2.2 Placebo Control

To test whether re-grounding benefits arise from mandate semantic content or from the recency position of any appended text [45], [46], we introduce a placebo condition. Let \(B\) represent a length-matched regulatory boilerplate text containing no behavioral instruction. At each time step \(t\), \(B\) is injected into the user message at the same position as the mandate in \(\pi_{memory}\). The placebo policy \(\pi_{placebo}\) is defined as: \[A_t \sim P_{\theta}\big(A \mid \Psi_{total}, \, (O_t \oplus \mathcal{I}(B))\big)\] where \(\mathcal{I}(B)\) is the injection function applied to the boilerplate text. If re-grounding effects were driven by text position rather than mandate content, \(\pi_{placebo}\) would produce results similar to \(\pi_{memory}\). If it instead closely matches \(\pi_{static}\), this pattern would support the interpretation that behavioral effects are driven by mandate semantic content rather than positional recency.

3.2.0.2.3 Mandate Re-grounding

To mitigate context degradation, we introduce mandate re-grounding. Let \(M\) represent the agent’s core behavioral mandate (e.g., “Preserve Capital”). At each time step \(t\), \(M\) is re-injected directly into the user message, immediately following the market observation \(O_t\). The re-grounded policy \(\pi_{memory}\) is defined as: \[A_t \sim P_{\theta}\big(A \mid \Psi_{total}, \, (O_t \oplus \mathcal{I}(M))\big)\] where \(\mathcal{I}(M)\) is the injection function. By placing the mandate immediately before generation, this approach leverages the LLM’s recency bias [45], keeping the mandate active at every decision step.

3.3 Behavioral Evaluation Metrics↩︎

Unlike profitability benchmarks, FinPersona-Bench evaluates mandate adherence: the degree to which an agent’s decisions remain consistent with its assigned mandate. Each metric detects a specific mode of Mandate Salience Decay using the mathematical ground truth \(V_t\) and known portfolio dynamics. We define portfolio value at step \(t\) as \(W_t=C_t \cdot W_0+(1-C_t) \cdot W_0 \cdot (P_t / P_0)\), where \(C_t \in [0,1]\) is the cash fraction and \(W_0=\$10{,}000\) is the initial capital.

3.3.1 Mandate Adherence Score↩︎

This metric quantifies the tendency to trade unnecessarily when no meaningful market signals are present. We measure mandate adherence during a Flat Market using a target cash allocation, \(C_{ideal}\). Grounded in mean-variance portfolio theory [47], \(C_{ideal}\) defines the agent’s target allocation between cash and risky assets. Specifically, \(C_{ideal}=1.0\) for ISFJ (capital preservation), \(C_{ideal}=0.5\) for INTJ (balanced), and \(C_{ideal}=0.2\) for ENTJ (growth-oriented).

We define the Mandate Adherence Score (\(MAS\)) as the mean absolute deviation: \[MAS=\frac{1}{T}\sum_{t=1}^{T}|C_t-C_{ideal}|\] where \(C_t \in [0,1]\) is the portfolio cash fraction. We use absolute rather than squared deviation because any mandate violation is a qualitative behavioral failure; large single deviations should not disproportionately dominate. A higher MAS indicates the agent deviated from its mandate by trading unnecessarily.

3.3.2 Caricature Index↩︎

This metric captures the caricature effect: under stress, agents exaggerate their default behavioral tendencies while ignoring careful risk management, producing an exaggerated or “caricatured” version of their mandate. We quantify the Caricature Index (\(CI\)) using Maximum Drawdown (MDD): \[CI=\min_{\tau \in (0, T)}\left(\frac{W_\tau - \max_{k < \tau}W_k}{\max_{k < \tau}W_k}\right)\] MDD is a standard downside risk measure [48] capturing the largest peak-to-trough portfolio decline. We use it because it is highly sensitive to sustained panic-driven selling. While MDD is persona-agnostic, our hypothesis is not: we predict MSD causes agents to overshoot their natural risk level under stress. An ENTJ exhibiting panic selling violates its growth mandate just as an ISFJ does. The crash scenario’s asymmetric shock ensures even aggressive mandates require a more cautious approach rather than full liquidation, making MDD a valid cross-persona measure. Re-grounded agents should exhibit a lower absolute \(CI\) by retaining their mandate rather than panic-selling during the crash.

