Every pitfall below is a recurring finding in reliability prediction reviews; the guards are correspondingly practical.
| Pitfall | What it looks like | The guard |
|---|---|---|
| Selling the number as a field forecast | "The prediction says 48,600 h, so the fleet will see it" | Quote predictions as comparative estimates; track field MTBF separately |
| Unit mix-ups | FIT and f/10⁶ h blended in one roll-up: a factor of 1,000 | Declare units at the top; convert at the boundary, once |
| One environment for a multi-phase life | A missile electronics box predicted entirely in GF | Segment the mission; time-weight per-phase results |
| Extrapolating the models | Stress ratios above 1.0, temperatures past the tables | Treat out-of-table parts as derating findings, not inputs |
| Forgetting the unglamorous contributors | No PCB, solder-joint, or connector terms in the roll-up | Follow the handbook's roll-up procedure to the letter |
| Optimistic quality claims | Commercial parts carried at military π_Q "because the vendor is good" | Evidence for the claimed screening, or the honest factor |
| Ignoring duty cycle and dormancy | Calendar hours fed into operating-hour models | Define operating time explicitly; profile on/off phases |
| Quoting MTBF without its clock | Operating-hour prediction argued against calendar-hour field data | State the clock with every MTBF; publish both views when duty cycle < 100% |
| Wear-out items left in the λ sum | A life-limited item carried at a constant rate with no replacement plan | Screen against mean life; pull flagged items into conditional-reliability treatment |
| One-shot items on the hourly clock | A squib or cartridge carrying a per-hour failure rate | Per-demand probability, combined at the demand point, never in the λ sum |
| Silent zero from an unsupported environment | A model asked for an environment it does not define returns a perfect part | Treat any zero rate as a defect in the analysis until proven deliberate |
| Double-counting field evidence | Field-based specified rates further "corrected" by field-data methods or factors | One evidentiary basis per rate, recorded with the rate |
| Chasing the number, not the design | Assumptions tuned until the target is met | Declare and freeze the assumption set before the run; review changes to the assumptions, not only changes to the result |
| Frozen-handbook blindness | 1995 models applied to parts that did not exist in 1995 | Say so in the assumptions; consider newer model sets or field data for modern parts |
Most of these are process failures, not mathematics failures, and the last two deserve a closer look. The gaming problem is a contracting problem before it is an analysis problem: the foreword of MIL-HDBK-217F Notice 2 states in capitals that the handbook is for guidance only and shall not be cited as a requirement, and a prediction whose assumptions were tuned to clear a threshold has spent its credibility for one review meeting in order to satisfy a clause the handbook never asked anybody to write. The defence is procedural rather than mathematical: the assumption set (model, environment, temperatures, quality claims, exclusions) is declared before the run and any later change to it is reviewed as a change, not absorbed quietly into a better answer. And the frozen-handbook problem is structural: MIL-HDBK-217F Notice 2 predates three decades of component evolution, which is precisely why the serious modern practice is to treat handbook predictions as one pedigree among several, blended with field data where it exists, cross-checked against newer model sets where it matters, and always documented with their assumptions attached.