> ## Documentation Index
> Fetch the complete documentation index at: https://trygradient.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Assessments

> How Gradient assessments work

# Assessments

An **assessment** is a structured evaluation that measures a candidate's ability to use AI tools effectively for knowledge work. Think of it as a realistic work sample test - but instead of a whiteboard coding exercise, candidates produce real deliverables with AI assistance.

## What's in an assessment?

Each assessment contains:

* **Phases**: A fixed three-phase model set by Gradient (see below). You edit the **Task** phase; the **Explore** and **Reflection** phases are fixed.
* **Task brief**: The realistic work prompt for the Task phase (for example, "Create a product strategy deck for entering the European market").
* **Deliverable type**: The output format the candidate produces: presentation (`pptx`), document (`docx`), email, or an AI `workflow`.
* **Connectors & data**: The data sources a candidate's AI can search (some useful, some distractors). Configured in the build wizard's Data step. See [Connectors](/concepts/connectors).
* **Candidate permissions**: What candidates may do during the assessment (see [Candidate permissions](#candidate-permissions)).
* **Admin guide**: Reviewer-facing notes (overview, what good looks like, pitfalls, buried info) that calibrate scoring.
* **Scoring rubric**: The criteria used to evaluate the candidate's work. See [Scoring](/concepts/scoring).

## The phase model

Every assessment runs on the same fixed three phases. Gradient sets the shape so results stay comparable across candidates and assessments.

| Phase          | Time                        | Graded | Editable by admin |
| -------------- | --------------------------- | ------ | ----------------- |
| **Explore**    | Fixed 5 minutes             | No     | No                |
| **Task**       | Admin-set                   | Yes    | Yes               |
| **Reflection** | Fixed 10 minutes (optional) | Yes    | Toggle on or off  |

Explore is an ungraded orientation window. Task is the graded work the admin designs. Reflection, when enabled, asks a locked, identical prompt so reflections are comparable. See [Candidates & Sessions](/concepts/candidates-and-sessions#the-phase-model) for the candidate's view.

## Assessment lifecycle

<Steps>
  <Step title="Draft">
    Starting from a [role](/concepts/roles), you design the assessment in the build wizard: the task brief, the Data step (connectors and seeded documents), settings, the optional reflection, and the rubric. Gradient can suggest a first draft from the role's job description.
  </Step>

  <Step title="Active">
    The assessment is ready to accept candidates. You can invite people and they'll receive unique links.
  </Step>

  <Step title="Scoring">
    After candidates submit, the scoring engine evaluates their work automatically.
  </Step>

  <Step title="Review">
    You review scores and adjust if needed.
  </Step>
</Steps>

## Templates

Gradient includes pre-built assessment templates for common roles:

| Template                                     | Role                                   | Deliverable |
| -------------------------------------------- | -------------------------------------- | ----------- |
| Retail Chain Digital Transformation Strategy | Strategy Consultant / Business Analyst | PPTX        |
| Sales Territory & Headcount Planning         | Sales Strategy / RevOps Analyst        | DOCX        |
| Product Go-to-Market Strategy                | Product Manager                        | PPTX        |
| Sales Transition: Account Takeover Strategy  | Account Executive / Sales              | DOCX        |

Pick a template as a starting point when you build an assessment under a role, or start from a custom brief.

## Signal vs. noise

A key feature of Gradient assessments is the **signal/noise pattern** in data sources. When you attach documents to an assessment, you mark some as **signal** (relevant to the task) and leave others as **noise** (distractors). This tests whether candidates can identify which information actually matters - a critical AI fluency skill.

<Tip>A good assessment has a healthy mix of signal and noise. Too little noise makes the task trivial; too much makes it frustrating.</Tip>

## Candidate permissions

Each assessment controls what candidates may do in their workspace. There are five permissions:

| Permission                 | Default | Effect                                         |
| -------------------------- | ------- | ---------------------------------------------- |
| Browse connector documents | On      | See individual documents within each connector |
| Add custom MCP servers     | Off     | Connect their own MCP servers                  |
| Manage skills              | On      | Add, edit, or toggle skills                    |
| Edit agents.md             | On      | Edit the AI's system prompt                    |
| Add memory                 | On      | Add persistent memory entries                  |

The first two govern [connectors](/concepts/connectors); the rest govern the AI workspace. These map to the `candidate_permissions` object in the [Assessments API](/api-reference/assessments#candidate-permissions).