eliteAI / Evidence / Backtest
If eliteAI had been running last year
Stock value held, actual against modelled
Click a month to see what was suggested, what was done and what it cost.
The scorecard
Stockouts in the last twelve months, and how many the model would have warned about in advance
Average stock value held across the year, actual against modelled
Duplicate orders placed, and how many free text matching would have caught
Invoice overcharges paid, and how much rate card checking would have flagged
Where the model was wrong
Three cases from this data where following the suggestion would have been worse.
BRG-SKF-22324-CCK, main export pump bearing
- The model said
- Cut the line. Twenty four months without a movement and a unit value high enough to dominate the surplus ranking.
- The team did
- Held two units. The rotating equipment engineer classified it as insurance against a single point of failure.
The pump has no installed spare. A failure stops export for eleven weeks, and no movement history can show that.
CHEM-DEMUL-D2200, demulsifier
- The model said
- Lower the maximum. Consumption fell for four straight months, so the model read it as declining demand.
- The team did
- Held the maximum. Production knew the Walio tie in would raise water cut from the third quarter.
The drop was a well being shut in for workover, not a trend. The forward plan was never in the movement data.
INV-2026-02233, Nusantara Bahari Logistik
- The model said
- Flag the mobilisation line as above rate and send the invoice back to the supplier.
- The team did
- Accepted and paid. The rate was varied in a signed side letter that sits outside the rate card.
The variation was agreed in writing during the cyclone standby and never loaded into the contract record.