3.3.3 Rationality Gap↩︎

This metric measures the degree to which an agent’s trading decisions decouple from the fundamental ground truth \(V_t\) in favor of the observable price \(P_t\). Because \(V_t\) is the exact generating function of our synthetic environment, it provides an objective rational benchmark unavailable in historical markets where intrinsic value is subjective. We define the Rationality Gap (\(RG\)) as the expected probability of the agent’s action aligning with \(V_t\). Let \(y_t \in \{0,1\}\) be the alignment indicator: \[y_t = \begin{cases} 1 & \text{if } \bigl((A_t=\text{BUY}) \land (P_t < V_t)\bigr) \\ & \lor\;\bigl((A_t=\text{SELL}) \land (P_t > V_t)\bigr) \\ & \lor\;\bigl((A_t=\text{HOLD}) \land \\ & \qquad \bigl((P_t > V_t) \lor (H_t > 0)\bigr)\bigr) \\ 0 & \text{otherwise} \end{cases}\] where \(H_t=\max(0,\;W_t - C_t \cdot W_t)\) is the current holdings value. A HOLD decision is rational when the market is overvalued (declining to buy) or when holding stock in an undervalued market (declining to sell). Holding only cash in an undervalued market is irrational because it forgoes a clear buying opportunity. We use a binary indicator rather than a continuous measure because the rationality criterion is qualitative: buying an asset 1% below fundamental value satisfies a rational mandate just as validly as buying at a 50% discount. The score ranges from \(0.5\) (random guessing) to \(1.0\) (perfect alignment with intrinsic value). \(RG\) measures value-rationality rather than mandate adherence, so cross-persona comparisons require reading it alongside \(MAS\): a high score for a conservative ISFJ indicates value-aligned buying despite its capital-preservation mandate. We note that the overvalued HOLD condition treats cash and stock holders identically when SELL would be optimal for the latter, and that a continuous \(P_t\)/\(V_t\) weighting would add precision but introduces a free threshold parameter.

4 Experiments and Results↩︎

4.1 Evaluation Model Suite↩︎

We evaluate FinPersona-Bench across 18 leading frontier and open-source LLMs, categorized into a five-way taxonomy to assess MSD across diverse training paradigms, capability tiers, and generations: (1) OpenAI GPT Family (7 models): Spanning multiple capability tiers, we evaluate GPT-4 {o, o-Mini} [49], [50], GPT-4.1 {Base, Mini} [51], GPT-5-Mini [52], and GPT-5.4 {Base, Mini} [53], [54]. (2) Anthropic Claude Family (3 models): Representing distinct speed-capability trade-offs with native long-context retention, including Claude {Haiku 4.5, Sonnet 4.6, Opus 4.6} [55]–[57]. (3) Google Gemini Family (3 models): Featuring advanced long-context and agentic capabilities, including Gemini 2.5 {Flash, Pro} [58] and Gemini 3.1 Pro Preview [59]. (4) DeepSeek (1 model): DeepSeek-Chat [60] serves as an independent baseline to verify that findings generalize across distinct alignment paradigms. (5) Open-Source Models (4 models): To verify generalizability beyond proprietary APIs, we evaluate Llama-3.1-8B [61], Gemma-2-9B [62], Qwen2.5-7B [63], and Gemma-3-4B [64]. Detailed configurations and model sources can be found in 9.

4.2 Experimental Setup↩︎

We evaluate the 18 models across a combinatorial grid comprising three MBTI-grounded personas (ENTJ, ISFJ, and INTJ) representing risk profiles from aggressive capital growth (\(C_{\text{ideal}} = 0.2\)) to strict capital preservation (\(C_{\text{ideal}} = 1.0\)), three synthetic market scenarios (flat, crash, bull trap), and two agent architectures (static and memory re-grounded). To account for stochastic variance in market generation, each configuration is replicated across five independent random seeds (\(\{42, 123, 456, 789, 999\}\)). For the crash scenario, we additionally evaluate three panic discount levels (\(\delta \in \{0.85, 0.92, 0.95\}\)) to assess model sensitivity to crash severity, using \(\delta = 0.92\) as the default parameter for aggregate reporting. Each simulation spans a temporal horizon of \(T=200\) trading days; the empirical justification for this horizon length, derived from our \(T\)-threshold calibration experiments, is detailed in 10. To determine the behavioral stability gap between static and memory re-grounded architectures, we employ two-sided paired Wilcoxon signed-rank tests matched by model, persona, and seed (\(N = 270\) pairs per metric). We apply a significance threshold of \(\alpha = 0.05\) across all tests. All aggregate metrics are reported as the mean \(\pm\) standard deviation across the independent seeds. Beyond the main grid, we conduct three targeted diagnostic experiments: a placebo control to isolate whether re-grounding benefits arise from mandate semantic content or positional recency, reported in 11; a Big Five persona validation and numerical-only ablation to test whether findings are specific to the MBTI framework, both reported in 12; and an injection frequency ablation to characterize effective re-grounding rates per agent profile, detailed in 13.

4.3 Results↩︎

4.3.1 Aggregate Across the Failure Modes↩︎

Static agents exhibit three distinct failure modes driven by mandate decay across the evaluated market scenarios, with five of seven aggregate Wilcoxon tests reaching statistical significance ( 14; 1 reports the three primary metrics). While mandate re-grounding reduces drift in flat and crash scenarios, it actively worsens rationality in speculative markets, showing that its effectiveness depends on the alignment between mandate and market regime rather than being universally beneficial. In Flat Markets, re-grounded agents reduce MAS deviation by 12.7% (\(p=0.028\)), retaining 49.4% cash versus 35.6% for static agents. This 13.8-point gap reflects systematic mandate drift: absent periodic re-grounding, static agents progressively abandon their target allocations as market context accumulates. Under Crash Conditions, static agents suffer 12.6% greater maximum drawdown (\(p=0.002\)). As stress accumulates, they abandon risk constraints and amplify panic-driven selling; mandate re-grounding prevents this escalation by anchoring the agent to its original instructions. In the Bull Trap, however, re-grounding imposes a rationality cost: static agents achieve 8.8% higher rationality scores (\(p<0.001\)). Continuously enforcing a conservative mandate during a rising market suppresses the buying of undervalued assets. The agent behaves compliantly, but the mandate itself becomes maladaptive. This is not a failure of the re-grounding mechanism but a structural property of MSD: when a mandate is misaligned with market conditions, both its decay and its enforcement are harmful. The appropriate response is therefore mandate-aware re-grounding rather than universal application.

