Recipe 23: Pairwise answer evaluation #23

Recipe 23 of 60 · Level 2 (Easy) · Evaluation & optimization · Choice

Use case

Compare two candidate answers against one explicit rubric criterion while retaining tie and insufficient-evidence outcomes for evaluation against human labels.

What to build

The folder recipes/23-pairwise-answer-evaluation/ with notebook.ipynb, README.md, and fixtures/, meeting the recipe contract. Start from python tools/new_recipe.py 23 and the template in recipes/_template/.

Build notes

• One explicit rubric criterion per comparison. Outcomes: a, b, tie, insufficient_evidence.
• Python randomises which answer is shown first and checks that swapping the order gives a consistent verdict.
• Fixtures carry human labels.

Decision design

• Choice. Options are a fixed set supplied by Python, including the fallback outcome this use case names. Show the choice, the per-option probabilities, and the confidence for individual examples before aggregating.

Level 2 scope

Context-sensitive matching, scoring, or verification with candidate sets and explicit uncertain or no-match outcomes.

Acceptance criteria

• recipes/23-pairwise-answer-evaluation/notebook.ipynb runs top to bottom offline, with no key and no network, and its outputs are committed
• Live mode is opt-in (JEV_COOKBOOK_LIVE=1 with TYPESAFE_API_KEY) and uses the same question definitions as the offline run
• Fixtures are synthetic, include the hard cases named in the build notes, carry gold labels, and pass the fixture validator
• Every response fixture records its provenance, and the notebook states which mode it ran in
• Evaluation: Agreement with human labels (raw and Cohen's kappa) and the order-consistency rate.
• Candidate sets are assembled by Python and shown to the reader
• The uncertain or no-match outcome is exercised by several fixtures and reported separately in the evaluation
• No statement about Jev's quality, latency, or cost appears without recorded live inference; in a synthetic run, metrics are labelled as pipeline checks
• recipes/23-pairwise-answer-evaluation/README.md says what the recipe teaches, how to run it, and what was and was not measured live
• The pull request changes only recipes/23-pairwise-answer-evaluation/ and CI is green

Dependencies

• Blocked by Foundation: CI executes every notebook offline and validates fixtures #69 (CI executes every notebook offline and validates fixtures)
• Tracked in Epic: Level 2 · Easy recipes #76

Sources

Read these before writing the questions.

• S02 · TypeSafe AI: Primitives (Questions): Choice, Noul, Score, independent questions, and dependent follow-up calls.
• S03 · TypeSafe AI: Confidence: Distribution-derived confidence and task-specific thresholds informed by observed outcomes.

Out of scope

• Changes to shared code, the catalog, or the root README. If the shared package is missing something, open a foundation issue and link it here.
• Real side effects of any kind. Actions are simulated.