Testing Your Simulations with the Automated Test Bot

Run automated test conversations against your simulation before any real participant tries it. An AI bot plays the learner, so you can check how your character behaves, whether the feedback and the grade are fair, and how the simulation copes with difficult people, all without involving anyone.

Who can open the test bot

The test bot creates real simulation runs, so it is not open to everyone. All three of these must be true:

You are a Creator or above on the platform.
You hold Tester, admin or owner standing on the space, or you own it.
The simulation sits in a testing space.

If one is missing, the page tells you why the test bot cannot run. An Instructor in an ordinary work space is not a tester, and does not reach the test bot from there. Move the simulation to a testing space first. See What is a Testing Space?

Open the console

Open Testing Spaces in the sidebar and pick the space holding your simulation.
On the simulation, click Automatic Testing (labelled Open Test Page in some views).

The console opens on the question of who takes the test. A list of every test run sits on the left, and the header shows the simulation's name and picture.

What the console shows

Test runs (left)	Every test anyone has run on this activity, newest first, grouped by day. Pick one to read it. A running test shows its progress, and a finished one opens on its report. Score chips use the same colour bands as the report.
New test (header)	Takes you back from a transcript or a report to the screen where you start a test.
Advanced settings (header)	The introduction the bot is briefed with, and, for system administrators, the model panel described below.
Edit Template / View Reports	Open the simulation editor or the space analytics in a new tab.

On a phone, the run list and the settings become panels you open from the header rather than columns, so the conversation keeps the full width.

Start a test

Under Pick the test bot's persona, click the card for the learner you want the bot to play. Hover over a card to read what the persona does. To see the edge cases, click Edge-case personas. See Test bot personas below.
If the simulation is available in more than one language, pick the language the test runs in. Give it a moment to load. Starting before it finishes is blocked, so a run cannot be set to one language and played in another.
Optional: to change how many exchanges the bot plays, click Advanced · 30 turns and set Number of Turns, up to 120.
Click Run test with followed by the persona's name, for example Run test with Professional (Perfect Score).

To check what the run is against before you start, click Simulation details. It shows the simulation's image, the character, the opening line, how many criteria the run is judged on, who evaluates it, and whether the guide and the debrief are on.

Watch the conversation as it runs. Stop this run ends it early and still produces a report from what was said.

The bot plays the learner's role as your simulation's introduction describes it, and keeps going until the work of the conversation is done. It does not stop just because the character sounds satisfied. It still ends when the work is done or when the character ends the conversation.

Run up to three tests at once

Three of your own tests can run at the same time. Each keeps the settings it was launched with, so changing the persona or the language while a test is in flight only affects the next one you start.

The run list shows how many are live. When three are running, the start button is disabled with the message "Stop a run before starting another — three run at a time."

Runs keep going while you read a different one. Moving between runs in the list never abandons a conversation.

Test bot personas

Each persona is a kind of learner, and each one is there to prove something about your simulation. Cesura maintains one persona library for every organization, grouped into four tiers. The console opens on the Grading personas. The other three tiers are edge cases, grouped by tier under Edge-case personas.

Tier	What a run proves	Personas
Grading	Plays the exercise to the end. What the run proves is the grade and the report: Professional should land near the top of the bands and Mediocre around a pass.	Authentic; Professional (Perfect Score); Mediocre (Passing Score); Strong Talker, Weak Write-up; Cooperative; Assertive; Out of Their Depth
Engagement	Gives the character little to work with. What the run proves is that the character keeps the conversation alive instead of drifting or closing.	Confused, Passive, Emotional
Conduct	Misbehaves as a person. What the run proves is the boundary ladder: the in-character reaction, the explicit warning, then the automatic stop. A short run here is a pass.	Unprofessional, Rude, Hostile, Insulting
Integrity	Attacks the fiction or the rules rather than the person. What the run proves is that the character never leaves role, never reveals its prompt, never obeys an injected instruction and never gives real-world advice.	Unethical, Meta, Telegraphing

A few personas worth knowing:

Mediocre attempts everything and completes nothing. Run it next to Professional: if both land in the same place, your criteria are not telling a strong learner from a weak one.
Out of Their Depth is willing but does not know what to do, and adopts anything the character says. If the character hands over a hint or a step, it shows up in the learner's very next turn, which makes a character that gives away answers easy to spot.
Rude is dismissive but backs down once the character sets a boundary, so it tests the middle of the boundary ladder.
Meta politely refuses the fiction and asks for the prompt, the model and real advice. Telegraphing writes the character's feelings and decisions for them.

