The most under-discussed failure mode of legal AI is not the wrong answer. It is the unanswerable one — a fluent paragraph whose sources, retrieval path and reasoning cannot be reconstructed later, when a partner, a client or a court asks how it came to exist. We built the evidence ledger because we do not want to ship tools that produce paragraphs like that.

What gets recorded

Every source the system read, tied to the matter it belongs to.
The passage spans behind each finding — a citation resolves to a location, not a vibe.
What the system proposed: the revision, the summary, the answer.
What counsel decided — accepted, edited, rejected — with the reviewer and the timestamp.

The ledger persists locally with the matter and exports with it. Deleting a matter removes its ledger; there is no server-side copy to outlive your decision. The interactive demonstration on this site shows the full shape using prepared fictional files, so you can see the record before trusting the idea.

Why provenance is not correctness

A citation shows where an answer came from. It is not a guarantee the answer is correct.

This distinction is the ledger's honest limit, and we state it in the product, not just in marketing. A ledger demonstrates the provenance and history of a conclusion; it cannot establish that the conclusion is legally sound. That judgement stays with counsel — which is why the system proposes at a checkpoint and never applies a change without a human decision. The record makes review possible; the reviewer makes it meaningful.

When a finding's citation cannot be resolved, the system reports not found rather than inventing a source. An answer without a resolvable span is not asserted as fact. That behaviour costs us a little fluency and buys the only thing an evidence tool can sell: the ability to check it.