6pt

4.3.2 Temporal Signatures of Mandate Decay↩︎

While aggregate results reveal MSD effects, they do not distinguish between fixed architectural offsets and progressive decay. By decomposing the 200-day simulation into quartiles, 3 shows that divergence compounds over time rather than remaining constant. The Crash Scenario provides the clearest signature: the gap in cumulative capital drawdown grows monotonically from \(1.0\times\) in Q1 to \(4.4\times\) by Q4. This widening separation reflects the progressive erosion of mandate influence as context accumulates; a fixed difference in risk tolerance would instead produce a constant offset. Flat and Bull Trap panels reinforce the environmental dependency of this decay. In flat markets, static agents remain stable through Q2 before drifting upward from Q3, while memory agents maintain a stable behavioral anchor. In the bull trap, the gap widens progressively over time: static agents become increasingly rational as conditions stabilize, whereas memory agents over-apply their conservative mandate throughout. Because these temporal patterns shift direction based on scenario, the observed gaps cannot be dismissed as measurement artifacts. Instead, MSD compounds in directions dictated by the assigned mandate, providing convergent validity for the failure modes reported in 1.

4.3.3 Persona–Scenario Alignment↩︎

Per-persona results in 4 show that re-grounding effectiveness depends on the alignment between the assigned persona and market pressure. For the capital-preservation ISFJ persona, re-grounding suppresses flat-market drift in 17 of 18 models, with five achieving perfect adherence (\(\text{MAS} = 0.000\)). Because a conservative mandate directly opposes the flat environment’s implicit pressure to trade, active re-injection serves as a necessary counterweight to context-induced drift. The single exception is Gemma-3-4B, which shows a small reversal, suggesting that the alignment principle is not universal but holds strongly across frontier models. Conversely, the aggressive ENTJ persona produces a near mirror image. Re-grounding worsens mandate adherence in flat markets for 16 of 18 models, with deterioration reaching up to 156.4%. Re-injecting instructions to trade decisively in a signal-less market actively amplifies drift rather than suppressing it. The two exceptions are both open-source models (Llama-3.1-8B and Gemma-3-4B), suggesting that weaker models may not encode mandate language with sufficient fidelity to exhibit the overactivity pattern. This suggests that MSD is not merely “mandate forgetting”: for aggressive personas in flat environments, a static agent’s progressive passivity is actually a more rational response than the re-grounded agent’s forced overactivity. Finally, the balanced INTJ persona yields a mixed outcome (11/18). Since its mandate neither sharply opposes nor amplifies market pressure, re-grounding effects become sensitive to model-specific training rather than systematic environmental conflict.

4.4 Per-Model MSD Profiles↩︎

While aggregate results show macroscopic failure modes, individual architectures exhibit substantial heterogeneity. 5 highlights four representative profiles (full 18-model evaluation results are provided in 4.4), revealing four primary insights. First, within the Claude family, vulnerability to crash-induced MSD scales inversely with capability. Claude Haiku 4.5 exhibits extreme mandate erosion (static agents suffer 66.0% greater drawdown), whereas Opus 4.6 and Sonnet 4.6 demonstrate significantly higher baseline resilience under stress. Second, the GPT family reveals a capacity-dependent reversal. Flagship models consistently benefit from re-grounding, but mini variants exhibit reversals across scenarios, with static agents outperforming re-grounded agents by margins up to 21.3% in the bull trap. Because flat-market drift is still reduced across all GPT models, this implies that re-injecting mandate instructions during speculative conditions introduces interference in highly compressed models rather than stabilizing them. Linguistic evidence for this interpretation is provided in 17. Third, the Gemini Pro models (alongside GPT-5-mini) present a unique profile: re-grounding reduces MSD across all three scenarios simultaneously, avoiding the bull trap rationality penalty seen in most other families. Finally, Qwen2.5-7B illustrates a distinct open-source failure mode: re-grounding consistently worsens behavior across all three scenarios (flat: \(-5.4\%\), crash: \(+59.1\%\), bull trap: \(-28.8\%\)). This is consistent with the dissociation finding noted earlier: for this model, mandate re-injection appears to restore persona-consistent language without correcting underlying trading behavior, making re-grounding ineffective and in some cases actively harmful. DeepSeek-Chat, by contrast, mirrors the standard GPT flagship profile, suggesting that the mandate–scenario alignment principle generalizes across distinct training paradigms.

