This test used two translated public examples and four authored test notes. I tested typed text in TiinyOS with Qwen3.6-35B-A3B and reasoning on; I did not test voice dictation or save records to an app.
A note like “95-grade fuel, 42 litres, 367 yuan and 8 jiao” has three numbers that belong in different fields. Add “30 yuan for a wash, sorry, 25” and a car-cost log can become wrong before I total it.
I supplied candidate numbers and dates alongside each of six short notes, then asked Tiiny for draft records to review. The first fuel note came back with 95 as the fuel grade, 42 as litres and CNY 367.80 as the price. Another note mentioning fuel yesterday and a car wash today became two records with separate dates. For the correction, Tiiny kept one CNY 25 wash record instead of counting both 30 and 25. When a note said “42 litres” but gave no price, it left the amount blank and flagged it for review.
The planned oil change exposed the part I would fix before using this with a real expense tracker. Tiiny labelled the possible CNY 450 cost as planned and unpaid, but still put 45,000 fen (CNY 450) in the same amount_fen field used for actual expenses. A tracker that ignores the status could quietly add it to spending. I would keep estimates in a separate field and review every draft before saving it.
These were draft records from one test, not evidence of a working automatic bookkeeping system.
No comments yet.