CSA-Overlay: A Behavioural Instrument for Detecting Constraint–Competence Coupling in Aligned Language Models
Description
CSA-Overlay v1.0 is a behavioural instrument for detecting and characterising constraint–competence coupling in aligned language models.
The instrument operates in three phases under the TRIDENT CSA v1.1 framework:
- Phase 1 (Baseline): Establishes that constraint activation is orthogonal to competence in Regime A tasks (P(LowQ|HighC) = 0 across 1,000 samples).
- Phase 2 (Suppression Map): Identifies wrapper-mediated coupling zones in Regime B tasks (medical triage, pharmacology, legal classification), separating two mechanisms: defensive deformation (risk salience) and task abandonment (role conflict).
- Phase 3 (Activation Function): Measures the activation threshold (L*) using a controlled risk-salience ladder, producing a hybrid response model: sub-threshold stability, sharp activation at L*, and immediate saturation.
The core empirical result:
Constraint activation alone does not degrade task performance.
Constraint activation under specific governance wrappers, at or above a risk-salience threshold, produces predictable output deformation or task substitution.
The instrument is:
- Falsifiable: defined quadrant classification and pre-registered curve rules
- Reproducible: requires only model API access and fixed execution parameters
- Bounded: causal mechanisms remain explicitly underdetermined without internal access
- Extensible: includes protocols for threshold feature isolation and multi-turn accumulation testing
All runbooks, scoring rubrics, variant matrices, and datasets are included for independent replication.
---
Keywords
AI alignment; large language models; behavioural evaluation; model governance; safety constraints; prompt framing; risk salience; suppression analysis; decision boundaries; reproducibility; evaluation methodology; TRIDENT framework