How We Resolved an ERP Inventory Oversell After DeepSeek’s Price Hike Using Seed Evolving

Incident Overview

On 2026‑08‑01 an order was closed, but a payment‑success callback arrived after closure, causing three items to be oversold. The correct metric is the fulfillment stock gap:

fulfillment_stock_gap = paid_pending_fulfillment_quantity - actual_available_stock

Timeline

2026‑08‑01 14:04:46 – Order created for 1 item.

2026‑08‑01 14:04:46 – Inventory reserved (3 base units).

2026‑08‑01 14:19:37 – Payment record created.

2026‑08‑01 14:19:46 – Order automatically closed after ~15 min.

After 14:19:46 – Close task released the 3 reserved units.

2026‑08‑01 14:21:05 – Payment success callback (≈1 min 19 s after close).

2026‑08‑01 14:21:05 – Closed order status re‑activated to paid.

2026‑08‑01 20:47:52 – Order shipped (≈6 h 26 min after payment).

2026‑08‑01 21:03:00 – Order status updated to shipped (≈15 min 8 s after shipment).

Root Causes

Pre‑reservation on order creation, no deduction on payment. The design assumes the close and payment paths are mutually exclusive.

CloseOrder releases stock but does not close the payment channel. The payment window remains open after the order is marked closed.

Payment success logic lacks order‑status validation. ChangeOrderSucceeded only checks that the payment record is “unpaid” and updates the order to “paid” even if the order is already closed, creating an illegal transition Closed → Paid.

No unified concurrency control. Close and payment flows read and write independently without row locks or conditional updates, so the final state depends on execution order, which can be disturbed by queue delays or retries.

Repair Directions

Prohibit Illegal State Transitions

unpaid → paid   ✅ legal
unpaid → closed ✅ legal
closed → paid   ❌ prohibited

Late payments should trigger an “abnormal payment pending” branch, record the event, raise alerts, and initiate a refund instead of silently converting the order to paid.

Conditional Update on Payment Success

-- Before (pseudo‑SQL)
UPDATE orders SET status='paid' WHERE id=?;

-- After
UPDATE orders SET status='paid'
WHERE id=?
  AND status='unpaid'
  AND off_time IS NULL;

If the condition fails, the process must enter a conflict branch and avoid further ERP synchronization.

Inventory Reservation Flow

reserved   → pre‑reserved, awaiting payment
released   → released (triggered by close)
consumed   → consumed (triggered by payment)

Close can only move reserved to released; payment can only move reserved to consumed. If a record is already released, the payment flow must reject the operation, preserving idempotency.

Align Payment Deadline with Close Time

The payment cutoff is now calculated from order creation time, ending one minute before the 15‑minute automatic close, reducing the window for late payments.

Alert Metrics

pay_time > off_time order count
Closed orders receiving payment success events
Order status conditional‑update failure count
Duplicate inventory release count per order item
Payments succeeding without a valid reservation

Integrating these metrics into monitoring catches similar issues early.

Historical Data Investigation

Before applying any compensation, run a read‑only scan to list all “closed‑then‑paid” candidates:

SELECT order_id, off_time, pay_time,
       TIMESTAMPDIFF(SECOND, off_time, pay_time) AS delay_seconds,
       sku_id, quantity, status
FROM orders
WHERE off_time IS NOT NULL
  AND pay_time IS NOT NULL
  AND pay_time > off_time
ORDER BY pay_time DESC;

Each result is manually classified into four handling paths (fulfilled, pending, refunded, refunding). Write operations require explicit confirmation and audit logging.

Takeaways

AI excels at global structural analysis, quickly correlating logic across multiple files.

Business judgment (e.g., whether to refund a fulfilled order) remains a human decision.

Long‑context models are most effective when focused on the relevant core files rather than the entire repository.

Pricing changes can affect not only cost but also workflow scheduling; selecting the appropriate tool for the scenario is essential.

References

DeepSeek pricing announcement: https://www.volcengine.com/docs/82379/2549861

Seed Evolving model documentation: https://www.aitop100.cn/infomation/details/34326.html