Case → role-slot preprocessing + an ultra-small logistic-regression classifier. Training and inference run entirely inside this browser tab; nothing is uploaded; no GPU; tens of KB. The routing demo below uses Japanese internal-regulation questions (Japanese marks roles with case particles); §4 shows the multilingual + injection story.
Loading WASM…
1. Train in your browser(teach the classifier typical questions per regulation clause, including word-order variants)
2. Question → clause routing(Japanese input; try changing the word order)
3. Audit log(each decision step + a hash chain = tamper-evident. Deterministic: same input → same result → same hash)
#
Question
Steps
Result
hash(prev+now)
4. Multilingual + attribute injection(case-marking languages get it natively; languages without case are made equivalent by injecting role attributes — all order-invariant)
Clause → answer (deterministic rule) is a separate layer. Here we demonstrate "question → clause" with a classifier trained inside the browser. Verify: open DevTools → Network — routing a question sends zero requests (only static code is fetched).