When a regulator asks how a firm used AI on a matter, most firms are in trouble before they open their mouth. Not because they did anything wrong — because they cannot prove they did anything right. The record is scattered: drafts in a version history, a fee-earner's notes in email, the AI tool's own logs in a separate vendor portal, if those logs still exist. Reconstructing what happened is an afternoon's work and, when it is done, it is still just an assertion. "This is what we think happened." An open question, not evidence.

The signed reasoning ledger closes that question.

What the ledger records

Every time the assistant produces something on a matter — a draft, a fact extraction, a critique, a citation — an entry goes to the ledger. And under it, every human decision: accepted, rejected, edited. The result is one readable record of how the matter was worked alongside AI. What the firm asked. What came back. What a human signed off, and why.

Note carefully what the entry contains, because this is where privilege and evidence had to be reconciled. The input is recorded as a one-way fingerprint — a hash, never the prompt text in the clear. The output is stored as an excerpt, alongside the call metadata: model, tokens, skill, timestamp. The ledger proves what happened and proves it has not been altered. It is not a verbatim transcript of privileged reasoning sitting in an exportable file waiting to leak. We were not willing to build a tamper-proof record that was itself a privilege hazard.

Why "signed" is the load-bearing word

A log you can quietly edit is not evidence — it is a claim with extra steps. So each entry is HMAC-signed under a per-firm secret, and the entries are hash-chained: each links to the one before it, so any tampering breaks the chain and shows. At the close of each UTC day, the chain's position is sealed under the firm key. That gives an auditor one line to check per day instead of a walk through every row, and it fixes that day's record: rewrite a sealed day afterwards and the seal stops matching, even if the hash chain itself is carefully repaired.

When the moment comes, you export the whole thing in one click — a readable audit pack, in markdown or CSV, verifiable end to end. You are not handing the regulator a story you have assembled. You are handing them a record they can independently check.

Honest about who this is for, and what is live

The ledger runs today, on every plan — append-only, HMAC-signed, daily-sealed, exportable in one click. There is no tier gate on it and we are not adding one. A compliance record you can only afford at the top of the price list is not much of a compliance record.

What is not built is per-office cryptographic isolation. A multi-office firm today keeps one ledger for the firm, not one sealed ledger per office. This page previously described that isolation as shipped, which is precisely the kind of claim this ledger exists to make impossible — so: it is on the roadmap, and it is not in the product.

The difference this makes

General-purpose tools keep session logs. That is the gap in one sentence: a session log lets you say "we have logs"; a signed, chained, exportable ledger lets you say "we have evidence the regulator can rely on". One is a shrug; the other is a document.

That distinction is exactly the kind of thing you only build if you have sat on the wrong side of a regulatory review and understood what "prove it" actually demands. Built by solicitors, for solicitors — so that the question of how AI was used on a matter is answered with evidence, and never left open.

← All posts