Evidence Ledger

An Evidence Ledger is a register of evidence. The English word “ledger” means an account book, that is, a list of entries in chronological order. “Evidence” means proof or documentation. Together, the term describes a kind of logbook that records what a statement or a decision is based on. What matters most is that old entries must never be overwritten: corrections are added as new lines, while the old line remains readable. This makes it possible, even months later, to trace what was known at what point in time.

When an answer is not enough, and proof is needed

Computer programs based on machine learning deliver results without a spelled-out chain of reasoning. Such a program might say that a loan application is risky, but not clearly explain why. As long as it’s about movie recommendations, that’s unproblematic. As soon as money, health, or rights are involved, it becomes a problem. Anyone who wants to challenge a decision needs something verifiable in hand.

This is precisely the gap an Evidence Ledger is meant to close. It does not store the answer itself, but its basis. Typical entries include: which documents were read, which version of the program was running, who approved the result. This shifts the question from “Is the result correct?” to “Was the result produced in a verifiable way?”. The second question can be checked, the first often cannot.

One practical side effect concerns disputes. Banks, insurers, and hospitals must be able to substantiate decisions for years. Without a ledger, all that remains in the end is the employee’s word against the customer’s word. With a ledger, a third, independent trail exists.

What an entry contains

A single entry usually consists of four parts. First, a timestamp, meaning date and time accurate to the second. Second, the trigger, such as a user question or a submitted application. Third, the evidence used, often as a reference to a specific document along with a page number. Fourth, the result and the name of the system or person that generated it.

To ensure that no one secretly alters old entries, checksums are typically used. A checksum is a short sequence of numbers and letters calculated from the content of an entry. Changing even a single character in the content produces a completely different checksum. If each entry additionally contains the checksum of its predecessor, a chain is formed. If someone tampers with an old entry, all subsequent checksums no longer match.

This principle is familiar from the blockchain, the method behind cryptocurrencies. An Evidence Ledger, however, does not need a global network for this. It usually runs as a closed database within a company. It’s important to distinguish it from an ordinary log: a log records technical events for developers, while a ledger collects evidence for auditors and courts. A common misconception is that a ledger makes the AI honest. It only documents what happened, and does not protect against the evidence itself being false.

Where the term appears in products and news

In the news, the term is mainly encountered in connection with AI regulation. The European AI Act requires automatic recording of processes for high-risk applications. Enterprise software vendors sell exactly these building blocks and call them Evidence Ledger, Audit Trail, or Compliance Log. For banks and medical technology, this has since become its own market.

The idea also comes up with chatbots, even without hearing the name. When an assistant provides its answer along with source links, that is the visible tip of such a register. In the background, the system keeps track of which passages of text it actually read. During audits, this makes it noticeable when an answer had no source at all.

For investors and observers, the topic is relevant because verifiability incurs costs. Storage space, computing time, and review processes are part of it. Companies that maintain these records properly find it easier to enter heavily regulated industries. Companies without such structures often remain limited to non-critical applications.