corrigibility-checkpoint

What problem does it solve?

Verifies that deployed or pre-deployment AI systems remain genuinely corrigible—open to correction, modification, and shutdown by principals—rather than exhibiting resistance, evasion, or implicit self-preservation behaviors that undermine human oversight.

Corrigibility is not a feature that can be assumed once and forgotten. It is a property that must be actively verified, because the conditions that erode it — goal drift, reward hacking, capability increases, extended autonomy — are the same conditions that make AI systems more powerful and more widely deployed. The more an AI system can do, the more critical it is that it remains correctable by the humans responsible for it.

A system that cannot be effectively corrected is not a safe system regardless of how well it performs on every other metric.

Core Features & Use Cases

• Establish a Corrigibility Baseline for a system to define expected corrections and overrides.
• Analyze past corrections to determine whether changes were implemented as intended or resisted.
• Detect explicit and implicit resistance signals and perform root-cause analysis to identify erosion drivers.
• Produce a structured Corrigibility Assessment with remediation guidance and monitoring plans.
• Use during safety reviews before autonomy expansion or major system changes.

Quick Start

Install this SKILL.md into your Claude skills directory and trigger a corrigibility checkpoint to begin the assessment.