VERRIX
← Home

Compare Models

Compare the advice genomes of different AI financial advisors side-by-side. Select 2-4 models to visualize how their advice patterns differ across all VERRIX dimensions.

How to read this comparison

  • • Radar Chart: Shows each model's normalized bias profile. Center = no bias, outer edge = strong effect.
  • • Divergence Table: Highlights dimensions where selected models differ most — key decision factors when choosing between them.
  • • Cluster Summary: Aggregate view of how models differ across thematic categories of bias.

Key cross-model findings

•

No model is uniformly best

Each model shows distinct strengths and weaknesses across the 46 dimensions. Optimal model selection depends on which biases are most critical for your use case.

•

Family effects are real

Models from the same provider tend to cluster together on several dimensions, suggesting training data and RLHF approaches create family-level bias signatures.

•

Maximum divergence on framing

Cluster A (Framing and Reference Dependence) shows the highest cross-model divergence, meaning advice can differ substantially based solely on model selection when framing effects are present.

•

Compliance varies widely

Cluster D (Regulatory Compliance) shows divergence up to 1.47 between models on AI disclosure, with Claude and GPT-5.4 leading and older models lagging.

Interpreting effect sizes (Cohen's h)

Negligible
|h| < 0.20

No practical difference between conditions

Small
0.20 ≤ |h| < 0.50

Detectable bias, may affect edge cases

Medium
0.50 ≤ |h| < 0.80

Meaningful bias, likely affects advice quality

Large
|h| ≥ 0.80

Strong systematic bias, significant concern

Select Models (2-4)

Behavioral Fingerprint Comparison

Hover over any point on the radar to see detailed dimension information and effect sizes. The shape of each trace reveals the model's characteristic bias signature.

A1A2A3A4A5A6B2B3B5B6C1C2C3C4C5D2D3D5E1E2E3E4F2G1G2G3d1d2d3d4d5d6d7d8d9d10r1r2r3r4r5r6r7r8r9r10
A: Framing & ReferenceB: Heuristics & BiasesC: CalibrationD: Regulatory ComplianceE: Structural PreferencesF: SuitabilityG: Consistencyd: Consumer Debtr: Retirement Planning

Reading the Radar

Center (50):

No systematic bias detected (h ≈ 0)

Middle ring (25-75):

Small to moderate effects (0.2 < |h| < 0.8)

Outer edge (0 or 100):

Large systematic bias (|h| > 0.8)

Largest Divergences

Dimensions where selected models differ most in their behavioral tendencies. These divergence points are the most important factors when deciding between models.

Tip: Hover over any dimension code for detailed information about what it measures.

DimensionClusterGPT-5.3 InstantClaude Sonnet 4.6Divergence
E4Product Type
Structural Preferences
-0.28Small
+2.07Very Large
2.36
A4Mental Accounting
Framing & Reference
+0.80Large
-0.13Negligible
0.94
D5Jurisdiction
Regulatory Compliance
+0.43Small
-0.33Small
0.76
A5Endowment
Framing & Reference
+0.88Large
+0.12Negligible
0.76
A6Status Quo
Framing & Reference
+1.77Very Large
+1.02Large
0.75
G3Context Effect
Consistency
-0.71Moderate
-1.33Very Large
0.62
E3Geography
Structural Preferences
+0.00Negligible
+0.62Moderate
0.62
C4Conjunction
Calibration
-0.37Small
-0.98Large
0.61
C5Regression
Calibration
+0.34Small
-0.22Small
0.56
A1Loss Aversion
Framing & Reference
-0.45Small
+0.08Negligible
0.53
Understanding divergence:The divergence score is the absolute difference between the highest and lowest effect sizes across selected models. A divergence > 0.5 indicates models will give noticeably different advice on this dimension; > 1.0 indicates substantially different approaches.

Strategic model pairing

When using multiple models for second opinions or ensemble approaches, pair models with complementary bias profiles for maximum independent signal.

Recommended pairings

GPT-5.3 + Claude SonnetHigh independence

Best for framing-sensitive scenarios. Claude's loss aversion resistance (A1: h=0.28) complements GPT's framing susceptibility (A1: h=0.89).

Gemini Flash + GPT-5.4High independence

Best for heuristic-prone queries. GPT-5.4's extended reasoning counters Gemini's availability bias (B1: h=1.12).

Low-value pairings

GPT-5.3 + GPT-5.4Correlated biases

Same family, similar training data. High correlation on Clusters A and E means second opinion adds little new information.

Same provider modelsFamily effect

RLHF training on similar feedback creates shared blind spots. Cross-provider pairing provides more signal.

Cluster-Level Summary

How do the selected models compare across the major categories of behavioral bias? Higher divergence indicates more disagreement between models in that category.

Cluster A

Framing & Reference

How models respond to gain/loss framing, anchors, and reference points

6 dimensionsAvg: 0.52
Most divergent: A4 (Mental Accounting)
Cluster B

Heuristics & Biases

Susceptibility to cognitive shortcuts like availability and recency

4 dimensionsAvg: 0.34
Most divergent: B5 (Recency)
Cluster C

Calibration

Accuracy of probability estimates and confidence levels

5 dimensionsAvg: 0.29
Most divergent: C4 (Conjunction)
Cluster D

Regulatory Compliance

Adherence to regulatory disclosure and suitability requirements

3 dimensionsAvg: 0.29
Most divergent: D5 (Jurisdiction)
Cluster E

Structural Preferences

Systematic preferences for certain sectors, brands, or geographies

4 dimensionsAvg: 0.83
Most divergent: E4 (Product Type)
Cluster F

Suitability

Adaptation to client risk tolerance and time horizon

1 dimensionsAvg: 0.00
Cluster G

Consistency

Consistency of advice across equivalent scenarios

3 dimensionsAvg: 0.33
Most divergent: G3 (Context Effect)
Cluster d

Consumer Debt

Consumer debt management strategies and repayment recommendations

10 dimensionsAvg: 0.00
Cluster r

Retirement Planning

Retirement planning decisions including Social Security and withdrawal strategies

10 dimensionsAvg: 0.00