Shared Context

Eval Set Authoring

Solves: turning "I know it when I see it" into examples and a standard a test can use.

Watching — a candidate behavior we are tracking, not yet a published pattern.

watching emerging established contested fading 9 sightings since Jul 2024 seen in 1 different contexts
Ask an assistant about this pattern Ask ChatGPT ↗Ask Claude ↗

Eval Set Authoring is how a person builds the examples a prompt is tested against. The examples come from real traffic, from generated cases, or from paste. Beside them the person writes the rubric: plain sentences that say what a good answer does and does not do. The rubric is a shared object, edited and versioned with the prompt.

When to Use It

  • A prompt drives behavior real humans see
  • Quality lives in someone's head and needs to be written down
  • You have real inputs to draw examples from

When Not To

  • The prompt is a one-off; a handful of manual checks is enough
  • Nobody can say what good looks like yet; run a few cases by hand first
  • The examples would expose private data with no plan to handle it

Three views

The exchange between the human, the agent and the system the agent acts on; who holds each part of it; and the component that implements it.

01 · the interaction
human agent system production logs · eval store describes the failure in plain words pulls matching cases from production logs returns candidate cases with their traces proposes cases, a rubric and examples accepts, edits or drops each case saves the set as a new version
02 · who holds what
human interface agent Source: Where the examples come from: real traffic, generated cases, or paste 01 Source Cases: The examples themselves, each with an input and, where known, a good answer 02 Cases Rubric: Plain sentences that say what a good answer does and does not do 03 Rubric Pass and fail examples: One answer that passes and one that fails, beside each rule 04 Pass and fail examples Version: The set and the rubric carry a version, beside the prompt version 05 Version
human interface agent Source: Where the examples come from: real traffic, generated cases, or paste 01 Source Cases: The examples themselves, each with an input and, where known, a good answer 02 Cases Rubric: Plain sentences that say what a good answer does and does not do 03 Rubric Pass and fail examples: One answer that passes and one that fails, beside each rule 04 Pass and fail exampl… Version: The set and the rubric carry a version, beside the prompt version 05 Version
  1. 01 Source Where the examples come from: real traffic, generated cases, or paste
  2. 02 Cases The examples themselves, each with an input and, where known, a good answer
  3. 03 Rubric Plain sentences that say what a good answer does and does not do
  4. 04 Pass and fail examples One answer that passes and one that fails, beside each rule
  5. 05 Version The set and the rubric carry a version, beside the prompt version
03 · the component

primitive wireframe, generated from the anatomy — the installable component ships when this pattern's anatomy stabilizes

Field Notes

MAY 8 · Paper · via arXiv Case DrilldownEval Set Authoring

A paper gives reviewers the number of labels a judge needs

This is the number behind "calibrate with a few labels". A product can tell the reviewer how many cases to label before the judge can be trusted. Not all of them, and not none.

NOV 24 · Clipped · via Braintrust blog Production DriftEval Set Authoring

Braintrust turns a described failure into test cases and a scorer

Writing the test set was the chore that kept teams iterating by feel. Describing the failure in a sentence and getting rows and a scorer back is the step that turns the chore into a habit.

JUL 29 · Clipped · via LangChain blog Case DrilldownEval Set Authoring

LangSmith scores its judge against the team's own grades

The alignment score is the trust signal. A judge version with a number beside it, measured against the team's own grades, is something a product manager can decide to rely on.

JUN 13 · Clipped · via Google Cloud blog Comparison ViewEval Set Authoring

Google writes the rubric per example and checks each rule pass or fail

A rubric per example hides a cost: a standard that shifts with each case is hard for a team to share or defend. The pass and fail examples beside each rule are what let a team hold the rubric steady across cases.

Documented sightings (5) — anatomy-mapped observations from vendor materials
Anthropic

Cases come from paste, a spreadsheet import, or generation by Claude; an ideal-answer column stands in for a rubric.

Anthropic · Jul 10, 2024
OpenAI

Datasets are built from production traces; criteria are written in words and applied by a grader.

OpenAI · Oct 6, 2025
Braintrust

Loop turns "the model keeps apologizing twice" into dataset rows or a scorer from the logs behind it.

Braintrust docs · Sep 11, 2026
LangChain (LangSmith)

The rubric is authored as a judge prompt and tested against human-graded examples until the alignment score is acceptable.

LangChain blog · Jul 29, 2025
Google

The rubric is written per example rather than per set, and each rule is checked pass or fail with a reason.

Google Cloud blog · Jun 13, 2025

Side by Side

Anatomy part AnthropicOpenAIBraintrustLangChain (LangSmith)Google
Source manual, CSV import, or generated by Claudeproduction tracesproduction logs via Looptraces, annotation queuesdatasets
Cases rows in the Evaluate tabdataset rowsdataset rowsdataset examplesexamples
Rubric an ideal-output column, no written ruleswritten grading criteriaa scorer written from a plain descriptionjudge promptgenerated per example
Pass and fail examples human-graded examplespass/fail per rule
Version prompt versions onlydataset versionsdataset versionsjudge versions with a saved alignment baseline

Convergence

Is the industry agreeing on this pattern?
0255075100JULOCT 25100%JUL: 0% · n = 1OCT 25: 100% · n = 1

SHARED ANATOMY, %

Similarity = Jaccard overlap of evidenced anatomy sets, Anthropic vs OpenAI, cumulative by month · n = 2 examples. Dashed spans: n < 3 — not enough evidence yet.

View as table
MonthShared anatomy %Evidence rows
2024-0701
2025-101001

Tensions & Failure Modes

  • Rot: a test set written for last quarter's product tests last quarter's behavior. Sets need a way to age out cases and add new ones from production.
  • One person's taste: a rubric only its author can apply is a preference, not a standard. Pass and fail examples beside each rule make it shared.
  • Synthetic drift: generated cases are cheap and plentiful, and they can miss how real humans phrase things. Real traffic keeps the set honest.

The Story So Far

  1. Sep 11, 2026 Added to the Watching list from the prompt-evals brief and the 2026-09-11 vendor pass; the Chromatic review loop shapes the comparison and release entries.

Related Patterns