Calvin

• Lesson content (frozen markdown drop): The course is a drop of markdown lesson files, grouped by level. Each lesson has micro-doses, one named Trap Factor and its Fix. The files are treated as frozen: the parser adapts to them, never the reverse.
• Loader and parser (server-only, cached): Reads the content folder with node:fs once per process, caches it and parses it into typed modules. Course packagings (a free Foundations on-ramp, a Sprint and a Mastery track) are defined as data over the same lessons.
• Lesson player (interaction stage): Each micro-dose runs an interaction (dragging a dot of light to form vowels, snapping syllable blocks, sorting sounds) inside one shared InteractionStage wrapper. The demo re-implements dose 0.3, syllable blocks.
• Supabase and Postgres (scaffolded): Supabase auth, a Drizzle schema and the session helpers are in the repo as stubs. The stats layer has one swap point for them, but no feature code runs through them yet.
• Learner stats (local storage today): Progress and streaks are kept by a local store behind one entry point, getStatsStore(). When sign-in lands, that function returns the Supabase backend instead, and the local store stays as the first-login import source.
• Audio clips (female voice set): Clip ids map to served audio files, so a new voice set is a file swap, not a content change. The live app plays audio. The demo here stays silent.
• Lesson content → Loader and parser (parse)
• Loader and parser → Lesson player (modules)
• Lesson player → Learner stats (progress)
• Learner stats → Supabase and Postgres (later)
• Audio clips → Lesson player (clips)

Outlined in the accent colour: the part the demo re-implements. Dashed: scaffolded in the repo, not a working feature yet. Design: one type face (Pretendard) and one accent hue, Cheong blue.

• Worker's phone (Expo / React Native app): A working app: camera capture, mark the spot, category, severity and location, in five languages (Vietnamese, Uzbek, Chinese, Korean, English). Coverage is uneven: Uzbek and Chinese fall back to English on several screens.
• Report server (Node, no dependencies): One Node file with an in-memory store: POST and GET /reports plus the AI endpoints. Restarting it clears everything. There is no authentication. The demo re-implements the report-to-record flow on sample data.
• Anthropic API (two real calls): Two genuine calls to claude-haiku-4-5: a short tip on the worker's description, and a vision draft of a description from the photo. Both fail safe: with no key, a timeout or an error, the report still submits.
• Loop back to worker (designed, not built): There is no notification to the worker when a hazard is resolved, and the app's My reports screen is a placeholder. Translation does not exist either: the console shows the worker's original text.
• Manager console (static HTML, polls): A standalone page that polls the report feed and shows each report as a structured, timestamped record. Its Generate with AI buttons are simulated: a short timeout and hand-built templates, with no model call.
• Government view (aggregates the feed): Reads the same feed and aggregates it in the browser. Showing only totals is a presentation choice, not an access rule enforced by the API.
• Worker's phone → Report server (POST report)
• Report server → Anthropic API (tip, photo)
• Report server → Manager console (polled)
• Manager console → Government view (same feed)
• Report server → Loop back to worker (not built)

Outlined in the accent colour: the part the demo re-implements. Dashed: designed or simulated in the hackathon build, not a working feature.