WOLF AI field guide
Methodology for source-linked outputs.
A source-linked output carries enough provenance for a reviewer to return to the supporting document, page, and passage instead of accepting a generated statement on trust.
Published · Last reviewed
Written and maintained by the WOLF AI product team and checked against the public sources cited on this page. It has not been reviewed by outside counsel, and it is not legal advice.
What do the numbers say?
The NIST AI Risk Management Framework organizes trustworthy AI around functions it calls Govern, Map, Measure, and Manage, and treats traceability and human oversight as design requirements rather than optional additions. Source-linked output is the practical form that traceability takes in a litigation workflow.
Comment 8 to ABA Model Rule 1.1 states that maintaining competence requires a lawyer to keep abreast of "the benefits and risks associated with relevant technology." Adopting an AI-assisted workflow and verifying what it produces are treated as two halves of the same professional obligation, not as separate choices.
Source: ABA Model Rule 1.1: Competence (including technology)
A qui tam complaint is filed in camera and, under 31 U.S.C. § 3730(b)(2), "shall remain under seal for at least 60 days" and is not served on the defendant until the court so orders. Courts routinely extend that period, so the confidentiality obligation on the record set is measured in months or years, not weeks.
Who is this workflow for?
Anyone who has to answer the question "where did this come from" about an AI-assisted output — the reviewing associate, the partner signing off, and the firm evaluating whether the tool is usable on client work at all.
When it is the wrong tool. Provenance is a verification aid. It is not a correctness guarantee, and it does not shift professional responsibility from the reviewer to the software.
What documents and inputs do you need?
- Source documents with page boundaries preserved through ingestion.
- OCR output with per-page confidence retained rather than discarded.
- A record of reviewer decisions: edits, rejections, and approvals.
How does the workflow actually run?
- 1.Preserve source identity and page boundaries during ingestion.
- 2.Attach extracted statements and events to precise source locations.
- 3.Show the cited passage alongside editable generated work product.
- 4.Record attorney edits, rejection, and approval in the working process.
What does this look like in practice?
Illustrative scenario
Illustrative only. A generated chronology entry reads: "18 March — service billed, no corresponding clinical note."
- The entry carries two links: the claim line on page 412 of the billing export, and the note range searched on pages 88 to 94 of the clinical file.
- A reviewer opens both. The claim line is right. The note range is wrong — the note exists, in a different file that was collected later.
- The reviewer corrects the entry and records the rejection, so the same wrong inference is not regenerated.
- The citation was accurate the whole time. The inference drawn from it was not.
What this does not show. That gap — accurate citation, wrong conclusion — is the failure mode source linking is meant to make visible, not the one it prevents.
What do you get out, and who reviews it?
- Work product where each statement names the document, page, and passage behind it.
- A retained record of what a reviewer changed, rejected, or approved.
- The reviewer owns the interpretation. Provenance makes checking it cheap; it does not do the checking.
Where does this approach break down?
- A citation can be accurate while the interpretation is still wrong.
- OCR, pagination, and document quality can weaken provenance and require manual correction.
- Source linking supports verification; it does not transfer professional responsibility to software.
- Provenance says nothing about what was never ingested.
Common questions
Does source linking prevent hallucination?
It makes an unsupported statement cheap to catch, because a reviewer can open the cited passage in seconds. It does not prevent a wrong inference drawn from a real passage.
What happens when OCR quality is poor?
Low-confidence pages are marked rather than silently accepted, and they need manual checking against the original document.
Are reviewer decisions retained?
Edits, rejections, and approvals are recorded as part of the working process, so the review history stays with the work product.