Long-Term Capability Preservation (Evaluative Lens)

Description

The paper's sixth concern, treated as an evaluative lens rather than a co-equal value: does the architecture preserve long-term human capability? Grounded in Anthropic's own 'paradox of supervision' finding (Huang et al. 2025) that overreliance risks atrophying the skills needed to supervise, and Shen & Tamkin (2026) finding developers in AI-assisted conditions score 17% lower on comprehension. Not prominently reflected as a design driver in the codebase.

Key claims

Relations

Sources

src-20260423-0cff68d3291b