Every RAM study produces a number — 96.2%, say — and that number promptly acquires a life of its own in investment memos and partner presentations. What rarely travels with it is the honesty check: would this model have predicted the losses you actually suffered over the last three years? Ask that question and many impressive models fall quiet.
Check one: whose failure data?
Generic reliability databases like OREDA are invaluable starting points and dangerous ending points. Your compressors do not fail at the industry-average rate; they fail at their rate, driven by your service conditions, your operating discipline, and your maintenance history. A defensible model blends generic data with site history — and says explicitly which is which.
Check two: repair time is not downtime
Models routinely use bare mean-time-to-repair figures while real events are dominated by everything around the wrench time: fault diagnosis, mobilization to a remote site, waiting for the spare that was never stocked, permits, and production ramp-up. If your logistics delays aren't modeled from actual events, availability is systematically overstated — usually by enough to change the investment decision.
Check three: the back-cast
Run the model against the past. If it cannot reproduce, within reason, the production efficiency and top loss contributors of the last three years, it has not earned the right to forecast the next ten. When a model does pass this test, it becomes genuinely powerful: you can rank debottlenecking options, sparing changes and maintenance strategies by barrels, with error bars you can stand behind.
Key takeaways
- Blend site failure history with generic databases — and disclose the mix.
- Model full downtime events (logistics, spares, ramp-up), not bare repair times.
- Back-cast against 3 years of actual losses before trusting any forecast.
- A calibrated RAM model turns investment debates into ranked, quantified options.