AI Teacher · the hyper-scoped teaching-agent concept · honest-off by default
A bank of narrow, single-purpose AI agents for teaching — not one big classroom chatbot. And none of them answers a live question by default.
A name like AI Teacher invites one assumption: a live assistant that can answer anything, for anyone, right now. That is not what this is, and we are saying so before we say anything else. The idea behind aiteacher.network is narrower and, we think, safer: instead of one broad model, a growing bank of small, single-purpose agents, each built for exactly one subject or one task, each grounded only in a school’s own adopted material, each built to refuse an answer it cannot ground rather than invent one. Two of those agents exist today as real engineering — a student-tutor seam and a teacher-toolbox seam — and by default, without a school-configured model behind them, both return an honest “no model available” result instead of an answer.
The in-VPC wall, the guardrail stack, and the suggest-only discipline are shipped and always enforced when either seam runs. The served model behind them is off by default until a school turns one on. We say so before we say anything else.
What this page is not claiming
There is no live AI teacher answering a real question for anyone by default today. Without an operator-configured in-VPC model server, the student-tutor and teacher-toolbox seams below both return the same honest “no model available” result — never a fabricated answer, never a guessed accuracy number, never an invented adoption statistic. This is not a general-purpose classroom chatbot, and it does not claim to know a student, a class, or a subject beyond exactly what a school has actually adopted.
The wider “bank” of subject- and task-scoped agents this concept describes is a direction, not a shipped catalog. Today there are exactly two seams that demonstrate the pattern. More narrowly-scoped agents are the stated intent behind the idea, named here as intent, not dressed up as a released feature list with names and dates.
What is built
Every item below is code that exists today. Where a card says the model is gated off by default, that is the honest default for every school unless an operator configures an in-VPC model server — the properties in the other three cards (the wall, the guardrails, suggest-only) are enforced regardless.
The grounded student-tutor seam
A student asks a plain-language question; the seam is built to answer USING ONLY the school’s own adopted lesson content and to end every answer with the citations it used. A question that cannot be grounded in anything adopted gets an honest refusal, never an invented answer. By default no model is served, so today the seam returns the same honest “no model available” result to everyone. Shipped engine — model gated off by default
The class-level teacher-toolbox seam
A separate, student-name-free seam assists a teacher draft class-level artifacts — a scoring-rubric scaffold, a general feedback phrasebank — never a per-student narrative. It runs the same guardrail stack as the tutor seam. Like the tutor, it is not wired to a served model by default. Shipped engine — model gated off by default
The in-VPC wall
Both seams require an in-VPC provider before any request is assembled. A cloud-only configuration is refused before the network call is ever made — an in-VPC model server is the only place this text is ever processed, or the seam does not run at all. This check runs every time, independent of whether a model is served. Shipped
Guardrails on every call
A person-image-generation refusal, a zero-retention assertion, PII redaction on the question before it is ever assembled into a prompt, and a content-safety check on anything that comes back — the same stack for both seams, enforced on every call, not an opt-in setting. Shipped
Suggest-only, never auto-applied
Neither seam writes to a gradebook, a report card, or a student record on its own. A result is something a teacher or student reads and decides about; nothing here is ever saved automatically. Shipped
How the bank is meant to grow — and where the platform stops today
The design point is one shape, repeated: pick one subject or one task, ground it only in what a school has actually adopted, wrap it in the same wall and the same guardrails, and make refusal the default when grounding fails. The student-tutor seam and the teacher-toolbox seam are the first two agents built to that shape. Neither is a general assistant, and neither reaches outside the school’s own material to answer.
What the platform does not yet do is serve a model behind either seam by default, or offer a wider catalog of subject- and task-scoped agents beyond these two. Growing the bank means adding more agents that fit the same narrow shape — a single-subject drill agent, a lab-report feedback agent, and similar — not widening any one agent into a general-purpose assistant. That is an intended direction, not a wired feature, and we are not going to describe it as available to make the page read better.
