Agency & Decisions

Eval Run Setup

Solves: choosing what to test and knowing what it will cost before pressing run.

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

watching emerging established contested fading 3 sightings since Jun 2026
Ask an assistant about this pattern Ask ChatGPT ↗Ask Claude ↗

Eval Run Setup is the moment before a test runs. The person picks the prompt versions, the models and the example set to test them against. The tool says what the run will cost and how long it will take before the person starts it. The run then carries a record of what was tested.

When to Use It

  • A run costs real money or real minutes
  • More than one version or model is in play
  • The team needs a record of what was tested

When Not To

  • Runs are free and instant; just run them
  • There is one version and one model; the choice is already made
  • The estimate would be a guess with no basis

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 dataset · models · price list picks the versions, the models and the dataset estimates cost from dataset size and model prices returns the estimate shows the estimate before the run starts the run runs; records the configuration with the results
02 · who holds what
human interface agent Prompt versions: The versions to test, at least a baseline and a candidate 01 Prompt versions Models: The models each version runs on 02 Models Example set: The eval set the run uses, at a named version 03 Example set Estimate: Cost and time for the run, shown before the person commits 04 Estimate Run record: What was tested, when, by whom, at what cost 05 Run record
human interface agent Prompt versions: The versions to test, at least a baseline and a candidate 01 Prompt versions Models: The models each version runs on 02 Models Example set: The eval set the run uses, at a named version 03 Example set Estimate: Cost and time for the run, shown before the person commits 04 Estimate Run record: What was tested, when, by whom, at what cost 05 Run record
  1. 01 Prompt versions The versions to test, at least a baseline and a candidate
  2. 02 Models The models each version runs on
  3. 03 Example set The eval set the run uses, at a named version
  4. 04 Estimate Cost and time for the run, shown before the person commits
  5. 05 Run record What was tested, when, by whom, at what cost
03 · the component

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

Field Notes

SEP 10 · Clipped · via Promptfoo on GitHub Release DecisionEval Run Setup

Promptfoo lets a pass percentage turn the release check red

A threshold is a decision made once, by whoever wrote the workflow file. The product does not ask whether this regression should block or only warn; that question needs a screen, not a setting.

JUN 3 · Clipped · via TheRouter (secondary) Release DecisionEval Run Setup

OpenAI retires its testing product and points builders elsewhere

A vendor retiring the surface moves the interaction into the team's own tools. That is an argument for patterns that outlive any one product, and for baselines the team owns.

Documented sightings (1) — anatomy-mapped observations from vendor materials
Promptfoo

Prompts times providers as a matrix; repeats for flaky judges are a setting. No cost or time estimate before the run.

Promptfoo on GitHub · Sep 10, 2026

Tensions & Failure Modes

  • Estimate honesty: a cost figure that is often wrong trains humans to ignore it. Show the range and the assumptions behind it.
  • Grid explosion: three versions times four models times two sets is a lot of runs. Defaults should make the smallest useful comparison the easy path.
  • Reruns: outputs vary run to run. One run of a small set can flip a verdict; the setup should say how many repeats it used.

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