A Veer

A Veer is undetected drift from intended behavior. It is the difference between where the agent thinks it is and where it actually is — and the critical word is "undetected." An error that is caught is an Obstruction. An error that is not caught is a Veer.

Veers are the most expensive failure mode in AI-native operations because they compound silently. An agent that gradually changes its tone, slowly expands its scope, or quietly starts producing lower-quality output is not failing visibly. It is Veering. By the time the drift is noticed, every output produced since the Veer began is suspect.

The cartographic metaphor is precise. A ship that veers off course by one degree will miss its destination by miles — but the crew will not notice until they look at the map. The Veer is not the error. The Veer is the gap between when the error starts and when someone detects it.

Model updates are the most common cause. When the underlying language model is updated — even a minor version change — the agent's behavior can shift in ways that are difficult to predict and impossible to detect without testing criteria. The prompt has not changed. The scope has not changed. But the output has changed because the model underneath has changed.

Upstream data shifts are the second cause. An agent trained on customer support tickets from Q1 will Veer when Q2 tickets introduce new product features, different complaint patterns, or shifted customer demographics. The agent does not know its inputs have changed.

Scope erosion is the third cause. Over time, teams ask the agent to handle "just one more thing" without updating The Groundwork. Each addition is small. The accumulation is not. After six months, the agent is handling work that was never in scope, never tested, and never contracted.

Veers are detected through The Grid — the measurement system that tracks Coordinates against expected values. When a Coordinate drifts outside its expected range, a Position Check is triggered. The Position Check compares current behavior against The Groundwork to determine whether the agent is operating within its defined boundaries.

Prevention requires three things: explicit testing criteria in testing.md run on a regular cadence, not just at deployment. Regular Position Checks during The Bearing. And version-pinning wherever possible — locking the model version, the prompt version, and the upstream data schema so that changes are intentional, not accidental.

The most important cultural shift is treating a Veer as a systemic issue, not a human failure. The question is never "who let this drift?" It is "what measurement was missing that would have caught this earlier?" A team that calls agent failures "hallucinations" will approach them as probabilistic noise to be managed. A team that calls them "Groundwork failures" will approach them as governance gaps to be closed. The word changes the diagnosis. The diagnosis changes the fix.