diff --git a/SCRATCH/2026-02-07_customer-id-schema.md b/SCRATCH/2026-02-07_customer-id-schema.md
new file mode 100644
index 0000000..f9f618e
--- /dev/null
+++ b/SCRATCH/2026-02-07_customer-id-schema.md
@@ -0,0 +1,110 @@
+# 2026-02-07 — Customer ID schema + "hashed id" confusion
+
+## Current reality
+
+### User table
+
+Prisma schema:
+
+```prisma
+model User {
+  id           String   @id @default(cuid())
+  email        String   @unique
+  username     String   @unique
+  passwordHash String
+
+  firstName    String?
+  lastName     String?
+  displayName  String?
+
+  createdAt    DateTime @default(now())
+  updatedAt    DateTime @updatedAt
+}
+```
+
+* When a user registers via the API (`POST /api/auth/register`), Prisma assigns a **CUID** (looks "hashed" in Prisma Studio, e.g. `cmk07b7n4000q...`).
+* That value is **not derived from the JWT token**.
+* During testing you manually created a user with an ID like `u000` (and other prefixed IDs), which is why you're seeing mixed formats.
+
+### ContainerInstance.customerId (what broke)
+
+You observed `ContainerInstance.customerId` being **polluted**:
+
+* sometimes `u000` (manual)
+* sometimes `u-dev-001` (hardcoded during testing/dev)
+* sometimes a Prisma CUID (new registrations)
+
+Root cause (confirmed):
+
+* Frontend / provisioning calls were hardcoded to use `u-dev-001` as `customerId`.
+* New registrations correctly used Prisma's default `User.id` (CUID), but container creation wasn't consistently using it.
+
+## Short-term decision (now)
+
+✅ **Use `User.id` as the authoritative customer identifier** everywhere.
+
+That means:
+
+* `customerId` in `ContainerInstance` should always be set to `req.user.id`.
+* Stop inventing or hardcoding IDs (`u-dev-001`, etc.) in the frontend.
+* Keep `username` / `displayName` for human-friendly display and troubleshooting.
+
+This keeps the system consistent immediately and avoids adding schema complexity right now.
+
+## Why this is OK
+
+Pros:
+
+* Guaranteed unique
+* Harder to guess (good for security)
+* Already produced by Prisma automatically
+* No more "what format is customerId?" problems
+
+Cons:
+
+* Not pretty/human-friendly in logs (CUIDs)
+
+Mitigation:
+
+* When displaying in UI or logs, show a composite:
+
+  * `displayName (username)` and optionally the first 6–8 chars of the id
+  * Example: `Jester (jester) • cmk07b7n4`
+
+## Long-term (optional) improvement
+
+If you later want a human-readable / sequential ID:
+
+Add a second field, e.g.:
+
+* `publicCustomerId` (`u001`, `u002`, etc.) generated by a counter table or dedicated allocator
+
+Then:
+
+* Keep `id` as the internal primary key (CUID)
+* Use `publicCustomerId` only for:
+
+  * display
+  * support tickets
+  * invoices / Stripe customer metadata
+  * external references
+
+But: don't block on this now.
+
+## Cleanup / migration note
+
+If you want to clean existing polluted rows later:
+
+* Update all `ContainerInstance.customerId = 'u-dev-001'` to the correct `User.id` (based on ownership mapping / audit trail / known associations).
+* If no reliable mapping exists, leave historic test rows as-is and only enforce the rule going forward.
+
+## Implementation rule (what to enforce)
+
+* Provisioning endpoint should **ignore any client-supplied customerId**.
+* Always derive from auth context:
+
+```js
+const customerId = req.user.id;
+```
+
+* If you need admin tooling to create for others, that should be a separate privileged endpoint that explicitly sets ownership (and logs it).