Make-or-break test: does a small encoder-only NER (token-classification, ~66–110M, no generation, no KV-cache) RUN on this Motorola G where the 256M decoder crashed — and does it cleanly tag PERSON / LOC spans in real lease prose? NER handles the hard spans (names, address); MONEY + DATE are the cheap deterministic regex pass (shown alongside).
5 real lease templates filled with fake persona data — each in its own distinct prose style. Tap one, then Run.