Technical Definition: AI model, thinking level & templates (admin)

The Technical Definition section holds advanced settings. These pages appear only for system administrators — regular users will not see them, and their simulations simply use the platform defaults. There are two pages.

💡 The voice settings page (once called Audio) used to live here. It is now Delivery Mode, under Simulation Display, because what it configures is what a learner sees and hears — and it is no longer administrator-only. See Set up a voice simulation.

AI Model

The AI Model page chooses which AI model powers the simulation. The default is Auto (Default), which uses the system’s default model chain with automatic fallback — the platform picks the best available model and falls back to another if one is unavailable. Administrators can select a specific model instead.

Thinking level per stage

Beneath the model sits Thinking level per stage — how hard the model reasons at each point of the simulation. It is set separately for four stages, because they are different jobs:

Stage	What it covers
Simulation chat	The learner’s conversation turns.
Guide	In-simulation guide tips. Applies whether or not this template pins a model.
Feedback report	Generating the feedback report.
Feedback chat	The conversation about the feedback report.

Each is one of Not set, Minimal, Low, Medium or High. Not set states no preference here, so the platform-wide level from the admin Models page applies. Minimal is the floor, not off — there is no way to switch reasoning off entirely. Higher levels cost more and take longer, so raise them where the task is genuinely hard (feedback reports, complex evaluation) and leave the simple stages alone.

Which setting wins

The rule is short: whoever names the model owns the reasoning on it. A template that pins its own model sets its own levels; a template that pins nothing follows the organization’s default model and that organization’s levels. The levels are available once a model is pinned — on automatic, the organization’s own level applies when the organization pins a model and sets one; otherwise the platform-wide level does.

A system administrator sets the same four levels for an organization beside its default AI model, in Organization Management → Policies.

The Guide is the exception. Nothing pins the guide’s model — it always runs on the platform’s assistant model — so its level applies with or without a pinned model, and falls back from template to organization to the platform task setting.

When a level does nothing

Only some Gemini models reason at a set level. Pick a model that does not, and the page tells you: the levels are saved, and no request sends them. Swap back to a model that honours them and the setting applies again — nothing is lost in between.

Organizations running on their own AI provider credentials (BYOK) are a deliberate exception: Cesura sets no thinking level for their simulations at all. The levels are still saved and apply again if that changes. See Run simulations on your own AI provider account (BYOK).

Pinning the conversation’s own model

System administrators can now set the conversation stage’s model and thinking level on its own, from the admin Models page, instead of reordering the platform fallback chain and moving every other task along with it.

What an unset thinking level now means

This reversed. Leaving a level unset used to mean “the model’s own default decides”. It now means “no preference is stated here, so the platform level applies”.

Two populations behave differently as a result: a template, or an organization, that pins its own model and leaves the Feedback report or Feedback chat level unset now runs those stages at the platform’s level for that task rather than at the model’s own default. Everything else is unchanged, and organizations running on their own AI credentials (BYOK) are unaffected — no thinking level is sent for them either way.

Templates

The Templates page shows administrators, read-only, the full prompts the simulation sends to the AI — for the simulation itself, the feedback report, the feedback chat, and the guide advice, with a separate guide prompt for each of the Beginner, Regular and Custom advice levels. This page is only available once the simulation has been saved.