ELIPS is embedded — production is your binary, on a server, with disks you trust.

What the layered stack looks like when an agent runtime owns its own retrieval.

Durability: paranoid on durable storage, relaxed with scheduled checkpoints on ephemeral disks
Replicas: open with access_mode='read_only' from N readers; shared flock coordinates
Checkpoint: schedule elips checkpoint via cron/systemd-timer; tune cadence to recovery time
Backups: snapshot the directory while it is closed, or while only read-only openers hold it
Telemetry: emit explain().strategy and candidate_count per query
Compaction: run elips compact after large delete bursts to reclaim tombstones

You've now seen every concept in the engine. Each one has its own page in /docs and its own chapter here. The roadmap (docs/roadmap.md) is what is coming next.