REAL DATABASE RUN · 2026-09-25

10 Million Products.
20 Codex Agents.

Twenty independent Codex tasks authored catalog-repair SQL. We reviewed and executed those plans on twenty isolated branches of a real 10-million-row MatrixOne table, then selected approved rows and merged the approval branch into main.

The video is a 60-second replay reconstructed from the completed run, with time compressed. The UI is a walkthrough of the script-driven experiment. The public SQL Playground is a separate, smaller tutorial.

How we ran the experiment → · 阅读完整中文实录 →

Actual results

What actually ran

• Synthetic catalog: 20% missing attributes, 5% wrong categories, 100,000 duplicate SKUs, one million inconsistent brands, and one million poor descriptions.
• Five repair types × four ID shards; each branch contains the entire logical 10M-row baseline. Codex tasks ran in waves of up to three. Database update workers were also capped at three.
• Repairs copy factual supplier fields. SKU repair restores the canonical identifier for distinct product IDs; it does not delete products.
• MatrixOne native diff counts were checked against SQL counts. Both quality counts and native diff confirmed that main stayed identical to the baseline before merge.
• Application policy holds source confidence below 0.80 and excludes price changes. Fourteen price edits were deliberately injected by the test harness after agent execution; they were not spontaneous Codex mistakes.
• Native DATA BRANCH PICK ... KEYS (SELECT ...) copies approved rows into an approval branch. Native DATA BRANCH MERGE ... WHEN CONFLICT FAIL then applies that branch to main.

Observed timing

On this single 8.0.30-MatrixOne-v4.2.1 run, the 20 sequential approval PICK operations took 643.39 seconds in total. The final merge took 840.81 seconds. These include client-observed query time; they are not generalized performance claims. The video compresses these waits.

Scope and boundaries

The fixture gives each task a disjoint set of affected primary keys. We separately verified that a conflicting merge is rejected and that selecting a safe row leaves an unselected price change out of main. The approval workflow and business checks are application code around the native database operations. Some Codex tasks declined low-confidence proposals upfront; those unchanged rows are additional to the reported proposed-review queue.

The small capability probe also observed that re-merging the original source after a prior PICK can raise a conflict in this instance. This measured workflow avoids that path: it assembles one approval branch and merges it once. It does not claim arbitrary retry or conflict-free concurrent-merge behavior.

Evidence and reproduction

Download measured results (JSON) Read the reproduction scripts

Try the live SQL Playground →

中文摘要

本次使用 1,000 万行合成商品数据和 20 个独立 Codex 任务，真实执行分支、差异、按行选择和合并。共提出 4,585,827 行修改，合并 4,552,565 行；33,248 行提案保留待审，14 条主动注入的价格违规被排除。主表价格改动和未经审查的改动均为 0。视频为实测结果的压缩回放，应用层负责业务校验与审批，数据库原生执行 PICK 和 MERGE。