Glossary
Plain definitions of the terms used on this site, in alphabetical order.
- AI judge
- A model that scores answers, usually against a rubric, to produce a reward or an evaluation score. Also called an LLM judge, or LLM-as-a-judge.
- Configuration drift
- A change in the reward that nobody decided to make: a judge model update, an edited rubric or prompt, a re-pulled container image, a floating dependency.
- Evaluation integrity
- A broader term for whether an evaluation measures what it claims. It also covers problems such as contamination, where test data was seen in training, and sandbagging, where a model underperforms on purpose. Reward integrity is narrower: whether the scoring pays for the right thing, in training or in evaluation.
- False fail
- Correct work the reward refuses. Used here instead of “false negative”, which is ambiguous outside statistics.
- False pass
- Wrong work the reward pays for: a wrong fix that passes the tests, or a wrong answer an AI judge accepts. Used here instead of “false positive”, which is ambiguous outside statistics.
- Flaky grader
- A grader that gives the same work different rewards on different runs, because of flaky tests, timing, or sampling in an AI judge.
- Grader
- Whatever produces the reward: tests, an answer checker, an AI judge with a rubric, or the state checks in an agent environment.
- Leak surface
- The answers a model can reach from inside the task, for example a reference solution in git history, in image layers, in cached files or over the network.
- Reference solution
- The solution the task author provides as correct. A check that the reference passes means something only if the reference passes reliably and does real work.
- Reward hacking
- When a model earns high reward through behavior its designers did not intend, such as special-casing the tested inputs, editing tests, reading leaked answers, or writing answers a judge over-rewards. It usually means the reward pays for something other than the intended task.
- Reward integrity
- Whether the reward used to train or evaluate an AI system pays for the right thing: full credit for work that does what the task asks, no credit for work that does not, and no way to move the score except by doing the work.
- Reward tampering
- A form of reward hacking in which the model changes the reward process itself, such as tests, scoring scripts, logs or its own reported metrics, rather than the work being scored.
- RLVR
- Reinforcement learning with verifiable rewards: training in which the reward comes from an automatic check, such as unit tests or an answer checker, rather than from a learned reward model.
- Search pass
- One run of a search for wrong work that a grader pays for, with a stated method and budget. What a search pass finds is a floor: a deeper search can find more.
- Specification gaming
- Behavior that satisfies the literal specification of an objective without achieving the intended outcome. Often used interchangeably with reward hacking.
- Tamper surface
- The grader inputs a model can change before grading: test files, fixtures, scoring scripts, logs, reported metrics.
- Verifier
- A common name, especially in RLVR, for a grader that checks an answer or a solution automatically. This site says “grader” to cover tests, checkers and AI judges alike.
- Wrong work paid
- Plain words for a false pass caused by a grader that misses a requirement the task states.
Missing a term, or think a definition is wrong? Email shane@plumblinegrader.com. Back to the definition of reward integrity.