Glossary — AI Evaluation

Covers Lesson 1 (the rest of this track is in progress). Grouped by where each term first lands. Shared AI terms are reused verbatim from the cross-track canon so vocabulary never drifts.

Shared canon (carried in from ../ai-agents)

• LLM (large language model) — the reasoning model (the brain in a jar); predicts tokens, has no memory/tools/actions on its own.
• Tool — a callable capability/API/function the model can invoke.
• Agent — an LLM using tools in a loop to accomplish a goal.
• Agent Loop — observe → think → act → repeat.
• Runtime — the orchestration layer around the model (loop, state, tool execution, retries, limits).
• RAG (Retrieval-Augmented Generation) — retrieve → assemble → LLM → answer.

Foundations (L1)

• Golden dataset — a versioned, curated set of (input, expected-output / acceptance-criteria) cases used as the ground truth for offline evaluation. Small, hand-checked, and treated like test fixtures: it changes only by deliberate review, never silently.
• Offline evaluation — scoring the system against a fixed dataset (the golden set) before shipping. Reproducible, runs in CI, gates a deploy. Like a unit/integration test suite for AI behaviour: same inputs every run.
• Online evaluation — measuring quality on live production traffic. Two shapes:
  • A/B — split real users between the old and new system and compare outcome metrics. Users see the variant; you measure who does better.
  • Shadow — run the new system on real traffic in parallel but throw its output away (never shown to users); compare against the live system offline. No user risk, no live signal either.
• Eval harness — the runner that loads a dataset, executes the system per case, applies a grader, and aggregates scores into a report. The test framework for AI: dataset + runner + grader + report.

Datasets & benchmarks (L2)

• Benchmark — a shared, public dataset + metric used to compare models/systems against each other (e.g. MT-Bench, HELM scenarios). A standardized exam; good for cross-model comparison, weak as a proxy for your task.
• Benchmark contamination — the benchmark’s test data leaked into a model’s training data, so a high score reflects memorization, not capability. Why a fresh, private golden set beats a public leaderboard for trusting a result.

Human eval & judges (L3)

• LLM-as-a-judge — using an LLM to score or compare outputs (faithfulness, relevance, preference) in place of a human rater, so evaluation scales. Cheap and fast, but biased (position, verbosity, self-preference) and must be calibrated against human labels before you trust it.

Task & tool success (L4)

• Task success rate — fraction of cases where the system achieved the user’s actual goal (end-to-end, outcome-based), not whether individual steps looked plausible. The headline metric for an agent: did it get the job done.
• Tool success rate — for agents, the fraction of tool calls that were the right tool, with valid arguments, that succeeded. Decomposes a task failure into “wrong plan” vs “right plan, broken execution.”

Retrieval eval & hallucination (L5)

• Hallucination rate — fraction of outputs containing a claim not supported by the provided context or ground truth (unfaithful / fabricated). Measured against faithfulness/groundedness; the inverse of “every claim is grounded.” (Retrieval-specific eval lives in ../context-engineering L6.)

Regression & continuous eval (L6)

• Regression evaluation — re-running the golden set on every change (prompt, model, retrieval, tool) to catch quality that silently got worse. The AI equivalent of a regression test suite: a green eval gates the merge.

(L7 — production evaluation pipelines — composes the above: golden-set regression in CI + shadow/ A/B online + judge scoring + dashboards/alerts. No new primitive terms.)