diff --git a/docs/runbook-migrations.md b/docs/runbook-migrations.md
new file mode 100644
index 0000000..9c1dcac
--- /dev/null
+++ b/docs/runbook-migrations.md
@@ -0,0 +1,65 @@
+# nsecbunkerd migration & DB-maintenance runbook
+
+Operational notes for applying schema migrations and ACL/pairing maintenance on
+deployed, **LNbits-connected** nsecbunkerd instances.
+
+## ⚠️ Never full-wipe `nsecbunker.db` on an LNbits-connected instance
+
+The nsecbunkerd ↔ LNbits pairing is **split across both systems**:
+
+- **Bunker** (`nsecbunker.db`): per-account `KeyUser` binding (keyed by LNbits's
+  stable client pubkey) + redeemed `Token` + a shared `Policy`.
+- **LNbits** (`accounts.signer_config`, `RemoteBunkerSigner`):
+  `{token, client_nsec, policy_id}`.
+
+`RemoteBunkerSigner.sign_event()` signs **directly with the stored `client_nsec`**
+— it does NOT re-connect/re-redeem, and there is **no auto-repair** on restart or
+sign-failure. `provision()` runs only at new-account creation and mints a NEW
+npub (which changes the user's nostr identity).
+
+**Consequence of a full `nsecbunker.db` wipe:** the `KeyUser` bindings are
+deleted → every LNbits account's stored config dangles → all signing fails, and
+the only standard "repair" (`provision()`) changes identities. **Do not do it.**
+
+### Correct way to strip the #24 materialized photocopies
+
+Post-#27, token grants are evaluated live via the ACL step-4 `Token → Policy →
+PolicyRule` join, so the old materialized `SigningCondition` rows are redundant
+— and, written with `expiresAt = NULL`, they would keep granting past a token's
+expiry (silently re-opening #24 for already-paired clients). Strip them with a
+**targeted delete** that preserves the pairing:
+
+```sql
+-- Verify first.
+SELECT COUNT(*) FROM Token WHERE redeemedAt IS NOT NULL;  -- bindings to preserve
+SELECT COUNT(*) FROM SigningCondition;                    -- photocopies to strip
+
+-- Keeps KeyUser + Token + Policy intact. Live-token clients keep working
+-- untouched; only the stale photocopies are removed.
+DELETE FROM SigningCondition;
+```
+
+Run against each instance's `nsecbunker.db`. If an instance was already
+full-wiped, recover by restoring the pre-wipe `nsecbunker.db` backup, then run
+the targeted delete.
+
+> Manual-override grants (`add_signing_condition`, web-approval) also live in
+> `SigningCondition`. On an LNbits-only bunker there typically are none, so a
+> blanket `DELETE FROM SigningCondition` is safe. If an instance uses manual
+> overrides, delete only the policy-derived rows you intend to strip.
+
+## Keys are never in the DB
+
+Key material lives in `nsecbunker.json` (`keys`), never in `nsecbunker.db`. A DB
+wipe loses ACL/pairing state, never keys. LNbits holds the bunker **admin nsec**
+(`LNBITS_NSEC_BUNKER_ADMIN_NSEC`) and is the sole admin client.
+
+## Schema migrations
+
+Migrations are applied by the deploy's `prisma migrate deploy`, **not** by the
+daemon on boot — the in-`start.js` `npm run prisma:migrate` step is a no-op
+(tracked in #31). After adding a migration, make sure the deploy applies it.
+
+Prisma on NixOS needs the engine env pinned to `prisma-engines_6` (the bare
+`prisma-engines` attr is now 7.x with no `libquery_engine.node`; devShell fix
+tracked in #30). The deploy's `package.nix` already pins `_6`.