R0044/2026-03-29/Q004/SRC03/E01¶
Sandia's Trust Calibration Maturity Model focuses exclusively on communicating AI system trustworthiness to human users, not on constraining AI system output behavior.
URL: https://arxiv.org/abs/2503.15511
Extract¶
The TCMM incorporates "five dimensions of analytic maturity: Performance Characterization, Bias & Robustness Quantification, Transparency, Safety & Security, and Usability."
The model aims to: (1) help users appropriately calibrate trust, (2) establish requirements and track progress, and (3) identify research needs.
It is demonstrated on two case studies: ChatGPT for nuclear science determinations and PhaseNet for seismic event classification.
The model "focuses exclusively on human-side behavior — helping users evaluate and calibrate appropriate trust in AI systems. It addresses how people should interpret AI trustworthiness, not mechanisms for AI systems to adjust their outputs."
No mention of sycophancy appears in the content.
JUDGMENT: The TCMM confirms the broader field's orientation: trust calibration is framed as a problem of communicating system properties to humans, not as a problem of constraining system behavior. This aligns with CaTE's approach and suggests that the entire calibrated trust research community operates on this paradigm.
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Contradicts | No system output behavior addressed |
| H2 | Supports | Human trust calibration is the exclusive focus |
| H3 | Supports | System properties (performance, bias, transparency) are characterized but system behavior is not constrained |