Student data and consent, said plainly
A student’s own question is treated as student text, scoped to exactly one school by the same FERPA wall the rest of the platform enforces — not a role check an admin can waive. For an in-VPC seam operating on a specific student’s own data, the platform’s consent substrate is fail-closed: a missing, unverified, or withdrawn consent denies processing, and a subject under a do-not-publish or suppression hold is never processed by either seam, in-VPC or not.
The in-VPC path is treated as internal use, not a disclosure to an outside processor, because the text does not leave the school’s own compute environment; a cloud-only configuration would still require the same opt-in disclosure consent the rest of the platform requires before a minor’s data reaches a third-party processor — and for these two seams specifically, that cloud path is refused before any network call is made at all. Nothing produced by either seam is used for advertising, sold, or handed to a data broker.
How the money works, honestly
Neither seam carries a price on this page. The engineering — the wall, the guardrails, the grounded-or-refuse policy, the suggest-only discipline — is part of the platform; a served model is an operator decision a school makes separately.
This is honest-off money: there is no pricing table and no checkout on this page, and nothing here takes a live charge. A conversation with a school works out what, if anything, applies.
Common questions
Is there a live AI teacher answering student questions right now?
No. By default, no in-VPC model is served for either seam below, so a student or teacher gets an honest “no model available” result today, never a fabricated answer. The engineering — the in-VPC wall, the guardrails, the grounded-or-refuse policy — is built; the model an operator would point it at is not turned on by default.
What is a hyper-scoped AI agent, and why not one big classroom AI?
A hyper-scoped agent does exactly one job: one subject, one task, grounded in one school's own adopted material. Two examples exist as engineering today — a student-tutor seam that answers only from adopted lesson content and cites what it used, and a teacher-toolbox seam that drafts a rubric scaffold or a feedback phrasebank at the class level, never a per-student narrative. A single broad assistant that could answer anything about anyone is a different, riskier shape, and it is not what is built here.
Does it ever make an answer up?
The grounded student-tutor seam is built not to. It answers using only lesson references a school has actually adopted and ends every answer with the citations it used; when nothing adopted grounds the question, the seam returns an honest refusal instead of inventing a plausible-sounding one.
Does a student's question ever reach an outside AI company?
Not through these two seams. Both require an in-VPC provider before any request is built: a cloud-only configuration is refused before the network call is ever made, so the student's text is either handled inside the school's own compute environment or not handled at all — there is no third path where it quietly goes to a cloud AI vendor.
How is this different from virtualteacher.network?
virtualteacher.network is the coursework / LMS facet for a class that is not meeting in one room — the course shell, assignments and turn-in, gradebook, and discussion board — and it makes no AI claim at all. aiteacher.network is a different, orthogonal idea: narrowly scoped AI agents assisting one task at a time, whether or not a class shares a room. The names sound adjacent; the products are not the same thing.
Is anything from these agents ever auto-applied to a grade or a student record?
No. Both seams are suggest-only: the result is something a person reads and decides about, never something the system saves on its own. Nothing here writes to a gradebook, a report card, or a student record by itself.
Does this cost anything?
Money is honest-off on this page: there is no pricing table, no checkout, and nothing here takes a live charge. A conversation with a school works out what, if anything, applies.
What this page is, and is not, claiming
AI Teacher names a design choice: a bank of narrow, single-purpose AI teaching agents rather than one broad classroom AI. Two agents exist today as real engineering — a student-tutor seam grounded only in a school’s own adopted lesson content, and a class-level teacher-toolbox seam — both suggest-only, both wrapped in an in-VPC wall and a guardrail stack that run on every call. Neither answers a live question for anyone by default: without an operator-configured in-VPC model, both return an honest “no model available” result, never a fabricated one. The wider bank of subject- and task-scoped agents beyond these two is a stated direction, not a shipped catalog. Student data is FERPA-scoped to one school, consent is fail-closed, and nothing here is used for advertising, sold, or handed to a data broker. This is a for-profit vendor, not a charity. Money is honest-off: no pricing table, no checkout, no live charge on this page. There are no invented stats, adoption counts, or testimonials here, and no competitor is named.