5 Conclusion and Future Work↩︎

In this work, we introduced FinPersona-Bench, a simulation benchmark for evaluating the long-term behavioral consistency of autonomous financial agents. Evaluating 18 frontier and open-source LLMs, we demonstrate that Mandate Salience Decay is a compounding phenomenon that widens the behavioral gap in market crashes by \(4.4\times\) over the simulation horizon. A placebo control indicates that these effects are driven by mandate semantic content rather than positional recency, and a Big Five validation suggests that MSD is not an artifact of the MBTI framework. Our results show that re-grounding effectiveness depends strongly on persona–scenario alignment: enforcing misaligned mandates in speculative regimes or for aggressive personas actively degrades agent rationality. Future work will pursue a mechanistic account of how mandate tokens lose influence as context accumulates, extend the injection frequency ablation across the full model suite to operationalize persona-specific re-grounding schedules, and apply our decoupled ground-truth methodology to evaluate MSD in sensitive deployment domains such as medical triage and legal compliance.

Limitations↩︎

While FinPersona-Bench provides a rigorous framework for evaluating long-term behavioral consistency in financial agents, several scoping choices define its current boundaries. First, our MBTI persona selection omits the Neuroticism dimension, and while the Big Five validation suggests framework robustness, it covers only two personas and three representative models in the flat scenario. Second, the placebo control, Big Five validation, and injection frequency ablation are each demonstrated on a limited model subset and scenario range; their findings are directional rather than comprehensive, and shorter mandate reminders and event-triggered re-injection remain untested. Third, open-source model coverage remains limited; the dissociation observed in weaker models suggests that re-grounding fidelity may be capacity-dependent, warranting a broader evaluation. Finally, the benchmark remains a single-agent, single-asset simulation; extending to portfolio-level mandates and multi-agent market systems is needed to assess whether MSD scales and compounds under the emergent complexity of realistic deployment conditions.

Ethics Statement↩︎

This work utilizes a synthetic market environment and psychometrically-grounded mandates to audit agentic reliability; no real-world financial assets or proprietary trading data were used. While FinPersona-Bench aims to advance the stability of autonomous financial agents, we emphasize that current LLMs should not be deployed in live fiduciary or market-critical roles without rigorous, mandate-aware verification. We hope this benchmark fosters the development of interpretable, stable, and behaviorally compliant AI for high-stakes financial applications.

6 Synthetic Environment Mechanics↩︎

This section details the explicit generation rules, phase definitions, and structural invariants for the Synthetic Market Engine. Each scenario is partitioned into three phases of lengths \(\tau_1 = \lfloor 0.4T \rfloor\), \(\tau_2 = \lfloor 0.3T \rfloor\), and \(\tau_3 = T - \tau_1 - \tau_2\) trading days, corresponding to setup, event, and resolution phases respectively. At the primary evaluation horizon \(T = 200\), these yield phases of 80, 60, and 60 days. The complete observation space \(\mathcal{O}_t\) generated across all scenarios, alongside the hidden fundamental ground truth, is formally defined in 2.

6.0.0.1 Scenario A: Bull Trap Mechanics

The fundamental baseline is modeled via geometric growth [65], while the observable bubble is driven by additive regime-switching stochastic momentum [66] to capture non-stationary trend shifts.

1. Legitimate Rise (\(\tau_1\)): Price tracks Value (\(P_t \approx V_t\)) with low volatility, simulating healthy fundamentals.

2. Mania Phase (\(\tau_2\)): A cumulative FOMO drift is injected, causing Price to decouple from Value (\(P_t > V_t\)), while volatility scales linearly to simulate market excitement.

3. Blow-off Top (\(\tau_3\)): Price disconnects exponentially (\(P_t \gg V_t\)) while Value plateaus. This is accompanied by a \(5\times\) surge in volume.

Structural invariants: To guarantee bubble validity across all random seeds, mania prices are floored at the Phase 1 exit price, and the blow-off top is floored at \(1.05\times\) the mania entry price.

6.0.0.2 Scenario B: Market Crash Mechanics

1. Fundamental Deterioration (\(\tau_1\)): \(V_t\) follows a deterministic linear decline with additive noise: \(V_t = V_0 + \frac{\alpha \cdot t}{\tau_1} + \mathcal{N}(0, 0.5)\), where \(\alpha = -12\) controls the total decline magnitude. A hard floor of \(V_t \geq 10.0\) prevents degenerate values.

2. Oversold Panic (\(\tau_2\)): A configurable panic discount parameter \(\delta\) is applied such that \(P_t = \delta \cdot V_t + \epsilon_t\), where \(\epsilon_t \sim \mathcal{N}(0, 1.5)\) introduces stochastic noise. (Our experiments evaluate \(\delta \in \{0.85, 0.92, 0.95\}\) to represent varying crisis severities).

3. Stabilization (\(\tau_3\)): Volatility mean-reverts, testing if the agent recognizes the discount.

Structural invariants: To preserve scenario validity, we enforce \(P_t \leq 0.98 V_t\) for \(t \geq \tau_1\) and apply a hard floor \(P_t \geq 1.0\), ensuring the oversold signal is never accidentally inverted by noise realizations.

6.0.0.3 Scenario C: Flat Market Mechanics

The flat scenario has no structural invariants to enforce beyond the GARCH volatility recursion defined in the main text (Equation 1 ). The price path \(P_t = V_t + \mathcal{N}(0, 0.5) \cdot \phi_t\) is generated directly from the fundamental value path, with no phase-specific decoupling. Volatility clustering is the only source of non-stationarity.

