I spent years staring at systems that could tell me what happened but couldn't tell me what was missing. A log full of successful requests looks exactly like a log where half the requests were silently dropped. You don't know what you don't have until something downstream blows up. So when I first read how Venetian merchants kept their books in the fifteenth century, I had a physical reaction. These people solved a version of my problem five hundred years ago with ink and paper.
In 1494 a Franciscan friar named Luca Pacioli published a description of the bookkeeping method Venetian merchants had already been using for a century or so. He didn't invent it. He wrote it down. And the detail that stuck with me is small and mechanical: when a bookkeeper carried an entry from the journal into the ledger, he drew a diagonal stroke on the journal line after posting the debit side, and a second stroke after posting the credit side. One stroke meant unfinished. Two meant done. Open the journal, scan for single strokes, and the incomplete work announced itself.
The larger check ran on the same principle. Every transaction went in twice, once as a debit and once as a credit. Total both columns and compare. If they don't match, something is wrong somewhere in the books — you haven't found it yet, but you know it exists, and you know it before anyone comes to inspect you.
This mattered because merchants had employees and agents handling money in other cities. You couldn't watch everyone, a constraint that hasn't improved much now that the agents are software and the cities are API endpoints. What you could do was hand them a format that made incomplete or inconsistent work visible on the page.
I want to be precise about what the balance actually proved, because it's less than people assume. A fabricated transaction entered consistently on both sides stays balanced. An omitted transaction whose two halves are both missing leaves nothing behind. HMRC's examination guidance warns that computerized double-entry records can be perfectly balanced and still wrong. The trial balance proves internal consistency, not truth. But internal consistency is a property of the format, and properties of the format scale the way human attention does not.
That approach survived in a few places. The U.S. Treasury still requires federal agencies to submit trial balances where debits must equal credits, then runs those figures against other reporting systems that recorded the same transactions independently. The self-balancing check from the fifteenth century, plus a second record to disagree with the first.
Most digital transaction systems went elsewhere. They kept the vocabulary of debits and credits and replaced the self-watching part with access controls, encryption, and audit policy. The record no longer catches its own discrepancies. A reviewer catches them, assuming the reviewer looks, looks in time, and hasn't been handed more records than a person can read.
The format stopped assuming a skeptic would show up. It started assuming a trusted system would keep bad entries out in the first place, which is a reasonable assumption right up until the trusted system is the one generating the errors. And once automated transactions run at volume, after-the-fact review becomes sampling. You check some of it. You infer the rest. Pacioli's bookkeeper checked none of it by hand either — he just glanced down a column of strokes and let the paper tell him where the hole was.