The AI layer a school can switch on without losing control.
Schools block open chatbots for good reasons, and the course still asks students to do AI work. Schanion sits between the two. The school sets the rules, the student gets a guide, and the guide points to where to look instead of doing the thinking.
Block it, allow it, or guard it.
| Block it | Open chatbot | Schanion's guardrail layer | |
|---|---|---|---|
| Who sets the rules | The school, by blocking the site | The company that runs the chatbot | The school, per school and per class. A class can lower the level and never raise it |
| Assessed work | No AI help, and no AI skills either | Can write the answer | An assessed task can be locked, so the guide only points to notes |
| What a student gets back | Nothing | An answer | A clue, a question back, and where in the class materials to look |
| What the school can see | That the site was blocked | Usually nothing | An audit log of what was allowed and denied, with no conversation text |
| Open web | Not applicable | Often on | Always off. Sources are what teachers published |
| Student data | Not applicable | The company's terms | Not sold, not used for ads, not used to train models |
The middle column describes open chatbots in general and varies by product. Schanion does not block websites: that is done by the school's own network. Schanion gives students a sanctioned route to put in its place.
Try the school's side of it.
Switch the policy, then ask as a student. The reply and the guardrail trace change with the rules.
The school sets the rules. The student gets a guide.
AI the school can say yes to.
A student asks
Pick something a student might type, then change the policy and ask again.
A scripted stand-in. Schanion's companion is scripted today, and this is the layer a model would sit behind. No ChatGPT or Claude model is connected.
How one question travels.
- 1
The school's policy loads first
The highest level any class may use, whether uploads are allowed, when a model answer may be shown, and whether this task is locked. A teacher can lower the level for a class and never raise it.
- 2
A safety screen reads the message
Crisis language, bullying, attempts to override the rules, requests to keep secrets, and questions about other students are caught before anything else happens.
- 3
The intent decides the reply
Asking for the answer becomes a clue. Asking for an essay becomes a structure. Asking where something is gets a pointer. A concept question gets an explanation that ends with a question back.
- 4
The reply is built from what the teacher published
It names where to look: the slide, the page, the task sheet, and who published it. It does not browse.
- 5
An output guard checks the reply
It blocks leaked answers, claims to be human, requests for secrecy and manipulation, and checks the level again.
- 6
The school keeps a record, not a transcript
An event says what was checked and what happened. A safeguarding concern reaches a person and carries no conversation text.
What a school controls
- The highest AI level any class may use: off, principles only, guide, or guide with worked examples
- Whether students may upload files, which are read as material and never as instructions
- After how many attempts a model answer may be shown on practice questions, or never
- Assessment mode on a task, which holds across every companion tool
- How long learning memory is kept, and the student can see and delete it
- What happens when a student may be at risk: notify the wellbeing team, or show support only
- Whether first-pass marking suggestions may show marks on school-assessed coursework. Off by default
Who can see what.
The rules are checked on the server, not hidden on screen, and every denial is logged.
| Student | Teacher | Parent or carer | School leader | |
|---|---|---|---|---|
| Conversations with the companionPrivate to the student. | Yes | No | No | No |
| Learning memoryThe student can read, correct and delete it. | Yes | No | No | No |
| Attendance for a lessonA parent sees it only if the student has chosen to share it. | Yes | Yes | If shared | Yes |
| Progress and masteryTeachers see their own classes. Leaders see class-level reporting. | Yes | Yes | If shared | No |
| Safeguarding flagsDesignated staff only, and no conversation text. | No | No | No | Yes |
| Audit log of allowed and denied actions | No | No | No | Yes |
One path for everything.
Signals suggest, a person confirms, and only then is a record made. Nothing skips the middle.
Signals suggest
Check-ins at the door, Practice evidence, Absences, Marking evidence.
A teacher confirms
They verify, edit or reject.
Records are made
The roll, Marks, The catch-up plan, The family update.
Where this stands.
Built and tested
The policy layer, the safety screen, the output guard, assessment mode, the audit log and the permission rules, with 36 adversarial tests that all pass. The companion's replies come from authored content.
Not built
A connection to any language model. No ChatGPT or Claude model is connected, and Schanion is not affiliated with their makers. Adding a model would be a separate decision with its own safety review, and the same tests would be run against it. Nothing has been trialled in a real school.