Convene

Four steps, in order. Skipping step 1's isolation destroys the point of the whole exercise.

0. Name the criteria FIRST

Before dispatching anything, write down the criteria the drafts will be judged on, and put them in the brief. Criteria invented after the drafts arrive get reverse-engineered to favor whichever draft you already liked.

1. Blind draft

Dispatch each member with sys_session_send as a separate session carrying the same brief. They must not receive each other's drafts, summaries of each other's drafts, or hints about who else is drafting.

This isolation is the anti-deference mechanism. It is structural: a member cannot defer to a stronger peer it cannot see. Members will infer each other's identity from writing style eventually — that is fine, because by then the drafts are already committed.

Titles are continuation keys — make them unique

sys_session_send does spawn-or-continue keyed on title. Reusing a title continues that conversation rather than starting a new one. So every dispatch gets a title unique to that decision and that member:

DR-0002-draft-sm-a      DR-0002-critique-sm-a
DR-0002-draft-sm-b      DR-0002-critique-sm-b

Never a bare task name like write-state-model shared across members, and never a title already used in this run for anything else. Prefer sys_session_create when you need a guaranteed-fresh session.

Validate that each draft answers the brief

A returned session is not proof the brief arrived. Put a short echo token in the brief — the DR id plus the number of named criteria — and require each draft to open with it:

DR-0002 / 5 criteria

Check the echo before counting anything as a draft:

• Echo matches → a real draft.
• Echo missing or wrong → the dispatch was misbound. That is not a draft and not an agent error. Discard it, re-dispatch on a fresh session with a new unique title, and record that it happened.
• Misbinds twice for the same member → halt. The session layer is broken and continuing will silently shrink the council below its threshold.

This guard is deliberately mechanism-independent. It catches a misbound brief whatever the cause — title collision, duplicated conversation, replayed completion signal — because it tests the only thing that matters: did this member answer this question?

Collect with sys_read_inbox. Never busy-poll.

Failure mode to avoid: helpfully passing draft A into B's context to "save a round trip." That converts three independent positions into one position with two endorsements.

Write each draft to disk the moment it arrives

Do not batch the writes until the end. Write draft-a.md when A returns, before collecting anything else. An interrupted convene that wrote nothing loses every draft it paid for — a session transcript is not durable storage, and it is the layer most likely to have failed in the first place.

2. Reveal and critique

Reveal all drafts under neutral labels — SM-A / SM-B / SM-C, UX-A / UX-B. Never name models, at any point, in any artifact.

Each member critiques every draft including its own, against the named criteria and only those. A critique names what breaks, under what input, with what consequence. "I prefer B" is a vote, not a critique.

3. Converge

Apply merge-guard. Pick one draft as-is, or merge only if a criterion can be named on which the merged idea beats every input.

Two full rounds without convergence is a deadlock — run escalate. Do not run a third round hoping it resolves; it will produce the average you were trying to avoid.

4. Record

Emit a Decision Record with decision-record, and write the convene's artifacts to .agentic/convenes/<id>/:

.agentic/convenes/<id>/
  brief.md          the task and the criteria, as sent
  draft-a.md  draft-b.md  draft-c.md
  critiques.md
  decision.md       the Decision Record

The record is the durable output. Six months from now the code will have moved and the reasoning will not be recoverable from it. If a convene produced no record, it produced nothing.