Same prompt for both models, medium reasoning, fresh isolated context, one shot each, no tools during generation, November 18 2026. Frozen outputs are unmodified. Evaluated on PC999 Windows. Model inference is hosted, not local offline inference.

Create a polished Japanese-language CSV sales analyzer, no external dependencies/network. User can paste CSV and click Analyze. CSV columns date,category,amount. Support quoted fields including comma and doubled quote; accept CRLF/LF; skip blank lines; validate real ISO dates YYYY-MM-DD (reject 2026-02-30), finite numeric amounts including negative refunds, disallow empty amount; reject invalid rows but summarize accepted rows with total, accepted count and category totals, and show row-specific errors without discarding good rows. Treat category as untrusted plain text (no HTML execution). Include a prefilled sample, accessible labels, keyboard operation, mobile layout, clear empty/error states. Provide CSV export of grouped totals. Keep implementation concise. Output solely the complete document. Do not claim tests were run.

Test harness: node common-harness.cjs
These are experimental generated examples, not production accounting tools. The Node harness measures logic and generated Blob text, not browser download completion. See published article for limits. No secrets or real business data are included.
