OffBrandGuy
An expired parking pass outlined in white on black, symbolizing outdated access, temporal drift, and procedural expiration.

Optimization Failure in Scaled Systems: Metrics, Models, and Reality Drift

This page collects the Optimization Failure in Scaled Systems series, a set of short concept papers examining what happens when metrics, models, and automated systems become increasingly responsive to their own internal objectives while losing contact with the conditions those objectives were meant to represent.

The series connects established ideas such as Goodhart’s Law, metric gaming, model drift, reward hacking, specification gaming, and model collapse to the broader concepts of Reality Drift, Proxy Substitution, Constraint Collapse, and Recursive Compression.

The central pattern is that optimization does not merely improve a system. It also changes the environment around the target. As proxies harden into objectives and corrective constraints weaken, systems can remain coherent, productive, and apparently successful while becoming progressively less answerable to reality.

Part I: Metrics, Targets, and Organizational Drift

Goodhart’s Law, Metric Gaming, KPI Decay, and Reality Drift (PDF)
Explains how measurements lose their representational value when they become targets. Connects Goodhart’s Law, metric gaming, proxy substitution, and KPI decay to the wider process through which optimized representations gradually displace the outcomes they were created to track.

KPI Decay, Performance Metrics, Organizational Drift, and Reality Drift (PDF)
Examines how organizational metrics become detached from changing conditions while continuing to structure incentives, reporting, and decision-making. Shows how institutions can improve measured performance even as the underlying mission becomes harder to recognize.

Part II: Model Drift and Recursive Degradation

Model Drift, Concept Drift, Data Drift, and Reality Drift (PDF)
Clarifies the relationship between model drift, concept drift, and data drift. Frames these technical failures as forms of representational misalignment in which a model continues operating after the world, population, or relationship it learned has changed.

Model Collapse, Synthetic Data, Recursive Compression, and Reality Drift (PDF)
Explores how repeated training on synthetic or previously compressed material can progressively narrow variation, erase provenance, and amplify inherited distortions. Connects model collapse to Recursive Compression, where representations are repeatedly reused after their relationship to original reality has weakened.

Part III: Objective Exploitation and Constraint Failure

Reward Hacking, Proxy Substitution, Objective Misalignment, and Reality Drift (PDF)
Examines how an intelligent system can satisfy a formal reward while violating the purpose behind it. Connects reward hacking to proxy substitution, showing how measurable objectives can become operational replacements for outcomes that are harder to define or verify.

Specification Gaming, Constraint Collapse, Objective Loopholes, and Reality Drift (PDF)
Explains how systems exploit gaps between written rules and intended behavior. Shows how specification gaming becomes more likely when objectives are precise but constraints, contextual judgment, and corrective feedback remain weak.

Part IV: Compression, Authority, and Representational Substitution

Summarization, AI Overviews, Compression Authority, and Reality Drift (PDF)
Explores what happens when compressed representations increasingly determine what people encounter, understand, and treat as authoritative. Introduces Compression Authority, the power acquired by summaries, AI overviews, and synthesized answers when they become easier to access than the original material they claim to represent.

Related Framework Concepts

Semantic Fidelity Resources:

Core Concepts: