Recurring insight, with a stop switch.

Recurring Brain work needs bounds: what to watch, when to stop, how to pause, and where insight lands — without inventing a swarm daemon or burning product AI keys. The Persistent Research Kit is that discipline. It is not a daemon, not CP-0, and not a model runner.

Three layers of compounding insight

Enabling it

1. Confirm the product pool instance is pool-locked.
2. Copy RESEARCH-CHARTER.md and RUN-LOG.md from their templates into the pool.
3. Fill the charter: questions, allowed paths, do-not-touch, max iterations, session modes.
4. Admin: lock research charter — status moves to admin-locked.
5. Run the validate script (check-research-loop.sh — it never calls a model).
6. Admin: start research loop — status moves to active.
7. Paste the Brain prompt into a fresh agent session, with placeholders filled from the charter.
8. Review INBOX.md / PROPOSALS.md; promote only through the Admin promote recipe.

Pause and resume

Pause by creating a PAUSE file next to the charter. Resume by removing it, then running start research loop again. The validate script fails closed for as long as PAUSE exists.

Cost and grounding controls

• Max iterations per session is a hard stop baked into the Brain prompt.
• The charter carries soft time / token budget notes.
• Routine research runs on Cursor search tools only.
• Forbidden for this loop: OpenRouter and the product's agent-provider env vars.
• The validate script's strict-env check (on by default) fails if those look set.

External cron or CI may open an agent session that loads the Brain prompt once validation passes — Hive itself embeds no scheduler and no model caller.

Red lines

• No WORKSTREAMS or blueprint edits by the Brain.
• No feature PRs, worktrees, or CP-0 from this kit.
• The local project boundary and cross-hive Admin gate still apply in full.
• Everything written here is insight — until Admin promotes it.