AI Provenance

AI You Can Put Your Name To

Legal work carries a person's name, so the record has to hold. Every ClauseMinds draft carries a deterministic provenance report — what the AI produced, what it grounded on, what a human changed. Authorship is attributed clause by clause, and a lawyer's sign-off is bound to the exact wording attested: edit the clause and the sign-off lapses.

Not a confidence score. Not a disclaimer. A record you can hand to whoever asks who wrote what.

Deterministic provenance report on every draft
Authorship attributed clause by clause
Sign-off bound to the exact wording
The provenance report

What the AI produced, what it grounded on, what a human changed

Every draft carries a provenance report built deterministically from the draft's actual history — not a summary written after the fact by another model. It shows which parts the AI produced and the sources each part was grounded on, which parts a human wrote, and which parts a human changed. When someone asks who wrote this, the answer is a record, not a reconstruction.

  • Deterministic: derived from the record of what happened, not generated about it
  • Shows what the AI produced, and what each part was grounded on
  • Shows what a human wrote and what a human changed, and where
  • One report per draft, available whenever the question comes up
Draft provenanceDeterministic report

Authorship by clause

2. Term

AI-drafted — grounded on your template

5. Non-solicitation

AI-drafted, then edited by you

8. Governing law

Written by you

Grounding

Every AI-drafted clause lists the template or library clause it was grounded on. Nothing in the draft is freeform generation.

Per-clause sign-off

Attestation binds to the wording. Edit the clause and it lapses.

Authorship is tracked per clause, so editing one clause reattributes only that clause — the rest of the record stands. When a lawyer signs off, the attestation binds to the exact wording attested. Change the wording and the sign-off lapses; it never carries over silently. And attestation is asserted by a named lawyer as an explicit act — it is never inferred from who touched the file or what role they held.

  • Authorship per clause — editing one clause reattributes only that clause
  • A lawyer's sign-off binds to the exact wording attested
  • Change the wording and the attestation lapses, visibly
  • Attestation is asserted by a named lawyer, never inferred
Clause attestationsWording-bound

4. Limitation of liability

Attested

Signed off by M. Alvarez. Wording unchanged since sign-off.

5. Indemnity

Attestation lapsed

Clause edited after sign-off. Needs a fresh review.

6. Confidentiality

Not attested

No lawyer has signed off on this clause.

Trust, as architecture

The same standard everywhere the AI speaks

Provenance is not a report bolted on at the end. Every AI surface in ClauseMinds carries the same invariants: show the source, record the decision, and never let silence pass as an answer.

Every finding quotes its source

Every extraction finding quotes the clause it came from, so you can verify it against the actual contract language before relying on it.

Silence is a claim

“Nothing found” is an explicit, reviewable claim — we checked for it and found nothing — not an ambiguous absence you have to interpret.

Full audit trail

Every accept, edit, and auto-accept is recorded — who, what, when. The history of a finding is part of the finding.

Grounded or gone

An AI read that cannot quote its source is dropped, not shown. The model never gets to be confidently sourceless.

Asserted, never inferred

A lawyer's attestation exists only because a named lawyer made it. It is never derived from roles, activity, or who happened to save the file.

Measured, not promised

Extraction accuracy is measured on adversarial held-out corpora with CI-enforced floors. The numbers, and their limits, are on the accuracy page.

Why it matters

The point of AI you can verify is to stop re-verifying everything

The hard part of using AI in legal work is not getting an answer — it is knowing what you would have to re-check before putting your name near it. ClauseMinds is built so the record answers that question: what was machine-written, what it stood on, what a person changed, and what a named lawyer actually attested to. Where a boundary exists, the product says so instead of rounding up.

  • Tell at a glance, per clause, what carries a sign-off and what needs your review
  • The provenance report answers “who wrote this?” without reconstructing history
  • A lapsed attestation is visible, never silently carried forward
  • Accuracy is measured on held-out corpora and published on the accuracy page
Draft summary12 clauses

AI-drafted, grounded on your templates

8 clauses

Edited or written by a human

4 clauses

Attested by a named lawyer

3 clauses

Attestation lapsed after an edit

1 clause

What this buys you

You review the 1 lapsed clause and the 8 you have not signed off — not the whole document from scratch, again.

The right to stop double-checking

A deterministic record of who wrote what, grounded on what, attested by whom — per clause, per draft. That is what it takes to put your name on AI-assisted work.

Provenance travels with the drafting rail — see Drafting & Negotiation. ClauseMinds is not a law firm and does not provide legal advice.