Scout — Discovery & Acquisition

Scout owns the unknown. It searches the open web for sources that are not yet in your brain, ranks them against your goals, and stages them as candidates for you to review. Scout never writes to the corpus — approved candidates are handed to the Librarian.

What Scout does

Scout runs a discovery loop:

flowchart LR
    POL[Set policy] --> RUN[Run Scout] --> Q[Ranked candidates]
    Q --> T[Target good ones] --> LIB[Librarian curation queue]
    Q --> R[Reject the rest]

It finds net-new sources only — anything already known to the brain is marked excluded so you are never shown duplicates.

The Scout dashboard

The header shows Discovery & acquisition with these actions:

• Run Scout — starts a discovery run (shows live Discovering… N%). (Requires the run_scout capability — Curator or Admin.)
• Add document — upload a file directly; it is staged as a candidate. (Requires workspace_upload.)
• Settings gear — per-agent configuration for Scout.

Below the header: stat tiles (Staged / Targeted → ingest / Rejected), the candidate queue on the left, and the policy summary on the right.

Configure the discovery policy

Click Edit policy (requires set_scout_policy) to open the policy drawer. It has:

Section	What you set
Goals	Free-text description of what this brain should learn. Scout ranks candidates against these.
Jurisdictions & Topics	Chip lists that focus the search.
Domains	Allow-list and block-list of source domains. Blocks win over allows; conflicts are flagged.
Discovery limits	A relevance threshold (0–1) and max results per run (1–50).

Find domains automatically

Inside the Domains section, Find domains from policy suggests candidate source domains (with sample titles and counts) derived from your goals via a live web search. Select the ones you want and Add them to the allow-list. If web search isn't configured, this step reports a clear error and you can add domains by hand.

Review candidates

Each row in the candidate queue shows a relevance score, the host, matched-goal chips, and a status pill. Use the Staged / Targeted / Rejected tabs to filter.

Click a candidate to open its detail sheet:

• Relevance score, matched goals, and the rationale for why it was surfaced.
• The source URL, the discovery query/seed, and a content snippet.
• Provenance — content hash and timestamp.

Actions on a candidate

(Require the target_candidate capability — Curator or Admin.)

Candidate state	Available actions
Staged	Target for ingestion or Reject.
Rejected	Re-target.
Targeted	Reject discovery (reverses the targeting).

The hand-off to the Librarian

Targeting is the bridge into the governed pipeline. When you target a candidate, Scout drops the document into the Librarian's curation queue, where it is extracted into a change set for human review — Scout itself makes no change to the corpus.

After a run you'll see a toast such as "N new candidates staged, M already known."

Role summary

Action	Minimum role
View candidates & policy	Reader
Run Scout, edit policy, target/reject	Curator

Read-only users see View policy and disabled actions rather than edit controls.