---
title: "Which pull requests should a code review trial include? | Proofline"
description: "A review trial measures the pull requests you put into it."
url: https://proofline.sh/learn/review-trial-change-mix/
---

[Learning center](/learn/)

# Which pull requests should a code review trial include?

A review trial measures the pull requests you put into it. Choose them by a rule fixed before the trial starts, and report results for each kind of change, so the outcome represents your team's work rather than the changes that made a reviewer look good.

Code review The Proofline team November 20, 2026 3 min read

## See how the sample can mislead

Suppose a team tries a reviewer on 40 pull requests over two weeks. Fifteen are dependency updates opened by a bot, ten change only documentation or tests, ten are features and five touch the database or authorization. The reviewer posts one comment on the dependency updates, several on the documentation and test changes and most of its findings on the last fifteen.

A single summary of that trial says the reviewer was quiet on most pull requests and useful on some. The decision the team actually faces is narrower: is the reviewer useful on the fifteen changes where a mistake would hurt? The dependency updates contributed nothing to that answer and made up more than a third of the sample.

The opposite distortion happens when engineers choose which pull requests to send. People tend to send the changes they are unsure about, or the ones they expect a reviewer to handle well. Either habit changes the result before the reviewer has read a line.

## Fix the selection rule before the trial

Pick one rule and write it down with the trial's start date:

-   **Every pull request** opened against the chosen repository's default branch during the trial window. This is the simplest rule and usually the best one for a single active repository.
-   **A fixed fraction**, when every pull request would be too much to judge. Select by something nobody controls, such as the last hexadecimal digit of the pull request's first head commit: `0` to `7` enters the trial, `8` to `f` does not. That keeps roughly half and cannot be steered by the author.
-   **Every pull request in named paths**, when the decision is about one area, such as `migrations/**` or `src/auth/**`. State that the result applies to those paths only.

Avoid rules that depend on a person's judgment of the change, such as "pull requests that seem risky". They are useful for scoping a reviewer later, but they bias the measurement.

## Label each change before reading the review

Give each pull request in the sample one kind, decided from the change itself and not from the review. A short list is enough:

| Kind | Example |
| --- | --- |
| Dependency or generated | Lockfile update, regenerated client |
| Configuration or infrastructure | CI workflow, Terraform, deployment settings |
| Feature or behavior change | New endpoint, changed validation |
| Refactor | Moved code with no intended behavior change |
| Data or access | Migration, query change, permission check |
| Tests or documentation only | New tests, README edits |

Record size as well, as changed lines from the pull request's own counts. A reviewer may do well on small changes and poorly on large ones, and the mix of sizes varies from week to week.

## Report by kind, then decide

Report each kind on its own line: pull requests, findings, accepted findings, dismissed findings and misses you know about. The [denominator guide](/learn/review-usefulness-denominator/) explains which counts to divide.

| Kind | Pull requests | Accepted findings | Dismissed findings |
| --- | --- | --- | --- |
| Dependency or generated | 15 | 0 | 1 |
| Tests or documentation only | 10 | 1 | 6 |
| Feature or behavior change | 10 | 5 | 4 |
| Data or access | 5 | 3 | 1 |

These fictional figures support a different decision from their total. The reviewer cost six dismissals on documentation changes and earned one accepted finding there, while on data and access changes it earned three accepted findings at the cost of one dismissal. Five pull requests are far too few to estimate a rate for that kind, as the [sample size guide](/learn/review-trial-sample-size/) shows, but they are enough to decide where the next two weeks of the trial should concentrate.

If one kind of change has almost no examples in the window, say so. A trial that saw no migrations has no evidence about migrations, whatever the reviewer's results elsewhere.

## One month of free code review.

New organizations get one calendar month of hosted code review for private and public GitHub repositories. No card or cloud connection required. [Eligibility and fair-use terms](/free-code-review/).

[Start free with GitHub](https://proofline.sh/orgs/plan?offer=free_month)

## Related reading

-   ### [How many findings does a code review trial need?](/learn/review-trial-sample-size/)
    
    Code review The Proofline team November 20, 2026 4 min read
    
-   ### [Keep, narrow or remove: deciding at the end of a code review trial](/learn/end-review-trial-decision/)
    
    Code review The Proofline team November 20, 2026 3 min read
