POLARIS / TECHNICAL EVALUATION
Better attribution.
Tested on unseen flights.
Inspect a recorded evaluation on public aviation data, including the baseline comparison and held-out checks.
THE BUSINESS GOAL
“Find why our flight-delay decisions are wrong.
Then improve them.”
A late departure was being charged to the carrier. Polaris investigated the records, identified where that rule failed, and delivered a better attribution method.
Public US aviation data · Historical replay
Attribution accuracy
Matches carrier-reported main causeControllable split
Controllable vs non-controllableSame evaluation set: 4,031 flights on held-out aircraft, May–Jun 2024.
PUT THE RESULT TO THE TEST
Better on data
it hadn’t seen.
The improvement held up in both frozen evaluations. Open the evidence behind the result.
Independent pilotPASS
300 flights. Valid outputs and accuracy at least as high as the strongest baseline on the same flights.
Expanded evaluationPASS
2,000 flights from other aircraft after the source-data period. Passed the accuracy, improvement and recall requirements on the first attempt.
What this demo shows+
This is a completed Polaris run using public US carrier on-time data. You are viewing recorded results, not launching a new run. The figures measure agreement with carrier-reported causes in historical replay.
Download the result data ↓YOUR TURN
Three records.
One wrong reference.
Can you spot it?
Change the sample data. Run the check. Take the evidence with you.
YOU KNOW THE GOAL. WE HANDLE THE AI.