# Design

## Product Boundary

The first user is a software job seeker. The product should begin with
technical interview preparation because it gives clear task loops:

- ask a question
- receive an answer
- grade against a rubric
- ask follow-ups
- extract memory
- recommend review
- show progress and next challenge

General students remain a future expansion path, but the first requirements
should not be diluted by full K-12 scope.

## Architecture Boundary

Use a Go backend as the product service boundary. Internalize `agent-farm-go`
workflow patterns and contracts inside that backend boundary while keeping agent
behavior configuration-first where possible. Use `third-one` as the LLM
execution kernel with `deepseek-v4-flash` as the default configured model
target.

The product backend owns durable user, learner, memory, ontology, and asset
records. External memory or graph projects may inform design or become adapters,
but they should not own product privacy or tenant semantics.

## Memory Boundary

The platform should not model memory as a flat RAG corpus. It should keep
structured learning state:

- learner profile
- concept mastery
- misconception
- evidence
- intervention
- review schedule

Every durable memory update must include evidence so the product can explain why
it believes a learner is weak or strong on a concept.

## Ontology Boundary

Uploaded materials should become source-backed learning graphs. Inferred gaps
and generated explanations are candidates until reviewed or otherwise validated.

## Visual Asset Boundary

Image generation should support diagrams and slide-like learning slices, but the
asset pipeline must preserve prompt lineage, source concept links, and review
state. The desired image provider key is `gpt-image-v2`, but production
implementation must verify the current OpenAI API model name before wiring.

## Gamification Boundary

Use game-design principles to create healthy persistence: adaptive challenge,
visible growth, clear goals, strong session endings, and long-term readiness
progress. Do not optimize for compulsion alone. Random rewards, punitive streaks,
and shame-based leaderboards are out of scope for the first product baseline.

The initial learning loop is:

- readiness goal
- interview question
- answer
- rubric feedback
- follow-up or correction
- memory update
- visible progress
- next best challenge