The 70% problem is a shorthand, not a measurement: AI handles familiar scaffolding and boilerplate, while the remaining work demands context-sensitive judgment. The percentage may rise with better models, but the unresolved remainder remains qualitatively different and often higher-stakes. section 1.6
The easy portion can create an illusion of completion, while production risk concentrates in the boundaries:
- Edge-case cartography: generated code follows the happy path more reliably than rare inputs, error conditions, and undescribed states. section 1.6
- Architectural integrity: locally correct code can violate project conventions, drift from established patterns, and fail at integration seams. section 1.6
When the imagined case is not the real case
The Gemini CLI failure worked for every operation its author imagined but remained undefined for the operation that actually occurred. The NX compromise similarly entered through a seam where locally reasonable code became catastrophic in context. section 1.6
