The promise you are making
When you give an agent a knowledge base, you are telling it "act on what you find here." That promise is only safe if a person looked at every fact before the agent did. The distiller is the part of Ship that holds the line. Imported sources are ingested by your workspace's per-workspace Lighthouse engine into the workspace corpus; the distiller's role is the human-review gate on synthesised articles before they become agent-readable knowledge. It takes raw text — a docs repo file, a web page that changed overnight, an upload, a note routed in from a chat — and turns it into a draft article, then stops. A human opens the draft, reads it against the source, and decides what to do.
Nothing becomes knowledge silently. That is the entire design.
How a draft is born
A source refreshes. Something changed. The distiller reads the new text and does three small jobs.
It routes the topic. Most documents are about one thing — the deploy procedure, the lint rules, the customer-tier ladder. The synthesiser reads the document, names what it is about, and routes it against the buckets you already have. Routing is rule-based today, so it leans on the source and bucket you already configured rather than inventing a new bucket on its own.
It writes a short article. A clear title. The fact as the source states it. An example only where the example removes ambiguity. A provenance footer that links back to the source so any reader, human or agent, can walk the pointer.
It stops. The draft is private. No agent sees it. No citation can land on it. It waits as a review item until you, or a teammate, decides what to do.
The two review moves
When a draft is ready you have two moves: accept it (the draft becomes the live article and the next agent reading the bucket picks it up — an update accepts as a new version that supersedes the prior one automatically) or dismiss it (the draft is archived — kept readable, removed from agent search).
Accept is the everyday move — the article looks right, you adjusted a word here and there, you accept it and citations begin to accumulate. Dismiss is for the draft that should not become knowledge: the source did not give the distiller enough to work with, the fact already changed, or the note was noise. The dismissed draft is archived rather than deleted, so the history stays readable; the archive is your memory of decisions, not a wastebasket.
Superseding is not a separate move you make. When you accept a draft that updates an article you already have, the new version supersedes the prior one automatically and the old version stays readable so the history is clean — anyone who needs to know why we changed our mind can read both.
Where drafts surface
A draft surfaces as a review item to dispose, one per draft. There is no separate filterable queue page to scope by bucket, source, or date — disposing a draft is the accept-or-dismiss decision above. Most teams find their knowledge base healthiest when someone spends ten to fifteen minutes a day on whatever drafts came in — long enough to clear what arrived overnight, short enough to fit between meetings.
The flow of drafts tells you something honest about your setup that no dashboard would. If drafts pile up, one of three things is true. Your sources are too aggressive — you are importing noise that does not belong. Or your buckets are too broad because reviewers cannot tell what goes where. Or your team is routing chat notes that should have been answered and forgotten instead of becoming articles. A calm flow means your knowledge base is in tune. A backed-up one means a tuning conversation is overdue.
If no drafts ever appear, you are probably not importing enough — or your team is answering many questions that nobody is writing down.
Why this shape
Other systems will offer to "auto-publish high-confidence imports." We do not. Auto-publish is how a vendor doc page change at 2 a.m. becomes Tuesday morning's wrong agent behaviour — with no audit trail of who decided the change was an upgrade. A reviewer who spends thirty seconds on a draft has done more for trust than any confidence score the model could attach to its own work.
The review path is the cost of trusting your agents with knowledge. The cost is small. Ten minutes a day, divided across the team. What you buy with it is the line between "we collected some text" and "agents now act on it." That line is worth holding.