Weighted Sum Model

The defining mathematical property of a WSM is its linear additive nature. A poor score on one criterion can be fully compensated for by a high score on another. This is a direct consequence of the summation operation. For trust scoring, this means a source with low Citation Integrity can still achieve a high composite score if its Content Freshness and Authority Vector magnitudes are sufficiently large. This contrasts with non-compensatory models like Trust Score Thresholding, where a single failing criterion can veto the entire assessment.

For a WSM to produce a mathematically valid result, the criteria must exhibit preferential independence. This means the contribution of one attribute to the overall score must not depend on the value of another attribute. In practice, this is often violated in trust systems. For example, the value of a high Credibility Index may be conditional on a low Hallucination Risk Assessment score. If the interaction is significant, a WSM will produce a distorted trust score, and a Bayesian Trust Network may be more appropriate.

Unlike black-box machine learning classifiers, a WSM is a fully deterministic and transparent algorithm. The final Trust Score can be decomposed and audited by inspecting each input signal's normalized value and its corresponding weight. This inherent Algorithmic Explainability is a primary reason the model is favored in regulated environments. A stakeholder can trace a specific score back to the formula: Score = w1*s1 + w2*s2 + ... + wn*sn, making the decision logic fully auditable for Trust Score Governance.

In a standard WSM, the importance weights assigned to each signal are fixed constants determined by domain experts or derived from historical analysis. They do not adapt in real-time. This static nature is a key limitation addressed by Dynamic Weighting systems. In a WSM-based trust pipeline, if a signal like Reputation Decay Function output becomes more volatile, the model cannot automatically increase its weight. Recalibration requires a manual Trust Calibration process to adjust the coefficients and redeploy the model.

A WSM collapses a multi-dimensional problem into a single, scalar composite value. While this simplifies decision-making, it results in significant information loss. Two entities with radically different signal profiles can receive identical final scores. A source with high authority but low accuracy can appear equivalent to one with moderate scores across all signals. This limitation motivates the use of an Authority Vector, which preserves the multi-dimensional profile, as a more granular input for downstream Signal Fusion processes.

The validity of a WSM is entirely dependent on the Trust Score Normalization preprocessing step. Because the model sums values directly, all input criteria must be transformed onto a common, dimensionless scale (typically 0 to 1). Summing a raw citation count (e.g., 1,500) with a binary verification status (0 or 1) would render the weight coefficients meaningless. Max-min normalization or Z-score standardization is mandatory to ensure that the weight vector accurately reflects the intended importance of each signal in the Signal Aggregation Layer.

Before signals can be summed, they must be transformed to a common scale. Min-max scaling maps values to [0,1], while Z-score normalization centers data around zero with unit variance. Without normalization, a signal ranging from 0–1000 would dominate one ranging from 0–1, rendering the weights meaningless. The choice of normalization technique directly impacts the interpretability of the final trust score.

The process of determining the importance coefficient for each criterion. Common methods include:

• Direct rating: Experts assign weights on a 0–100 scale
• Swing weighting: Weights reflect the range of each attribute
• Analytic Hierarchy Process (AHP): Pairwise comparisons derive consistent weights Poorly elicited weights are the primary source of error in weighted sum models.

A critical assumption of the weighted sum model: the contribution of each criterion to the total score is independent of all others. This means there are no interaction effects. If signal A and signal B together provide more trust evidence than their sum suggests, the model is misspecified. Violations require multiplicative or non-linear aggregation functions instead.

The systematic testing of how changes in input weights affect the final ranking or score. A robust trust model should not flip its classification of an entity as 'trusted' or 'untrusted' due to minor weight perturbations. Techniques include:

• One-at-a-time (OAT) variation
• Monte Carlo simulation across weight distributions
• Tornado diagrams to visualize impact

The weighted sum model assumes a linear relationship between each signal's value and its contribution to utility. A doubling of a signal's normalized value exactly doubles its weighted contribution. If the real-world relationship is non-linear—for example, diminishing returns after a certain threshold—a multi-attribute utility function (MAUT) with curved single-attribute utility functions should replace the simple weighted sum.

Weighted sum is a fully compensatory model: a very low score on one criterion can be offset by a very high score on another. In trust scoring, this means a source with excellent citation integrity might still pass a threshold despite poor content freshness. If certain signals are non-negotiable, a non-compensatory conjunctive rule (must-pass-all) or disjunctive rule (must-pass-any) should be applied before summation.