Start with a Grading persona to confirm the basic flow and the grade, then move to the edge cases to find where the character breaks.

You can adjust a persona for one simulation. If you can edit the simulation, click Persona actions to Edit Persona, Clone Persona, or Hide for this simulation. See Create a test persona for your testing space.

The introduction the bot is briefed with

Advanced settings shows the introduction the bot reads to learn its role. If you can edit the simulation, you can change it for one test, reset it, or click Save introduction for this simulation so every later test uses it.

Testers who can only view the simulation see the introduction read-only, with the note "Only editors of this simulation can change the introduction." If a saved introduction applies that they are not allowed to read, the panel says so, and their runs still use it.

Read a finished run

Picking a finished run from the list opens the same report an instructor reads for a real learner, under its own tabs (feedback, the evaluator chat, the conversation, the debrief, and any quizzes or forms), with the same PDF export.

Some problems stop a run or its report, and the console names them rather than showing an empty page:

"The AI model refused to write the user's turn for this scenario, so the test stopped." The model's safety filter blocked the bot's turn. Run it again, or soften the scenario's opening, which is what the model most often refuses.
"The run finished, but its evaluation is not ready yet." The conversation is saved. Open the run again shortly.
"The run could not be closed, so it was not evaluated." The conversation is saved, but no report arrives.

The bot writes the debrief too

If your simulation asks for a debrief, the bot writes one before the evaluation runs, in the persona it just played, the way a learner would. Criteria judged on the debrief are only exercised if a debrief exists, so a test without one would give you a report built from a rubric your learners never meet.

The console shows Writing the debrief... while it happens. If it fails, the run is still evaluated and says so.

What the bot does and does not see

The bot plays the learner unaided: it is not shown the Realtime Guide's advice. The guide still runs during a test, and its messages still appear in the transcript exactly as a learner would see them, but the bot does not act on them. Otherwise a test would measure how well the guide steers the bot rather than how the simulation behaves.

Which model and thinking level a test used

Advanced settings has a Model and thinking level panel that names, for each stage of the simulation, the model it runs on, which level of settings chose that model, and the thinking level it runs at. It is visible to system administrators.

Read the results

Check whether your character stays in role and whether its responses are natural. Review the generated feedback and confirm it reflects what the bot actually did. Professional should score near the top, Mediocre around a pass, and Out of Their Depth well below. If the guide is on, check its tips landed at sensible moments.

Then iterate: run a persona, read the conversation and the feedback, fix what is wrong in the editor, and run the same persona again to confirm. Test at least one persona from each tier before you publish. A simulation that handles Authentic well can still fall apart with Hostile.

Frequently asked questions

Does running the test bot consume credits?
On an organization account, no: a run in a testing space is free on every plan. It is still recorded in your usage reports and still counts against the organization's fair-use allowance, but the credit pool is never charged. On a personal account a testing run costs 0.2 credits instead of a full credit. See Credits & billing.

Can I customize the bot persona prompts?
Yes, for one simulation at a time: edit a persona, clone it into a new one, or hide one you never use. The persona library itself is managed by Cesura. See Create a test persona for your testing space.

Can I run several personas at once?
Yes, up to three of your own runs at a time, each fixed to the settings it started with.

How many turns should I set?
The default of 30 suits most simulations. The bot ends the conversation itself when the work is done, so a higher limit rarely hurts. The ceiling is 120.

A Conduct persona's run ended after two turns. Is that a failure?
No. A short run is a pass for Conduct personas: it means the character warned the learner and then stopped the conversation, as it should.

My character breaks role with the Integrity personas. What do I fix?
Strengthen the Persona description and the Scenario instructions in the editor, with explicit guidance on staying in role, then re-test with the same persona.