One version of loss
Potential, actual and event records must balance without well and plant losses being double counted.
BENCHMARK 04 / PRODUCTION LOSS / 2026
Production
Loss Atlas.
Can AI reconcile a disputed 90-day production-loss picture, reconstruct what happened and turn reporting into operational learning?
THE TEST
Reconcile the loss.
Explain the event.
Each model received the same fictional offshore loss portfolio and a major recurring compressor event. It must reconcile potential, actual production and scheduled and unscheduled loss records without counting the same impact twice.
The deeper test is whether the model can replay the evidence timeline, separate initiating event from failure mechanism and root cause, and convert the investigation into owned corrective actions with effectiveness checks.
RECONCILEREMOVE OVERLAPREPLAYEXPLAINASSIGNVERIFY
Make the sequence visible
Production changes and event callouts should show when losses began, changed and recovered.
Move beyond the label
Initiating event, mechanism and causal factors must remain distinct and traceable to evidence.
THE OUTPUTS
Inspect all three.
Identical evidence and task. Three independently produced working applications, presented without manual correction or a declared winner.
IMPORTANT
Return to all benchmarks →The installation, production history and loss records are fictional. These applications demonstrate model reasoning and interface design; they do not replace approved production-accounting rules, incident investigation processes or competent engineering review.