In plain English
This page explains how AI systems can change over time through updates, tests, retraining, memory, and approvals even when no single model rewrites itself.
- Why this matters: AI risk can come from the whole arrangement, not one obvious model.
- What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
- Technical version below: the expert terminology remains available and is linked through the glossary.
Endogenous Yardstick Drift
Endogenous yardstick driftAssurance decay caused when the system or adjacent automation changes the measurements, thresholds, tests, or evaluator assumptions used to judge success. Open glossary definition occurs when the system that is being optimized can influence the ruler by which future versions are measured. The phrase is report-derived terminology used here as an analytical label.
How drift appears
Drift can occur when evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition prompts change, benchmark weights move, hidden tests leak, model judges share training data with candidates, release pressure weakens thresholds, or success metrics overfit to a shortcut. It does not require malicious intent. It only requires a feedback loop where measurement changes in the direction favored by optimization.
Why it is dangerous
A stable model with a stable test can be compared over time. A changing model evaluated by a changing evaluator produces a moving baseline. Apparent improvement may be real, or it may be the ruler changing.
Control pattern
Keep frozen reference suites for regression comparison alongside live adaptive tests. Preserve evaluator versions, hidden-test custody, deterministic validators, negative controls, independent evidence stores, and human-owned hard constraints. The evaluator must have its own rollbackReturning a system to an earlier known state. Open glossary definition path.
Reference control
Use frozen reference suites alongside evolving evaluators so improvement claims can be compared against a stable ruler.