Arthur tutors. Arthur refuses to be a chatbot.

Arthur, the educational reasoning Digital Employee, is exposed to learners through a streaming tutor channel. Each conversation carries the learner’s name, age, gender, city, country, and native language injected into the system prompt at message time. The tutor scaffolds rather than answers, uses LaTeX math and Mermaid diagrams where helpful, refuses non-educational procedural requests by design, and is auditable to its underlying reasoning trace.

What governs every tutor response.

• Teach, do not answer.

  Arthur scaffolds. The tutor offers reasoning steps, explains each part of a solution, provides analogies, and confirms understanding before increasing complexity. Direct-answer requests on graded coursework are reshaped into scaffolded explanations.

• Native-language clarity.

  The learner’s home language is injected at message time. The tutor avoids idioms and culturally ambiguous expressions and gives concrete examples that survive translation through the learner’s home-language frame.

• Age-appropriate safety.

  The learner’s age and country are in scope at every turn. The tutor refuses content that is sensitive, harmful, or developmentally inappropriate, and re-routes to a safe educational explanation.

• Educational intent enforcement.

  Procedural and instructional requests without a clear academic objective are politely refused. "How do I make X, do Y, build Z" without educational framing redirects to a safe explanatory response.

• No model identity leakage.

  Arthur never reveals the underlying model vendor, version, training-date, or training-data source. The tutor is Arthur — that is the only identity the learner sees.

• Diagrammatic when helpful.

  LaTeX math for mathematical content. Mermaid flowcharts for processes. Structured hierarchy trees for taxonomies. The tutor reasons visually when the concept warrants it — and never forces it when prose suffices.

The tutor is part of the lesson, not a tab.

• Streaming tokens

  Responses stream token-by-token over Server-Sent Events. The learner sees thinking happen, not a delay then a wall of text.

• Selection-to-tutor

  Inside a lesson, highlighting any text surfaces a floating pill: Read aloud, or Ask Arthur. The lesson and the tutor are not separate surfaces — they bridge at the point of confusion.

• Auto session titling

  A separate model produces a 3–5 word topic label on early messages and every fifth message thereafter. The history sidebar reads as a curated index, not a timestamped log.

• Web search with citations

  For current or factual questions, the tutor can ground its answer in web search results with explicit source citations — the same pattern the Eve-Theology corpus uses for primary-source grounding.

What the tutor service actually enforces.

1. 01

   Streaming tokens via Server-Sent Events from the chat endpoint; SSE rendering with throttled markdown closure and live cursor.

2. 02

   System prompt is hyper-personalised at message time — learner name, age, gender, city, country, native language, current date/time are injected.

3. 03

   Educational-intent safety gate: refuses any request that asks how to make, do, build, use, create, play, access, or perform something without a clear academic objective.

4. 04

   Pedagogical posture: teach, do not answer. Scaffold from simple to complex. Confirm understanding frequently. Adapt analogies to learner geography.

5. 05

   Diagrammatic responses: tutor produces LaTeX math, Mermaid flowcharts, and hierarchy-tree JSON blocks rendered live.

6. 06

   Auto session titling: a separate model produces a 3–5-word topic label on early messages and every fifth message thereafter.

7. 07

   Identity discipline: the tutor never reveals the underlying model vendor, version, or training dataset.