realitydriftproject/goodharts-law-and-proxy-optimization

Mechanism Notes — Reality Drift Framework

This section documents specific mechanisms through which systems drift away from the realities they are meant to represent.

Each paper isolates a recurring pattern observed across AI systems, organizations, and decision environments.

Included Papers

• How AI Systems Exploit Objectives
  Reward hacking and specification gaming as failures of proxy alignment.

• How Algorithms Amplify the Wrong Signals
  Signal distortion in digital platforms and feedback-driven environments.

• When Metrics Become Targets
  Goodhart’s Law and the shift from measurement to optimization.

• Why Institutions Drift from Their Mission
  Goal displacement and bureaucratic drift in organizations.

• Why Systems Optimize Metrics Over Real Outcomes
  Proxy optimization as a general failure mode across systems.

Core Pattern

Across all examples, a consistent structure appears:

1. A measurable signal is introduced
2. The system begins optimizing toward that signal
3. The signal becomes easier to satisfy than the underlying reality
4. The system improves the signal while drifting from what it represents

Relation to the Framework

These mechanisms are expressions of:

• Reality Drift
• Optimization Trap
• Proxy Optimization

They also inform the design of:

• Drift Diagnostics (detecting misalignment)
• Cognitive Workflows (reducing reasoning drift)
• Reality-Constrained Systems (preventing drift structurally)

Usage

These notes can be used to:

• identify failure modes in AI and decision systems
• analyze where metrics no longer reflect reality
• design systems that maintain alignment under optimization pressure

Core framework and sources

• Substack (articles)
• GitHub (full library)
• DOI (research paper)
• Glossary & Definition

Each document includes both a Markdown version and a PDF version for distribution and indexing.