Every business's paperwork has quirks. Thermal paper fades. Two codes on the same receipt look alike. One store prints a till counter where you'd expect the receipt number; another splits a single purchase across two pages.
None of this shows up in a demo. All of it shows up in the first week of real use.
What we've seen
On one store's receipts, a model kept picking up a short till counter instead of the receipt number — confidently, dozens of times. In another case, a receipt listing two identical items came through with only one line, because an early duplicate check compared lines rather than whole receipts.
Both were easy to fix once seen. The lesson was to look for them before launch, not after.
What we test
Before we build any screens, we take a sample of your real documents — the messy ones included — and run them through extraction. Then we check, field by field, what came out right.
Each failure gets one of three answers: fix it with a rule, send it to a person for a quick check, or accept it because it doesn't matter.
You see an honest picture of what AI can do with your documents before you commit to anything.
It also shapes what gets built. Knowing where the reading is weak tells us where the review screen needs to be strongest.
每家公司的单据都有各自的毛病：热敏纸会褪色；同一张收据上，两个编码长得几乎一样；有的门店在单号的位置印的是收银流水号；还有的把一笔消费拆成两页打印。
这些情况在演示里一个都看不到，却会在正式使用的第一周全部冒出来。
我们遇到过的情况
有一家门店的收据，模型一再把一串很短的收银流水号当成单号，而且错得理直气壮，前后几十次。还有一次，一张列有两件相同商品的收据只录入了一行，原因是早期的查重逻辑逐行比对，没有按整张收据判断。
两个问题一经发现，修正都不难。教训是：这类问题要在上线前找出来，别等上线之后。
我们测什么
在做任何界面之前，我们会取一批你的真实单据，连那些皱巴巴、看不清的也算在内，交给 AI 识别，再逐个字段核对，看哪些读对了。
每一类错误，都有三种处理办法：用规则纠正，交给人快速确认，或者无关紧要、直接放过。
在你做任何决定之前，就能清楚看到 AI 处理你的单据的真实水平。
测试结果也决定了系统的做法：知道哪里识别得不稳，就知道核对界面该在哪里下足功夫。