01Where AI assists
Used carefully, models shorten the distance between a question and the material that might answer it. In our work that assistance is limited to discovery, organization, retrieval, and drafting support — always under analyst direction, never as a substitute for reading.
- Finding candidate sources and related documents
- Organizing large evidence sets
- Comparing terminology and themes
- Supporting translation, extraction, and classification
- Testing alternative explanations and counterarguments
02What AI never does
Some acts are not delegable, because they are the product.
- AI never makes the recommendation. A recommendation is a judgment, and judgment belongs to a person.
- AI never grades the credibility of evidence on its own. Source quality is weighed by an analyst who understands where the source sits and what it wants.
- AI never signs a judgment. Every conclusion we publish carries the name of the person prepared to defend it.
03Human responsibilities
- Verify material evidence
- Resolve conflicts and source quality concerns
- Understand domain and regulatory context
- Separate fact from interpretation
- Approve the final recommendation
04Human sign-off protocol
Every brief names the analysts who own it. Before anything is published, the lead analyst reads each material claim back against its source, checks that confidence levels reflect the evidence rather than the desired answer, and signs the recommendation. If a claim cannot be traced, it does not ship. Readers can see who is accountable because the accountability is printed on the work.
05Evidence provenance
Sources are dated, logged, and re-checkable. When model-assisted retrieval surfaces a candidate source, an analyst re-verifies it against the primary document before it is allowed to support a claim — a model can point, but only a verified source can carry weight. The provenance trail is kept so that any claim can be walked back to what was actually published, when, and by whom.
06Auditing the assistance
We periodically review where model output influenced a draft: which passages it touched, which sources it surfaced, and whether anything it contributed survived unexamined. Discrepancies are logged and corrected in the open, under the same discipline described in our corrections policy. The point of the audit is not to prove the tools are right; it is to catch the moment they are wrong before a reader does.
07Versioned record
When evidence moves, the record moves. Updates to a brief keep their history: what changed, when, and why. Accountability survives those updates — the analyst who signed a judgment remains answerable for its entire history, including the corrections. A changed conclusion is not a quiet edit; it is a visible event.
08What we should not do
Publish unsupported model output as fact, hide AI-generated analysis from reviewers, expose confidential data to unapproved systems, or let automated tools make final high-consequence recommendations without human accountability.