Image
Workflow
Operations

Eval harness runs on every model change

Teams change prompts and models often and check the results by trying a few examples. An agent runs the whole evaluation suite on every change, compares it with the last approved run and writes the report. Whether the change ships is a person’s call.

The short answer

Short answer

An agent runs your evaluation suite on every change by watching for a new model version, prompt, retrieval setting or tool definition, running the full set of evaluations you defined, and comparing the results with the last approved run.

It reports what improved, what regressed and which cases failed. A named person decides whether the change ships.

Today

Why do model changes ship after a few spot checks?

Someone tweaks a prompt to fix one bad answer, tries three examples, and ships it. Next week a model upgrade arrives, and the same thing happens. Nobody reruns the cases that worked before, so a fix in one place quietly breaks another, and the first person to notice is a customer. The evaluation set exists, but running it is slow and manual, so it runs before big launches and not much else.

Who is this for?

An ML engineer, AI product lead or engineering manager whose product depends on a model, prompts or retrieval settings that change often. It suits teams with at least a starting evaluation set: test inputs with expected outputs, or criteria for judging them. Deciding what to measure, and what is good enough to ship, stays with the team.

What changes when the full suite runs every time?

Part of the releaseSpot-checked by handEvaluated on every change
What is testedA few examples someone remembersThe whole evaluation suite you defined
WhenBefore big launchesEvery model, prompt or setting change
ComparisonAgainst memoryAgainst the last approved run, case by case
RegressionsFound by customersListed in the report before release
The ship decisionThe person who made the changeA named approver, with the report in hand

How it works, step by step

The agent runs and compares. The team defines the measures and decides what ships.

01

Define the suite with the team

The team writes the evaluation cases and how each is scored: exact checks, rule checks, or a rubric a reviewer applies. The NIST AI RMF Playbook observes that the development and utility of trustworthy AI systems depends on reliable measurements and evaluations of underlying technologies and their use, and the suite is where that starts.
02

Watch for changes

The agent watches the places changes happen: model version settings, prompt files, retrieval configuration and tool definitions. A change starts a run.
03

Run the full suite

It runs every case against the changed configuration, in a test environment, never against live users.
04

Compare with the last approved run

It compares each case with the last approved baseline and lists improvements, regressions and new failures, with the inputs and outputs side by side.
05

Send rubric cases to reviewers

Cases scored by a rubric go to the reviewers you name, who score them. The agent does not score rubric cases itself.
06

Human gate: a named approver decides whether the change ships

The report goes to the approver for that product. They decide whether the change ships, needs work, or ships with a known regression they accept in writing. The agent deploys nothing and moves no baseline on its own.
07

Record the decision and the new baseline

When a change is approved, its run becomes the new baseline, and the decision is recorded with the report.

Never with a model change

What does the agent never do with a model change?

Ship it

Deployment follows the approver's decision, through your normal release.

Score judgment cases. Rubric cases go to people.

Move the baseline

Only an approved run becomes the new baseline.

Test on live users

Runs happen in a test environment.

Pick the metrics. The team decides what is measured. Google's Rules of Machine Learning put it early in the list: first, design and implement metrics.

Where this fits

Workflow / Automate a Workflow

One change taken from detection to an approval decision is a single process: Tier 1, Automate a Workflow. The readiness checklist will look at whether your evaluation set is written down well enough to run. The same practice of comparing new output with an approved baseline before trusting it runs through shadow mode, and the approval follows approval gates. More in the use case library.

How do model evaluations differ from unit tests?

QuestionUnit testsModel evaluations
What is checkedCode behaviourModel output quality
Pass or failUsually exactOften graded, sometimes by people
What triggers a runA code changeA model, prompt, setting or data change
Who decides on a regressionThe build failsA named approver decides
Tier 1 / Prove
How It Works

Questions, answered

It can mark the run as failing your thresholds, and your release process can treat that as a hold. Whether to ship anyway is the approver's decision, recorded with a reason.
AI Opportunity Audit · Model evals

Run the whole suite on every change

Bring the last model or prompt change you shipped and the cases you checked before shipping it.