Argus

live

A JD Edwards style data warehouse, broken on purpose, with SQL that finds the damage and a score for how well it does.

Argus is a PostgreSQL data warehouse that stores data the way JD Edwards EnterpriseOne does, loaded with 18 months of simulated order-to-cash and procure-to-pay activity. I broke the data on purpose in 10 ways ERP shops actually see, wrote SQL to find the damage, and scored that SQL against a record of exactly what I broke. A dashboard shows the findings in JDE or SAP terms.

argus.demo/reconciliation

loading demo…

Key results

Ten defect classes injected with a logged ground truth, and a SQL detection library scored on precision and recall against it. All data is synthetic.

0/1,077

Findings that were false positives

889/889

Core defects found. An upper bound, since I wrote both the injections and the queries

188/207

Holdout defects found (90.8%), injected after the queries were frozen

0.67 s

Full query library, down from 3.2 s with identical output

Naive checks catch 27% to 85% of the defects the library catches at 100%.
Six of the 10 defect classes have 100% recall. The other four are documented with every miss and what a fix would cost.
The query freeze is verifiable: the queries were committed before the holdout set existed, and the diff against the scored commit is empty.
Slowest query rewritten from 1,699 ms to 67 ms, and the G/L is partitioned by fiscal period.
CI rebuilds everything from an empty database on every push and fails if a score changes.