Topic 5 of 20 · Part II : Grounding, Evaluation and Orchestration

Evaluation

Evaluation data determines what the organisation can see, and therefore what it can improve.

The control pattern

Data asset

Test datasets and scoring rules

Knobs

Scenario mix, Difficulty, Time split, Segment coverage, Scoring method, Pass thresholds

Outcome

Reliable measurement of system quality

Measurement

See how to measure this below

What it is

Evaluation data determines what the organisation can see, and therefore what it can improve. A test set is a control surface because teams naturally optimise toward the failures and metrics it exposes.

Its knobs include scenario mix, difficulty, coverage of rare conditions, scoring method, reviewer rubric, time window and release threshold. A single average score is rarely sufficient: it can hide severe weaknesses in certain sectors, document types, market regimes or user groups. Evaluation should be sliced by the properties that matter operationally.

Evaluation becomes a true control knob when a model, prompt, retriever or scoring change cannot ship unless it improves the intended metric without crossing a guardrail elsewhere.

The knobs in detail

Each row is one adjustable property of the data asset, and what moving it tends to do.

KnobWhat you adjustLikely effect
Scenario mixWhich tasks and cases are representedDetermines what failures are visible at all
DifficultyShare of hard and rare conditionsPrevents easy averages from hiding weakness
Time splitPoint-in-time boundaries on evidenceStops future information leaking into tests
Segment coverageSectors, document types, user groupsExposes concentrated failure
Scoring methodMetrics and reviewer rubricDefines what counts as correct
Pass thresholdsBar for release, by risk levelTurns measurement into a gate

Applied: Stocks Assistant

Stocks Assistant should maintain separate evaluation suites for factual extraction, earnings summarisation, sentiment interpretation, ranking, portfolio explanation, options analysis and conversational question answering. Financial evaluations must be point-in-time: a model answering a historical question may use only information that was available at that time, or revised filings, later analyst estimates and future prices leak into the test. Ranking evaluations should include transaction assumptions and survivorship-bias controls.

How to measure it

Evidence that the knob produced the intended behaviour, rather than shifting the problem elsewhere.

  • Numeric accuracy and citation precision
  • Unsupported claims and completeness
  • Confidence calibration
  • Latency and cost
  • A frozen core benchmark for comparability, plus a rolling set of new production failures

Common mistakes

Reporting one average score. It hides exactly the weaknesses that matter operationally.

Applying the same pass threshold to high-risk and low-risk tasks. One threshold for everything protects nothing.