> ## Documentation Index
> Fetch the complete documentation index at: https://docs.anyreach.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Reading the dashboard
> Make sense of metric scores, filter by who graded, and drill into exceptions.
Once you've run some evaluations, the numbers live in the dashboard under the **Analytics** section, on the **Overview** tab. This page walks you through how to read what you see there — what each score means, how to filter by who did the grading, and how to drill into the conversations that failed.
Before we start, two quick definitions you'll need throughout:
* A **metric** is one thing you measure about a conversation (for example, "Task Resolution"). Each metric groups one or more **criteria** — individual yes/no checks.
* A **verdict** (or **outcome**) is the answer to one criterion for one conversation. Every verdict is one of four values: **true**, **false**, **abstain** (the grader couldn't decide), or **na** (the criterion didn't apply to that conversation).
## The per-metric cards
The Overview tab shows one **card per metric**. Each card is a scoreboard for how that metric is performing across all the conversations you've evaluated.
The headline number on a card is the **compliant rate**. To understand it, you first need to know what "compliant" means.
A criterion has an **expected value** — the answer that counts as good. Some criteria are phrased so that **true** is good ("Did the agent confirm the caller's identity?"), and some so that **false** is good ("Did the agent give wrong information?"). A verdict is **compliant** when its outcome matches that expected value. So a "false" verdict can absolutely be the good outcome — it depends on the criterion.
The card shows four counts that add up to the full picture:
| Count | What it means |
| ------------- | --------------------------------------------------------------------------------------------------------------- |
| **Compliant** | Verdicts where the outcome matched the expected value — the "good" answers. |
| **Answered** | Verdicts that came back as either **true** or **false** — a real decision was made. |
| **Abstain** | The grader read the conversation but couldn't decide. |
| **Na** | The criterion didn't apply to that conversation (for example, an audio-only check on a chat with no recording). |
The **compliant rate** is then:
```
compliant rate = compliant ÷ answered
```
The key thing to notice is the denominator: **answered**, not the total number of verdicts. **Abstain** and **na** verdicts are deliberately left out of the math. A criterion that didn't apply, or that the grader skipped, shouldn't drag a score up or down — only real decisions count.
The compliant, answered, abstain, and na totals on each card are exact. They come straight from the server across **all** matching verdicts — they are not capped or sampled, so the rate you read is the true rate for the filters you've applied.
### How to read a metric card — a worked example
Say the **Information Accuracy** card shows these numbers over the last 30 days:
```
Information Accuracy
compliant rate: 94%
compliant 47 · answered 50 · abstain 2 · na 8
```
Read it like this:
* There were **60** verdicts in total (47 + 3 + 2 + 8 — the 3 is the non-compliant share of the answered group).
* **8** were **na** — the criterion didn't apply to those conversations, so they're ignored.
* **2** were **abstain** — the grader saw them but couldn't decide, so they're ignored too.
* That leaves **50 answered** verdicts. Of those, **47** were compliant.
* 47 ÷ 50 = **94%** compliant rate.
So three conversations had a real accuracy problem worth looking at — and those three are exactly what you'll find when you drill in (below).
## The assessor toggle — whose verdicts am I looking at?
At the top of the Overview is an **assessor toggle** with three options:
* **AI + Human** — verdicts from both the AI judge and human reviewers, combined.
* **AI judge** — only verdicts produced by the AI judge.
* **Human** — only verdicts produced by human reviewers.
This toggle changes **whose** verdicts every card is counting. The same metric can have very different numbers depending on which option you pick, because AI and human reviewers don't always agree.
The assessor toggle is **not** the same thing as a metric's scoring-mode badge (auto / hybrid / human\_only). The toggle controls **what you're currently viewing**; the badge describes a metric's **capability** — whether the AI judge is trusted to score it automatically. Both AI and human verdicts can exist for a metric no matter what its badge says. See [Graduation](/evaluation/graduation) for what the badge means and how a metric earns it.
If a metric you think of as "Human only" is showing surprising numbers — a suspiciously high count, or a rate that doesn't match what your reviewers told you — check the assessor toggle first. You're very likely viewing **AI judge** or **AI + Human** verdicts rather than just the human ones. Flip the toggle to **Human** to see only what your reviewers actually graded.
## The date-range filter
Next to the assessor toggle is a **date-range filter**. It scopes every card to verdicts from conversations in that window, so you can compare last week to this week, or zoom into a specific incident. The counts and the compliant rate recompute for the range you choose.
## Drilling into a metric's exceptions
A compliant rate tells you *how many* conversations had a problem. To see *which ones* and *why*, click into a metric to open its drill-down.
The drill-down lists that metric's **exceptions** — the individual verdicts that were **not** compliant. Each exception is a real conversation that failed the check, shown with the supporting evidence:
* The **criterion** that failed.
* A verbatim **quote** from the conversation — the exact line the grader anchored on.
* The grader's **reasoning** and **confidence**.
This is where a score becomes actionable. Instead of "94% accurate," you see the three specific moments where the agent said something wrong, in the agent's own words, so you can decide whether it's a prompt fix, a knowledge-base gap, or a one-off. You can open the underlying conversation from here to read the full [transcript](/conversations/conversation-detail).
## Run tiles and the runs-by-day chart
Below or alongside the metric cards, the Overview also summarizes the **runs** themselves. A **run** is one assessment of one conversation by one assessor — the unit of work behind every verdict.
You'll see three **run tiles**:
| Tile | What it counts |
| ------------- | ------------------------------------------------------ |
| **Total** | All evaluation runs in the current range. |
| **Completed** | Runs that finished successfully and produced verdicts. |
| **Failed** | Runs that errored out and produced nothing. |
A healthy **failed** count is near zero. If it climbs, runs aren't finishing — a good sign to investigate before you trust the rates above them, since failed runs contribute no verdicts.
The **runs-by-day chart** plots run volume over time, so you can see when evaluation activity spiked or went quiet across the selected date range.
## Permissions
Viewing Analytics requires the `assessments:read` scope. Creating, editing, triggering, or graduating metrics requires `assessments:manage`.
## Next steps
Learn whether the AI judge actually agrees with your humans — using Cohen's κ, not raw percentages.
Understand the scoring-mode badge and how a metric earns automatic AI scoring.