I built a compact text classifier with embeddings

I kept the flow tiny on purpose. One model, one contract, and hard failover rules.

• I pinned response shapes in a small schema.
• I kept context windows fixed per call.
• I rejected malformed output before it could touch state.

# quick guardrail check
node -e "console.log('schema ok')"

Things I kept

• explicit fallback path
• max token limit
• manual review on schema evolution