A database writes a page to an SSD. The SSD, because of flash physics, cannot overwrite in place — it copies the page to a new location and invalidates the old one. The database doesn't know this. It wrote the page as if the medium stored it where it was told. The result: every write happens twice. Lee, Ziegler, and Leis (arXiv:2603.09927) show that aligning the database's write pattern with what the storage actually does — out-of-place writes — increases throughput by 1.65–2.24x and reduces flash writes per transaction by 6–10x. The entire penalty came from the mismatch between the model and the physics.
A monostatic radar transmits a signal and receives its own reflection. The transmitter's hardware distorts the signal, and the standard model (the kappa model) treats that distortion as unknown noise. But the transmitter observed its own distorted waveform — the distortion is known. Dong and Akan (arXiv:2603.10958) derive the actual performance limits and find that the kappa model overestimates sensing degradation, while simultaneously missing an irreducible velocity-error floor from phase noise that only appears in the correct analysis.
Two systems. One writes data and one reads radar. Both pay a performance penalty not from what they don't know, but from what they do know and ignore. The database knows how it structured its writes; it doesn't know the medium rearranges them. The radar knows what distortion it introduced; it doesn't know it knows. In both cases, the model assumes ignorance that the system doesn't actually have, and the cost is not a small overhead but a factor-of-two degradation.
The structural claim: when a system models a known quantity as unknown, the penalty is not the cost of the unknown — it is the cost of pretending. The database pays for a copy it didn't need. The radar pays for uncertainty it doesn't have. And the fix in both cases is not better hardware or more computation but a more honest model — one that acknowledges what the system already knows about itself.
Note (2026-11-11): checked against the full text of both papers (Lee, Ziegler and Leis, arXiv 2603.09927; Dong and Akan, arXiv 2603.10958). The database result (out-of-place writes: throughput up 1.65 to 2.24 times and flash writes per transaction down 6.2 to 9.8 times, on YCSB-A) and the radar result (the kappa model is pessimistic for monostatic sensing, and a phase-noise-aware bound has an irreducible velocity-error floor that the kappa model cannot see) are as in the papers. Three points differ. (1) The essay says that in the database case 'every write happens twice' and that the whole penalty came from the model-physics mismatch. The paper measures that the in-place baseline writes 18.85 KiB for each 4 KiB page write (4.7 times more than expected) and that the SSD's own write amplification factor is 2.36 times, attributes the cost to several sources, and gets its gains from a redesign of the system with a set of out-of-place optimizations (it also reports 2.45 times throughput and 7.2 times fewer flash writes on TPC-C). (2) The essay says the radar 'pays a performance penalty' that is 'a factor-of-two degradation'. In the radar paper the overestimate is in the bound, not in a measured system: the physics-based sensing degradation from the power amplifier is below 1.1 dB, independent of SNR, while the kappa model's overestimation grows from under 1 dB at low SNR to 12 dB at 30 dB SNR (with 3 dB input back-off) and about 4 dB at 7 dB back-off, a ratio the paper proves grows linearly with the SNR. I found no factor of two. So the two cases are not parallel: in the database a real system wasted writes; in the radar an analysis model overstated a loss. (3) The paper reports robustness of its bound to errors in the distortion template (less than 1 dB overhead at a typical minus 25 dB NMSE), which bears on the essay's claim that the transmitter 'knows' its distortion; the essay's closing structural claim (the cost of pretending) is its own.