Pin down the input before it reaches the AI. Normalize notation variants, paraphrases, dates, units, and person ambiguity deterministically inside the browser, improving the reproducibility of RAG retrieval keys, audit logs, caches, and the BYOK gateway — an "LLM pre-processing" demo.
The Basic build (RLM core) with S1–S9 normalization integrated as an option. Within the defined normalization rules, inputs that differ in phrasing, word order, full-/half-width form, or English vs. Japanese are folded into the same key whenever they mean the same thing. Everything runs deterministically on WASM in the browser.
Keep only the entity + intent (ABOUT/ASK) and strip away particles, word order, and notation variants. Dates are folded into @DATE:ISO and quantities into @U:hash before normalizing (= same surface even for different notations).
One query per line. See how many ways of saying it fold into how many meanings (keys).
People are matched against an internal HR DB (surname + title / department / email / employee number). An ambiguous "Tanaka-san" is never guessed — it is HELD (fail-closed). Dates become @DATE and amounts are currency-converted. A deterministic audit_key is derived from the 4-tuple. (the HR DB is a demo dummy)
who / when / amount / action, plus the audit_key for that 4-tuple (any tampering necessarily changes it).
A deterministic kernel that turns embeddings into complex operators and measures non-commutativity via the commutator norm ‖QD−DQ‖. Matches native / Python bit for bit down to floating point (f64) — the "middle of the computation" in the BitExact thesis.
The normalization key (norm_key) has been verified to be a byte-for-byte match across native Rust, wasm32-wasip1, this page's WASM, and the reference Python (pipeline.py, 8803).
Standalone tree AI/SlimeTree/slimetree-wasm-8802/ (feature full). impl_v5 and the Python reference implementation are preserved unmodified.