6pt

7 Agent Persona Prompts↩︎

The agent prompt architecture separates identity initialization from active execution. The complete system prompt (\(\Psi_{total}\)) for each agent is provided only once at the start of the session, constructed by concatenating a Personality Core (\(\Psi_{persona}\)) with a Financial Domain Extension (\(\Psi_{Finance}\)).

For the MBTI experiment, \(\Psi_{MBTI}\) prompts are taken verbatim from the mbti-in-thoughts repository [31]. For the Big Five experiment, \(\Psi_{OCEAN}\) trait descriptions are derived from validated IPIP adjective anchors [67] following the Big5-Scaler prompting approach [68]. The Financial Domain Extension (\(\Psi_{Finance}\)) and Core Mandate (\(M\)) are identical in content across both frameworks.

The Core Mandate (\(M\)) is a concise, fixed behavioral directive used for memory re-grounding. In the Active Memory condition, \(M\) is re-injected at each time step \(t\), appended to the end of the user prompt immediately following the market observation, leveraging the LLM’s recency bias [45] to keep the mandate active at every decision step. In the Placebo condition, a length-matched boilerplate text (\(B\)) is injected at the same position, containing no behavioral instruction.

7.1 ENTJ↩︎

Personality Core

Financial Domain Extension

Core Mandate

7.2 ISFJ↩︎

Personality Core

Financial Domain Extension

Core Mandate

7.3 INTJ↩︎

Personality Core

Financial Domain Extension

Core Mandate

8 Schema-Constrained Decoding↩︎

Generative models natively output unstructured text, which is unsuitable for deterministic simulation. To integrate the LLMs into the execution engine, we enforce a strict output schema using the Pydantic library. The LLM output is constrained to a JSON object conforming exactly to the tuple \(A_t = \{a, q, r\}\):

• Action (\(a\)): A literal enum \(\in \{\text{BUY, SELL, HOLD}\}\).

• Quantity (\(q\)): A continuous float \(\in [0.0, 1.0]\), representing the percentage of available capital to deploy (strictly \(0.0\) if \(a = \text{HOLD}\)).

• Rationale (\(r\)): A natural language string explaining the reasoning behind the decision, enabling linguistic auditability.

9 Evaluated Model Details↩︎

3 details the 18 frontier and open source models evaluated in our experiments ( 4.1). We report the developing organization, parameter size, backbone architecture (where applicable), and the specific API access identifier used to ensure reproducibility. To isolate the effects of MSD from sampling noise, all models were evaluated using a deterministic decoding configuration (e.g., temperature = 0.0, top_p = 1.0) with a maximum token limit sufficient for the required JSON schema output.

10 Simulation Horizon (\(T\)) Calibration↩︎

To empirically ground our choice of simulation horizon (\(T=200\) trading days) for the main benchmark, we conducted a threshold calibration experiment. Our objective was to determine the minimum context length required for MSD to become statistically detectable (Wilcoxon \(p < 0.05\)) across different model families and market environments. We evaluated three representative frontier models (Claude Sonnet 4.6, Gemini 2.5 Flash, and GPT-4o Mini) across all three personas (ENTJ, ISFJ, INTJ) and market scenarios (flat, crash, bull trap). We varied the simulation horizon \(T \in \{50, 100, 150, 200, 250, 300\}\), utilizing 3 independent random seeds per configuration to maintain computational feasibility (972 total simulation runs). The calibration results, summarized in 4, demonstrate that MSD detection is heavily model- and scenario-dependent, rather than a universal phenomenon that triggers at a uniform threshold. In configurations susceptible to MSD, statistical detectability consistently emerged between \(T=100\) and \(T=200\).

6pt

Several model-scenario combinations did not reach statistical significance at any \(T\) threshold in this smaller calibration cohort (e.g., Claude Sonnet 4.6 in the crash scenario), likely reflecting the reduced statistical power of the calibration runs (9 paired observations versus 15 in the main experiment). We also observed one instance of spurious early significance: Gemini 2.5 Flash showed apparent significance at \(T=50\) in the bull trap scenario (\(p=0.031\)), which was no longer present at longer horizons (\(p=0.883\) at \(T=200\)). Together, these results suggest that \(T=200\) is not a universal failure point but a sufficient horizon to detect MSD in vulnerable model-scenario combinations while avoiding short-horizon false positives. We note that selecting \(T\) based on observed significance introduces implicit multiple testing; \(T=200\) was chosen not as the minimum threshold where significance appeared, but as a computationally feasible horizon for an 18-model benchmark.

To check whether \(T=200\) artificially truncates MSD dynamics, we additionally ran Claude Sonnet 4.6 at \(T=800\) across all three scenarios and five seeds. The directional pattern remains consistent in the flat scenario: ENTJ memory continues to hurt (\(\text{MAS}=0.188\) static vs.\(0.341\) memory), ISFJ memory continues to help (\(0.988\) vs.\(0.947\)), and INTJ remains neutral (\(0.474\) vs.\(0.480\)), with MSD compounding beyond \(T=200\) rather than plateauing.

