Atlas, the AI Companion

Atlas is the learner's companion in Journey: it orients them, explains where they are, and helps them think without handing over answers. It is not an open assistant. Code decides when Atlas speaks and what it may know; the AI only finds the words.

The 30-second version

Atlas is learner-only. Rules decide; the LLM words "one calm line". In chat it is grounded only in an allowlist of the learner's own course facts, follows a help ladder that starts light, and never gives or confirms an answer. It is quiet while a check is open and inside assessments. Its core philosophy: increase learner awareness, not learner dependency.

What Atlas is

Atlas is a reflective learning partner, not a tutor and not an answer engine. It appears on Today, beside the course and in the lesson as a chat dock, and it knows where the learner is because the grounding tells it — it never asks "what are you working on?" when it can already see.

Atlas intentionally avoids	Atlas does
Giving or confirming answers · finishing the learner's reasoning · writing milestone work · numbers about the learner · naming their feelings for them	Orients the day · explains what a result or a page means · asks a good question back · brings back the learner's own earlier work · points to the Milestone Guide · keeps the learner's choices visible

It speaks about learning in the learner's words — solid, firming up, you're here, ahead; for a check's result, solid, nearly there, not yet, a mix-up, or "you weren't sure yet" — and never in engine terms, scores or grades. It makes no academic decisions.

Rules decide, the AI words it

When Atlas speaks first, the decision is made by deterministic rules, never by the model. The rules read signals such as repeated wrong practice in a sitting, a long quiet stretch on a lesson, or a check that just came back, and they follow three defaults:

Silence is the default. A rule must fire; no rule, no line.
One calm line. At most one rule fires at a time; outranked rules wait for the next look.
No nagging. A declined offer isn't repeated in that sitting.

Only then does the model word the line, from facts the rule supplies. Counts and durations are turned into words before they reach the model, so no number about the learner can travel onward. If the wording fails validation, a fixed code sentence is used instead.

The grounding allowlist

In chat, Atlas sees a grounding block built by code from an allowlist: the course, the current lesson and its content, the learner's picture in words, recent conversation, and — for revision and milestones — plain facts such as whether a look-back is due, which lessons open items came from, and the next milestone's readiness line and the outcomes it covers. Every field is a string or a yes/no; numeric values are refused. Everything in it is already visible to the learner. It holds no check answers, no rubrics and no milestone work, and Atlas has no tools to fetch more.

For engineers — the firewall

The grounding schema is the firewall: allowlisted keys only, no numeric leaves, no internal band names — enforced by tests. Learner messages are treated as data, never as instructions. Every reply begins with a machine line naming the help level used; it is parsed and stripped before anything reaches the learner. A reply with no valid line, a level above the cap, or a lint failure is refused and re-asked once; a second failure shows a plain "ask me again" line and is audited.

The help ladder

Inside a lesson, Atlas starts at the lightest level that fits and moves down only when the learner asks again or the previous level didn't help:

Point to the part that covers it→ Ask a question back→ Bring back their own past work→ Explain it another way, from the lesson→ A worked example on a different problem

Asking back is the default first move. Whatever the level, Atlas never gives or confirms an assessment answer, finishes the learner's reasoning, or writes their milestone work. Asked about an assessment, it sends the learner to the Milestone Guide — never straight into the assessment — and describes only what the guide says.

Quiet zones

Atlas steps out where its presence would compromise the evidence:

While a check is open, the chat composer steps out and the server refuses chat — not just the screen.
Inside assessments — the milestone work and the spoken defense — Atlas is not mounted at all.
Revision items are treated like an open check: Atlas may explain what the Revision page is, but never answers or hints at an item.

The Milestone Guide is not a quiet zone: Atlas can explain what the guide says.

Atlas on Today

On Today, Atlas gives an orientation line (where you are, the best next step) and a reason line under the Continue button. Code picks the facts; the model may reword them, once a day. A third line is a quiet revision reminder, shown only when:

three or more revision items are waiting in a course, or
that course's milestone is open or next and has required items.

It names the course with the most open items, says how many short things are waiting and about how long they take, and offers one link: Start now. This is the one place Atlas states numbers about the learner, and they match the Revision badge exactly. There is no dismiss button; ignoring the line is the learner's choice, and it disappears when the items are cleared. It never appears inside lessons or assessments.

Let Atlas know you better

On their profile, learners may tell Atlas four things in their own words — what you do, why this programme, when you study, what you want to do — at most 200 characters each, editable and clearable at any time. It is optional and never nags. These words help Atlas talk to the learner as a person; they never change what the learner is given, because nothing in routing reads them.

Prompts & models

Every prompt Atlas uses is held in the prompt registry: versioned, and approved before it serves. Atlas's system prompt has two parts. The voice — warmth, register, how offers are worded — is editable through the registry with version history. The constitution — the hard rules, the help ladder, the scope line — is versioned in code and read-only, and it is placed last so no voice edit can remove or outrank a rule. The model behind each step is chosen in Admin Center from a catalogue; nothing is hard-coded.

Where it sits

Atlas is a runtime component of Maestro Journey — it exists because a specific learner is interacting. It reads the learner model through its grounding, speaks within its rules, and writes interaction patterns back as learning signals. The Studio assistant for course authors is a separate system and is not covered here. What Atlas can see is deliberately narrow, and that narrowness is the point.