---
title: "Evaluation"
description: "How agents are performing across every review: acceptance, coverage, and Merger's track record."
canonical: "https://baz.ai/docs/insights/evaluation"
---

# Evaluation

How agents are performing across every review: acceptance, coverage, and Merger's track record.

**Evaluation** aggregates every review Baz has run and shows how the agents are actually performing. It sits under **Insights → Evaluation** and splits into the same three stages as the [Software Factory](/docs/agents/baz-agents#the-software-factory):

| Tab | What it shows |
|---|---|
| **Build** | Acceptance, coverage, and trends for the code review agents, plus Fixer stats |
| **Deploy** | Merger Agent Stats and Verification Agent stats, with an author filter |
| **Maintain** | SRE Agent, Full Repo Security Scan, and Skill Maintainer stats |

Each tab shows stats only for agents that are on. **View Evaluation** on a stage page opens the matching tab.

## Build

### Overview

The top of the tab is a snapshot of review activity over the selected window.

| Metric               | Description                                                          |
| --------------------- | -------------------------------------------------------------------- |
| PRs Reviewed          | How many changes Baz analyzed                                        |
| Accepted Suggestions  | How many suggestions were accepted into code                         |
| Rejected Rate         | The share of suggestions explicitly rejected by reviewers            |
| Fixed via Fixer       | How many findings were resolved by a Fixer commit, where Fixer is enabled |

Below the stats, **Accepted Suggestions Per Agent Type** breaks acceptance down by agent category, such as security, code quality, and code correctness, so it is clear which classes of agent are driving accepted feedback. One accepted suggestion can be associated with more than one agent type. Where Fixer is enabled, the same breakdown can be switched to show suggestions Fixer fixed instead of suggestions developers accepted.

### Coverage

**Coverage** answers a different question from acceptance: not how much of what Baz said was taken, but how much of what was there Baz found. It shows the total number of comments alongside a per-category count, with a radar chart of the same categories, so gaps in coverage are visible next to the categories Baz reports heavily on. Where other AI reviewers are active on the same repositories, their real-bug recall for the window appears beside Baz's as a set of vendor badges.

### Acceptance rate

The acceptance rate section carries a trend chart and a per-agent table, filtered together.

| Column          | Description                                                                    |
| ---------------- | ------------------------------------------------------------------------------ |
| Agent            | The agent that produced the comments                                           |
| Accepted         | Comments counted as accepted. See [Definitions](#definitions)                  |
| Unaddressed      | Comments with no or neutral feedback that were not fixed or marked accepted    |
| Acceptance Rate  | Accepted comments as a share of all comments                                   |
| Rejection Rate   | Rejected comments as a share of all comments                                   |
| Fixed via Fixer  | Comments resolved by a Fixer commit, where Fixer is enabled                    |

Three filters apply to this section: **Agents** and **User**, both multi-select, and **Time**, a single choice of 1 week, 2 weeks, 1 month, or 6 months, defaulting to 1 month. With no agent filter applied, per-vendor acceptance rates for other AI reviewers appear alongside Baz's, which is how a team compares Baz against another reviewer running on the same code.

Historical findings from the earlier individual reviewers are counted under the agent that now covers them: Logical Bugs, Breaking Changes, and Type Inconsistency under **Code Correctness**, and Code Dedup and Conventions, Conciseness, Naming and Typos, and Code Hygiene under **Code Quality**, with rates recalculated from the combined counts. Findings from the retired REST API Best Practices agent are not included.

Below the review metrics, the Build tab shows Fixer stats: fixes Baz made on pull requests.

### Definitions

| Term         | Definition |
| ------------ | ---------- |
| Accepted     | A comment that meets any of: a developer marked it accepted in Baz, a later commit fixed it (judged by an AI check of the new code), or the replies and reactions on its thread were judged positive |
| Rejected     | A comment whose replies and reactions were judged negative, and that was not fixed or marked accepted |
| Unaddressed  | Every other comment: no or neutral feedback, not fixed, not marked accepted |
| Real-bug recall | For each finding category, the accepted code comments on merged pull requests from one reviewer, divided by the accepted code comments from all reviewers on those pull requests. Human reviewers count as one group, and each other AI reviewer is identified by its app account. Comments are not deduplicated across reviewers |

Reactions such as 👍 count only through the judgment of the thread's replies and reactions. Resolving a thread on its own does not make a comment accepted.

## Deploy

The Deploy tab tracks the agents that decide whether a change is ready. **Merger Agent Stats** shows evaluated pull requests, the share marked ready to merge, the share that merged with no human follow-up, a breakdown of recommended pull requests by current status, and throughput over time. Verification Agent stats sit beside it. An **Author** filter narrows both to specific pull request authors. See [Merge Automation](/docs/capabilities/merge-automation) for what the Merger metrics mean.

## Maintain

The Maintain tab tracks the agents that work on shipped code: SRE Agent issues and pull requests opened and merged over time, Full Repo Security Scan findings and scans, and Skill Maintainer updates and its last scan.

## How to use it

Open **Insights → Evaluation** and start on Build. Narrow to the agents, users, and window you care about, compare acceptance across agent types to see where the value is concentrated, then read Coverage to find the categories nobody is catching. Track the trend over successive windows to see whether review quality is improving, and switch to Deploy to check whether Merger's calls are holding up.

## FAQ

<details>

<summary>What metrics does this dashboard provide?</summary>

Changes reviewed, accepted suggestions, rejected rate, fixes applied by Fixer, acceptance broken down by agent type and by agent, acceptance trend over time, coverage by finding category, and Merger's evaluation and merge outcomes.

</details>

<details>

<summary>How do I read "Accepted Suggestions Per Agent Type"?</summary>

It shows which classes of agent, for example correctness, conventions, or security, are driving the most accepted suggestions. A single accepted suggestion can count toward more than one agent type.

</details>

<details>

<summary>What is the difference between acceptance and coverage?</summary>

Acceptance measures what developers did with Baz's comments. Coverage measures how much of the real issue space was found in the first place, by category, which is where a blind spot shows up even when acceptance looks healthy.

</details>

<details>

<summary>Can I filter by user, time, or agent?</summary>

Yes. The acceptance rate section filters by Agents, User, and Time, and the time window also drives the overview stats and the trend chart.

</details>

<details>

<summary>How does Baz compare against another AI reviewer?</summary>

When other AI reviewers comment on the same repositories, their acceptance rate and real-bug recall for the window show up beside Baz's, in the acceptance rate section and in Coverage respectively. Applying an agent filter hides that comparison, since it no longer covers the same population.

</details>