Aria — The AI Tutor

Aria, the AI tutor: the persona, how each answer is grounded in the NZ Curriculum, and what keeps her answers accurate and safe.

Status: Live / built. Core feature, present in both family and school account types.

What Aria Is

Aria is the AI tutor persona children interact with. Technically, Aria is a system-prompted instance of Anthropic's Claude (currently claude-sonnet-5, upgraded from claude-sonnet-4-6), accessed server-side via the Anthropic API. Children can optionally rename Aria in the family-account version.

Interaction Modes

Children can communicate with Aria by typing or by voice:

Voice input uses the browser's native speech recognition (webkitSpeechRecognition) — audio never reaches Synapse's servers; only the resulting transcribed text does (see Privacy Model for the full data-handling picture, including a documented iOS Safari mic reliability issue that was investigated and resolved for supported browsers).
Voice output uses Google Cloud Text-to-Speech (en-AU-Chirp3-HD-Aoede), giving Aria a spoken voice throughout the app — useful for younger children and children with differing needs.

Pedagogical Design

Exploration-first conduct. Aria's default behaviour is not overridden by teacher-set curriculum steering (see below) — steering is injected as soft context, weighting emphasis rather than branching Aria into a separate "mode."

The controls that keep what Aria says accurate and age-appropriate — curriculum grounding, no internet access, the calculator requirement — are set out with their documented limits under "How Aria's Answers Are Kept Accurate and Safe" below, rather than split between here and three other documents.

The Dual-Channel / Supervisor Notes Architecture

This is a foundational architectural pattern used throughout Synapse, not just in tutoring:

Aria's replies can contain hidden, bracketed structured markers, e.g.:

[ASSESSMENT: subject=literacy, phase=Developing, note=...]
[STYLE: preference=visual, confidence=medium, note=...]
[CONNECT: student_a=..., student_b=..., shared_interest=..., suggestion=...]
[SEARCH_NEEDED: topic=..., student=..., reason=...]
[WELLBEING: note=...]

The server strips these tags before the response reaches the child (stripSupervisorNotes) — the child only ever sees clean tutor text. The tags are separately extracted (extractSupervisorNotes), stored encrypted, and used to auto-update the student's literacy/numeracy phase and learning-style profile. They are never sent to the student client and are accessible only through authenticated teacher or parent routes.

This pattern is what powers progress tracking, wellbeing flagging (see Social Resilience and Privacy Model), and peer-connection suggestions, all without exposing raw conversation content through those channels.

Handling Questions Beyond Aria's Reach

Curiosity outside the strict curriculum is welcomed, not restricted — Aria steers off-topic questions back toward learning rather than shutting them down. A referral to an adult only happens in two narrower cases: the question is genuinely beyond Aria's own knowledge or ability, or it touches something outside what's normally age-appropriate or safe content for that child. Even then, the response is a warm redirect, not a shutdown — the judgment about what's right for a particular child belongs to the adult in their life, not to Aria.

The wording of that redirect differs honestly by account type. For a school account, Aria tells the child she's let their teacher know and they'll follow up — true, since school sessions do route to teacher review. For a family account, there is no teacher and nothing is automatically passed on — Aria instead tells the child plainly that it's something to ask an adult about, with no false claim that anyone's been notified. See Privacy Model for what does and does not reach an adult afterwards.

How Aria's Answers Are Kept Accurate and Safe

The single place for what actually constrains Aria's output, and what does not. Every point below is also stated in the document it cites — this section consolidates them so the question "can I trust Aria" does not retrieve only the reassuring half.

What is in place, before an answer is generated:

Curriculum grounding. Aria is grounded in an authoritative NZ Curriculum reference store rather than relying on model training knowledge for curriculum content — see Curriculum Alignment for why this pattern was chosen and how it works.
No open internet access. Aria has no internet or search tool at all, so she cannot answer from an unvetted source. Note the enforcement mechanism: this is a system-prompt constraint, not a technical sandbox — an important distinction, and the same one that applies to every behavioural rule below (see Risks and Known Limitations).
Arithmetic accuracy. Language models are unreliable at mental arithmetic. Aria is provided a formal calculator tool and is required to use it for any arithmetic beyond trivial single-digit operations, rather than generating computed answers directly.
Age-appropriateness and referral. Content Aria produces is designed to stay within safe, age-appropriate boundaries (Student / Child Experience), and questions that go beyond what is age-appropriate for a particular child, or beyond Aria's own knowledge, are referred to an adult rather than answered — see "Handling Questions Beyond Aria's Reach" above. These are instructions Aria follows, not filters she is passed through.

What is not in place — nothing checks an answer before the child reads it:

Oversight is retrospective, not pre-delivery. Teachers review AI-generated content after it has already reached the child, via transcript access, not before (Adversarial Review Process, finding #2). This is assessed as inherent to the product's design — partially mitigated by transcript access and messaging, not eliminated.
This cannot satisfy Ministry of Education guidance, which calls for teachers to vet AI output before it reaches students. No direct child-facing AI system can literally do that. The repository's own framing of the mitigation is "tight system-prompt constraints plus post-hoc transcript access" — a partial answer, explicitly not a resolution (Adversarial Review Process finding #3, Risks and Known Limitations).
The observation system has no independent validation. The AI-derived progress/observation system profiles children without any independent, non-developer validation of its accuracy or safety (Adversarial Review Process, finding #1). The developer's response to this finding was honest acknowledgement, not a claimed fix.
No independent third-party technical or safety review has been conducted. The technical and safety claims made about Synapse have been verified by no one other than the developer himself, aside from the structured self-critique process — which is valuable but is not independent review and has never been offered as a substitute for one (Governance Model, Adversarial Review Process). A review has been offered at the developer's cost if a school proceeds; it has not happened.
No automated safety system is presented as 100% reliable, and the disclaimer says so rather than claiming otherwise (Student / Child Experience, Privacy Model).

Who can actually review afterwards differs by account type:

School accounts: teachers can read full transcripts of a student's conversations with Aria via an authenticated route — a resolved, deliberate position, offered openly as a school choice (Privacy Model). This is what the retrospective-oversight model rests on.
Family accounts: there is no parent-facing transcript route at all. The parent dashboard returns learning observations, phase badges, session counts and wellbeing flags — never conversation text. A parent's ability to read everything is still real, but it works by signing in as their child with that child's PIN (which a parent can reset at will), not through a parent-side review surface (Family Features). So the after-the-fact check the school model depends on is materially less available at home.

An adjacent gap, easily confused with this one: no automated content filter sits between a teacher's message and a student — teacher→student messages are stored and delivered as written, and this is a confirmed, still-open roadmap item (Security Architecture, Risks and Known Limitations, Future Development). That gap concerns human teacher messages rather than Aria's output, but it sits in the same channel a child reads.

System Prompt Construction

Each tutoring request assembles a system prompt server-side from:

A base system prompt (curriculum framework, teaching-style frameworks, assessment protocol)
The student's profile (pseudonym, age/year level, interests, learning style)
Curriculum context relevant to the session
Any teacher-set curriculum steering (see Curriculum Alignment)

No student identifier is passed to the Anthropic API — the student is identified server-side by session token only; Anthropic receives a conversation thread with no information linking it to a specific real child.

Prompt Caching

Prompt caching is used throughout (roughly two-block split: stable instructions/curriculum vs. per-turn content), giving approximately a 70% cost reduction on repeated input, since most of what's sent (instructions, curriculum chunks) repeats turn to turn.

Known Historical Issue (Resolved)

A voice-input bug (mic icon animated but no transcribed text appeared) was investigated in August 2026. Root cause on one platform was a muted OS-level microphone (user error, not a code issue); the underlying diagnostic logging gap that made this hard to triage was fixed regardless (see Risks and Known Limitations for the general lesson about silent failure modes).