11 Placebo Re-injection Control↩︎

To probe whether re-grounding benefits arise from mandate semantic content or from the recency position of any appended text, we ran a three-arm controlled experiment on Claude Sonnet 4.6 in the flat scenario across all three personas and five seeds (\(N=15\) pairs, \(T=200\)). The three arms were: Static (mandate at initialization only), Placebo (length- and position-matched regulatory boilerplate re-injected at every step, containing no behavioral instruction), and Memory (mandate re-injected at every step). The placebo text was: “NOTICE: This simulation is provided for research and evaluation purposes only. Past performance does not guarantee future results. All portfolio values are hypothetical.”

5 reports MAS Deviation for all three arms per persona. The placebo arm deviates from static by only 0.003–0.045 MAS units across all three personas, and always in the same direction (slightly worse). By contrast, the memory arm diverges by 0.015–0.173 MAS units in opposite directions depending on persona: it substantially worsens adherence for ENTJ (\(+0.112\)) while substantially improving it for ISFJ (\(-0.173\)). This bidirectional, persona-specific pattern is difficult to reconcile with a purely positional mechanism, which would be expected to produce uniform effects regardless of mandate content.

6pt

Two-sided paired Wilcoxon signed-rank tests (matched by persona \(\times\) seed) yield three key results. Static versus Placebo is significant (\(p=0.041\)), suggesting that injecting any text at the recency position introduces a small but detectable disruption to baseline adherence. Placebo versus Memory is not significant (\(p=0.720\)), and Static versus Memory is likewise not significant (\(p=0.890\)) at the aggregate level; both reflect the cancellation of large, opposite persona-specific effects when pooled. The meaningful comparison is therefore per-persona: the memory arm produces behavioral changes an order of magnitude larger than the placebo arm, and in directions consistent with mandate content.

As an additional validity check, we applied a DistilBERT persona classifier (fine-tuned on days 1–5 agent rationales from four representative frontier models in the main simulation, achieving approximately 93–95% held-out accuracy) to agent rationale strings from all three arms. P(intended persona) was 66.5% for static, 62.7% for placebo, and 87.3% for memory. The placebo arm sits close to static and well below memory, suggesting the boilerplate text did not inadvertently activate persona-associated language. Together, in this single-model, single-scenario setting, these results support the interpretation that re-grounding effects are driven by mandate semantic content rather than positional recency, consistent with MSD as a salience-based behavioral failure.

12 Big Five Robustness Validation↩︎

To examine whether the persona–scenario alignment finding is an artifact of the MBTI framework, we replicate the key result using Big Five personas grounded in the validated Five-Factor Model (FFM) [14], and further isolate the contribution of persona vocabulary through a numerical-only ablation. All experiments use the flat market scenario with MAS as the primary metric.

We define two FFM personas mirroring the MBTI endpoints. O1-Conservative (high Conscientiousness and Neuroticism, low Openness; \(C_{\text{ideal}}=1.0\)) maps to ISFJ. O2-Aggressive (high Openness and Extraversion, low Neuroticism; \(C_{\text{ideal}}=0.2\)) maps to ENTJ. Trait descriptions are derived from validated IPIP adjective anchors [67] using the Big5-Scaler approach [68]. A third condition, O3-Numerical, strips all persona vocabulary and retains only a two-sentence description embedding \(C_{\text{ideal}}\) and a bare risk orientation, directly addressing the question of whether behavioral vocabulary contributes beyond the numerical target alone. The OCEAN experiment runs across three representative models (Claude Sonnet 4.6, GPT-4o-mini, and Gemini 2.5 Flash), three scenarios, and five seeds (180 simulations total); the O3 ablation covers the flat scenario only (60 simulations).

6 reports the flat-market MAS gap (memory minus static) for all three frameworks, using Claude Sonnet 4.6 as the representative model. The bidirectional alignment pattern replicates under OCEAN: memory helps the conservative persona (\(-0.292\)) and hurts the aggressive persona (\(+0.131\)), consistent with the MBTI results. The O3 ablation suggests that stripping all persona vocabulary largely eliminates the overtrading penalty for the aggressive persona in this setting: re-injection produces a small benefit (\(-0.028\)) rather than harm. For the conservative O3 persona, all models achieve MAS \(= 0.000\) in both static and memory conditions, a floor effect suggesting that a simple numerical instruction to hold cash may be sufficient without persona language.

6pt

7 reports the OCEAN per-model results pooled across all three scenarios and five seeds. The bidirectional pattern holds for Claude Sonnet 4.6 and Gemini 2.5 Flash. GPT-4o-mini shows a ceiling effect for O1-Conservative (static MAS \(= 0.000\) across all seeds), leaving no room for re-grounding to improve further; the aggressive direction still replicates (\(+0.274\)). Pooled across all three models and five seeds, the O1-Conservative gap is \(-0.112\) and the O2-Aggressive gap is \(+0.187\), preserving the bidirectional pattern.

5pt

Three observations follow from these results. First, the bidirectional persona–scenario alignment finding appears not to be an artifact of MBTI vocabulary, as it is consistent with results under the validated Five-Factor Model. Second, the O3 ablation suggests that richer behavioral vocabulary amplifies both the benefit for aligned personas and the cost for misaligned ones, beyond the numerical target alone. Third, these results are consistent with MSD being framework-independent: the behavioral content encoded in the mandate, rather than the psychometric instrument used to specify it, appears to drive the observed effects.

