Trust is the whole product
Students only speak freely to an AI they trust, and professors only adopt a tool they can rely on. Archimedes is engineered around both — anonymity for students, enforceable control for faculty.
Anonymity for students
Professor dashboards never display student names or IDs. You see aggregate trends and pseudonymized transcripts — never a name attached to a chat.
Control for faculty
A four-layer prompt architecture enforces the intent of each professor’s rules and resists prompt injection, persona attacks, and jailbreak attempts.
Institution-ready
Native Canvas LTI 1.3, single-use launch nonces, and role-based access built for the realities of a university environment.
Decoupled by default
Aggregate over individual
Insights describe the class — “68% of students struggled with the chain rule” — rather than any one person. When a professor opens the chats behind an insight, sessions appear as “Chat 1, Chat 2…” with no identity attached, and the numbering isn’t stable across views.
Honest about small classes
In a tiny class, even anonymized content can hint at who wrote it. Rather than pretend otherwise, Archimedes shows professors a clear caution when participation is low — so transcripts are treated with appropriate care.
Answer keys stay sealed
A confidential answer key is used only to check a student’s reasoning and guide them. It’s guarded by explicit directives, and professors can probe those guardrails themselves to confirm the key never leaks.
No surveillance, by design
No browser locking, no screen monitoring, no anti-cheat spyware. Archimedes earns compliance by being the AI students actually want to use — not by watching them.
The four-layer guardrail
Every student message is answered through a prompt assembled from four layers — so your rules are enforced by intent, not just wording.
Core persona
A consistent, encouraging academic tutor identity.
Intent-based security
Enforces the spirit of your rules and refuses to reveal its own instructions or be jailbroken.
Your dynamic constraints
Your allowed/restricted toggles, custom rules, and any assignment context — injected per session.
Contextual refusal
When a boundary is hit, it redirects to an approved way to help instead of a flat “no.”
Adopt AI you can stand behind
Give students a tutor they trust, and give yourself insight you can act on.