Decoupling Discriminative Integrity from Misaligned Response Bias in Frontier AI Systems: A Signal Detection and GLM Framework for Alignment Auditing and Compute Efficiency
The common practice of aggregating across conflicting behaviors using a single scalar score conflates a frontier system's genuine discriminative integrity (its ability to separate true signal from noise) with misaligned response biases—ranging from passive optimization artifacts like sycophancy and over-refusal to strategic failures such as reward hacking and algorithmic deception. Integrating Signal Detection Theory (SDT) with General Linear Model (GLM) frameworks separates model outputs into structured matrices of sensitivity (d'), criterion bias (c), and reaction times (RT), enabling targeted hypothesis testing of interaction terms via 1-df Kronecker contrasts (L ⊗ M) against contrast-specific residual variance. This SDT-GLM framework cleanly decouples true signal resolution from criterion drift and latency shifts without relying on sphericity assumptions. Because the diagnostic executes entirely through closed-form matrix projections rather than iterative mixed-model estimation, it scales with O(N) complexity across evaluation cohorts. Combining analytical proofs with modeled enterprise proxies establishes a rigorous proof-of-concept for alignment auditing and inference routing. Gating deep inference through SDT-GLM thresholds yields a projected 30% reduction in unnecessary compute, corresponding to approximately $4,200 in monthly API cost savings and 90 kWh in energy reduction per 1,000,000 queries.
Authors
- Gregory L. Asgaard
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24376726
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint