How New Zealand Curriculum grounding works — what is injected into each turn, and what alignment does and does not guarantee.
Status: Live / built (NZ Curriculum). A Japan-curriculum equivalent has been discussed only speculatively — see Future Development.
A deliberate architectural decision, applied consistently across Synapse: curriculum content is grounded in an authoritative reference store, not left to the AI model's own training knowledge. This pattern was explicitly identified as reusable across the developer's other AI products (Waypoint and others), not something specific to Synapse.
A SQLite database (nz_curriculum.db) built directly on the server, containing structured reference data:
This is queried (via a dedicated curriculum.js module) and injected as year-appropriate context into every relevant conversation, rather than trusting the model to recall curriculum specifics accurately from training.
Beyond the structured curriculum database, Aria's base system prompt incorporates broader teaching frameworks:
This distinction was the subject of a real correction during development and is worth stating precisely, since an earlier presentation draft conflated the two:
The product's explicit position, confirmed during presentation preparation: AI-derived progress observations inform, but never replace or carry the authority of, a teacher's formal PAT assessment.
See Teacher Features (School Accounts) for the curriculum-steering slider and lesson-plan generator, which draw on this reference store plus PAT scores, class progress, and session history to produce per-child personalised lesson plans, injected into Aria's behaviour as soft context rather than a rigid override.
The reference store does not gate what a child may ask. A question falling outside it is followed and steered back toward learning, not flagged. What gets referred onward is narrower — a question genuinely beyond Aria's knowledge or ability, or outside what is age-appropriate for that child — and on a school account that referral goes to the teacher for approval rather than being answered from the model's general knowledge or an internet search, which is the same grounding principle applied at the point of live tutoring. See Aria — The AI Tutor.