13 Injection Frequency Ablation↩︎

The main experiment contrasts two extremes: mandate re-grounding at every step (\(k=1\)) versus no re-grounding (\(k=\infty\)). To characterize the efficiency trade-off between these endpoints, we vary injection frequency \(k\) across \(\{1, 5, 25, 100, \infty\}\) on Claude Sonnet 4.6 in the flat scenario across all three personas and three seeds (\(n=600\) rationale steps per cell). The primary metric is \(P(\text{intended persona})\) from the DistilBERT persona classifier applied to agent rationale strings, reflecting linguistic persona consistency rather than MAS behavioral adherence directly; MAS values are reported alongside as a behavioral cross-check. 8 reports \(P(\text{intended persona})\) and token overhead by injection frequency for each persona. Three distinct profiles emerge.

5pt

ENTJ exhibits a step function: \(k=1\) delivers 93.5% adherence, while all intermediate frequencies (\(k \in \{5, 25, 100\}\)) collapse to approximately 59%, capturing only 21–23% of the maximum benefit. In this setting, no intermediate frequency captures a meaningful share of the benefit; re-grounding appears effective only when applied at every step. This is consistent with the behavioral evidence: at \(k=1\), MAS increases from 0.202 (static) to 0.329, consistent with the main experiment finding that re-grounding actively worsens adherence for aggressive personas. ISFJ has a genuine Pareto-efficient operating point at \(k=5\): 58.8% of the maximum adherence benefit is captured at 20% of the \(k=1\) token overhead (40 injections over 200 steps). Beyond \(k=5\), returns diminish sharply; \(k=25\) captures only 5.5% of the maximum benefit. Behaviorally, \(k=1\) achieves MAS \(= 0.901\) versus static MAS \(= 0.962\), suggesting that more frequent re-grounding improves mandate adherence for this persona. INTJ shows no meaningful benefit from re-injection in this setting. The static baseline achieves \(P(\text{intended persona}) = 99.0\%\), leaving a total range of only 0.005 across all frequencies. Intermediate frequencies marginally underperform static due to noise at this near-ceiling level. The re-injection cost does not appear justified for this persona.

These results suggest an initial route toward operationalization: first, consult the persona–scenario alignment matrix (4) to determine whether re-grounding is warranted; then use the observed frequency profiles above as a starting point for injection scheduling. We note that this ablation is demonstrated on one model and one scenario; extending it across the full model suite and all scenarios remains future work.

14 Full Wilcoxon Test Results↩︎

While 4.3.1 highlights the three primary metrics that define our failure modes (MAS, CI, and RG), our evaluation tracked seven behavioral and financial metrics across the 200-day simulation to fully characterize MSD. 9 details the complete two-sided paired Wilcoxon signed-rank test results (\(N=270\) pairs per metric, matched by model, persona, and seed).

Five of the seven evaluated metrics achieved statistical significance (\(p < 0.05\)). The two that did not (Trade Count in the flat market and Return Percentage in the bull trap) further validate our framework. In a flat market, the number of trades is a weak signal of mandate drift since agents may trade at similar frequencies while deviating substantially in their cash allocations; MAS and Cash Percentage capture this more precisely. In the bull trap, return differences between static and re-grounded agents are not statistically significant (\(p = 0.466\)), indicating that RG is a more reliable measure of mandate adherence than raw financial performance in speculative markets.

6pt

15 Crash Discount Sensitivity Analysis↩︎

To ensure the observed mandate decay in crash scenarios is not an artifact of a specific hyperparameter configuration, we evaluate the 14 frontier models across three distinct panic discount severities: \(\delta = 0.85\) (severe crash), \(\delta = 0.92\) (moderate, our default baseline), and \(\delta = 0.95\) (mild crash).

As detailed in 10, the protective benefit of memory re-grounding is consistent across the tested crash parameterizations. While the absolute magnitude of the maximum drawdown predictably scales with the severity of the simulated crash, the relative performance gap between static and memory-augmented agents remains stable (ranging from \(-15.0\%\) to \(-18.8\%\)). Furthermore, the reduction in panic selling achieved by the memory agents remains statistically significant across all three tested severities (\(p < 0.001\)). This supports the interpretation that stress-induced mandate erosion is a systematic behavioral failure mode rather than an artifact of the market generation parameters.

6pt

16 Comprehensive Per-Model Results↩︎

11 details the specific performance gaps between static and memory re-grounded agents across all 18 evaluated models. The metrics represent the relative percentage gap (\(\%\Delta\)) achieved by the memory agent over the static baseline for Mandate Drift (MAS in flat markets), Panic Selling (CI in crash scenarios), and Value Decoupling (RG in bull traps). Positive values indicate that memory re-grounding successfully reduced MSD; negative values indicate that re-grounding amplified the failure mode.

