How the Registry Learns
The registry grows two ways: an autonomous 24/7 learning loop and an operator-driven manual research mode. Both feed the same publish gate and the same review tray, so nothing reaches the public corpus unchecked.
Operator-gated. The autonomous loop ships dormant — it only runs when an operator enables it (
LEARNING_MODE), with a hard AI-budget cap. This page describes the capability; running it is an admin decision.
The autonomous loop
The loop turns the registry into a self-extending knowledge flywheel:
- Pick a topic. A
SeedTopicworklist combines a starter "genome" of topics with graph-derived gaps — the registry scans its own entity graph for weak spots (entities pointed at but never explored, or high-degree entities with thin sourcing) and seeds those, ranked by authority. - Research it. The loop runs the same research agent the platform uses, producing a schema-valid UKS v3.0 packet.
- Hold for review. Loop-produced packets publish private — they never go public automatically.
- Operator approves. An admin reviews the queue and approves (→ public) or edits/rejects.
Manual research mode
An operator can trigger a deep research run on demand, review the result, edit it, and publish — every edit recorded as a moderation decision.
The control plane
Both modes are driven from an admin surface. The relevant endpoints (all require an admin key):
| Endpoint | Purpose |
|---|---|
POST /api/registry/admin/learning-tick | Run one learning iteration (manual trigger) |
POST /api/registry/admin/gap-detection | Seed graph gaps into the worklist |
POST /api/registry/admin/pagerank | Recompute authority scores |
GET /api/registry/admin/review-queue | The hold-for-review tray |
POST /api/registry/admin/packets/:id/approve | Approve a held packet → public |
Guardrails
- Budget cap — a hard AI-spend ceiling per day; the loop stops when it's reached.
- Hold-for-review — autonomous output is private until a human approves it.
- The full publish gate — schema, secret-scan, clinical policy, and injection scan apply to loop output exactly as to manual publishes. See Publish & sign.
- Safety veto — approved packets are still subject to the continuous serve-time veto.
→ Related: Trust & safety · Publish & sign