diff --git a/CLAUDE.md b/CLAUDE.md
index d6a6cea..1af8267 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -685,6 +685,44 @@ in a review table. Consents expire annually, so expect it.
 
 Undo an import: `DELETE FROM transactions WHERE source = '<source>'`.
 
+### Zip, and the three things a BNPL account needs (2026-08-15)
+
+24 statements, Dec 2020 → Jun 2026, all balance-asserting at $0.00. Zip is a
+credit account, so importing it changes how three existing row types must read.
+
+**1. A repayment leaving the bank is a duplicate ONLY where a Zip statement
+covers the purchases.** Zip's own `Payment` rows are already `transfers`. On the
+bank side, `ZIPPAY*` debits are the same money — but only 2 of 8 fall inside a
+Zip statement period. Those two are now `transfers`; **the other six must stay as
+spend**, because for those periods the bank debit is the only record the purchase
+ever happened. Moving all eight would have deleted $703.35 of real spend. Same
+statement-coverage rule as the CSV importer, and the same trap.
+
+**2. `Account Credit` is cashback, and Zip labels its three money-in types
+differently.** `Payment` = repayment from the bank. `Refund <merchant>` = a
+genuine merchant refund. `Account Credit` = Zip cashback — 34 rows, $740, round
+amounts ($20 ×20, $30 ×10, $10 ×4), with no matching debit from any bank account.
+
+Gemini filed the identical description two ways — 23 as `other` (which *subtracts*
+from spend) and 11 as `transfers` (excluded entirely) — so one thing was doing two
+opposite things. All 34 are now `shopping` + `credit`, **netting against spend**:
+cashback lowers what a purchase actually cost, which is how `NET_SPEND_ROWS` /
+`SPEND_SIGNED` already treat a refund. Filing it as `income` was rejected for the
+reason the Up item sales note gives — it flatters `net = income − spent` and mixes
+non-earnings into the salary line. Ties to the **Cashback tracking (ING-6)** epic;
+a dedicated `cashback` category would need the CHECK constraint, the TS mirror,
+and a decision in every analytics fragment, so it was not done here.
+
+**3. A gift card stays spend.** Prezzee is $1,580 of the history. Stored value is
+tempting to call `transfers`, but that only works if every downstream purchase is
+logged — it is not, for 2020. Identical reasoning to the ATM-withdrawal rule.
+
+All 137 Zip rows are categorised; net spend is $11,135.11 against $10,085.11 of
+`transfers`. **Statement 200 (Sept 2021) came through incomplete** — no
+`billing_end_date`, no balances, 2 rows — so it is the one statement not covered
+by `uq_statement_identity` (the index is partial on `billing_end_date IS NOT
+NULL`) and a re-import would duplicate it. Worth re-processing paperless doc 296.
+
 ### Trusting extracted statement data
 
 **Balance assertions are the check that works.** `getStatements` computes