Neural networks increasingly play a crucial role in high-stakes domains such as medical diagnosis, credit approval, and energy bidding. Auditing in these settings necessitates case-level evidence: which training cases support an action and what outcomes they produced. Case-Based Decision Theory (CBDT) formalizes this reasoning by aggregating outcome support from remembered cases.
We demonstrate that an OLS action readout fitted on a fixed neural representation permits an exact case-based decomposition. Each action score is a weighted sum of training-case returns, with coefficients determined by empirical Gram geometry. We identify a sufficient regime for CBDT similarity semantics; outside this regime, coefficients should generally be treated as signed Gram-geometric influences.
The decomposition yields audit signals that trace scores back to training cases, measure action coherence, and identify weak support. Across synthetic CBDT, PJM, Adult Income, and Default Credit tasks, the method recovers case-level preference structures and achieves the highest mean Top-30 consistency among compared attribution baselines, while remaining competitive on support reconstruction.
The audit requires only fitting an OLS top-layer probe, without retraining the representation or accessing the original optimization trajectory; probe fidelity is measured by score reconstruction.
Blogger's Review: This paper provides a clear auditing pathway for neural network decisions via Case-Based Decision Theory, which is particularly significant in high-stakes applications. Its approach enhances decision transparency and achieves effective support reconstruction through a simple OLS probe, making it a valuable contribution for real-world implementations.