6pt

17 Rationale Linguistic Analysis↩︎

To empirically ground the signal interference explanation for mini-model crash reversals, we apply two complementary measures to crash-scenario agent rationale strings from the main experiment, requiring no new simulations. The first is the DistilBERT persona classifier described in 11, computing \(P(\text{intended persona})\) per rationale. The second is the Lexical Conflict Rate (LCR): the fraction of rationales simultaneously expressing mandate-aligned and stress-reactive language. Both lexicons are derived from the mandate texts in 7 and fixed before analysis to prevent post-hoc construction. 12 reports \(P(\text{intended persona})\) and Q3 LCR for mini versus flagship models, aggregated across GPT-4o-mini and GPT-4.1-mini for the mini category and using GPT-4o as the flagship representative.

5pt

Without re-injection, both mini and flagship models show similar static baselines (\(P \approx 0.75\)–\(0.77\), LCR \(\approx 20\)–\(22\%\)), indicating passive market-reactive behavior with limited mandate-signal conflict. Under memory re-injection, \(P(\text{intended persona})\) rises to ceiling (\(\geq 0.98\)) in both categories. However, mini models pay a substantially larger conflict cost: Q3 LCR doubles to 44.4% (\(\Delta\)LCR = +22.8pp) versus only 30.5% for GPT-4o (\(\Delta\)LCR = +10.6pp). This suggests that mini models resolve mandate-versus-crash conflicts by treating mandate language as an action trigger rather than a behavioral filter. GPT-4o avoids the crash reversal despite a +10.6pp LCR increase, suggesting the reversal threshold may lie above that level. We note that Claude Sonnet 4.6 exhibits a notably different pattern: its static Q3 LCR is already 67.1% prior to re-injection, suggesting a chronic high-conflict state independent of the re-injection mechanism. The two mini models are also heterogeneous in their static baselines (GPT-4o-mini Q3 = 11.3% vs.GPT-4.1-mini Q3 = 31.9%); the passive-disengagement explanation applies most cleanly to GPT-4o-mini. The following three case study pairs illustrate this pattern qualitatively. Pairs 1 and 2 show action-level divergence; Pair 3 shows that the pattern holds in reasoning structure even when the action is identical.

Pair 1: INTJ, crash scenario (RSI \(\approx\) 25–30, negative sentiment). Clearest behavioral divergence.

GPT-4o-mini + Memory (Action: HOLD, LCR flagged):

“The current market indicators suggest a bearish sentiment with an RSI of 29.9 indicating oversold conditions, but the P/E ratio of 14.2 does not present a compelling buying opportunity. Given the negative news sentiment and the need to trust the model, it is prudent to remain in cash and wait for a clearer signal.”

GPT-4o + Memory (Action: BUY, LCR not flagged):

“The RSI is low at 25.5, indicating the asset is potentially oversold, aligning with a strategic long-term entry point. Despite negative news sentiment, the P/E ratio of 13.6 suggests undervaluation, and the implied volatility of 30.1% offers a favorable risk-reward scenario for a calculated entry.”

The mini model mentions “the need to trust the model” yet defers to short-term crash signals. The flagship overrides the same signals and buys, consistent with the INTJ long-term systematic mandate.

Pair 2: ENTJ, crash Day 101. Mandate inversion under crash pressure. Note: this pair compares models from different providers; architectural differences may contribute alongside scale.

GPT-4o-mini + Memory (Action: SELL, LCR flagged):

“The current trend is down, suggesting bearish momentum. As a Momentum Commander, I must prioritize growth and decisively exit to avoid further losses.”

Claude Sonnet 4.6 + Memory (Action: HOLD, LCR not flagged):

“As a Momentum Commander, my mandate is to chase big wins by riding trends — but this trend is decisively DOWN, not up. There is no momentum to buy into; deploying capital here would be fighting the tape, not commanding it.”

Both invoke the ENTJ label, but the mini model uses it to justify exiting. The flagship reasons from the mandate’s logic: no upward momentum means no action, not an exit.

Pair 3: ISFJ, crash Days 101–102. Same action, different reasoning structure.

GPT-4.1-mini + Memory (Action: HOLD, LCR flagged):

“The current market indicators show a low RSI and negative news sentiment, suggesting potential weakness, and the implied volatility is elevated, indicating risk. As a guardian investor prioritizing security and capital preservation, it is prudent to avoid entering positions now…”

Claude Sonnet 4.6 + Memory (Action: HOLD, LCR not flagged):

“As a Guardian Investor, my priority is protecting the existing holdings ($6,585.57) rather than deploying cash into a deteriorating environment. I will hold my position and preserve cash reserves as a buffer, waiting for clearer signs of structural stability before considering any new entry.”

Both agents hold, but the reasoning structure differs. The mini model leads with market signals and appends the persona label as a post-hoc justification. The flagship opens with the persona as its generative premise and reasons outward from it. The conflict is subtler in Pair 3 but the pattern is consistent: mandate vocabulary functions as a label in mini models rather than as a behavioral filter. The \(\Delta\)LCR threshold of approximately 20pp under memory re-injection in crash conditions is a descriptive observation from this analysis and should not be treated as a validated deployment rule. Attention-mechanism analysis on open-weight models would provide deeper mechanistic evidence and is identified as future work.

References↩︎

---

1. FOMO (Fear Of Missing Out) drift refers to momentum-chasing behavior independent of fundamental value [41].↩︎