Percentage Used Is Derived, Not Measured
Percentage used is often read as a direct answer to how much life remains, but it is the output of a division: cumulative media writes over a designed write total. That denominator is not a physical constant — it is planned DWPD multiplied by warranty years multiplied by the nominal capacity of the SKU. J750 is planned at 1 DWPD and J770 at 3 DWPD, both over a five-year warranty. The same host write volume therefore climbs the counter an order of magnitude faster on a 1.92TB J750 than on a 12.8TB J770. The field is comparable only within one model at one capacity point; ranking models against each other by it just renders the difference in denominators.
Write Amplification Is the Comparable Quantity
If only two health fields could be retained, they should be media writes and host writes, whose ratio is write amplification. It is independent of the denominator and therefore comparable by construction. A ratio above expectation says the workload shape does not match the drive's internal policy: a high proportion of small random writes, overwrite hot spots that are too concentrated, or logical allocation that squeezes the schedulable spare pool. Whether a selection holds and whether the planned DWPD suffices is decided by measured write amplification times host writes against the endurance budget — not by where the percentage sits. The capacity-step difference between J750 at 1.92/3.84/7.68/15.36TB and J770 at 1.6/3.2/6.4/12.8TB is exactly spare area, and it acts directly on this ratio.
MTBF and UBER Bound Populations
The specified MTBF of at least 2 million hours and UBER no worse than 1 sector per 10^18 bits read are statistical bounds describing expected behaviour of a large sample under stated conditions. Reading 2 million hours as the life expectancy of one drive, or inverting UBER to predict when a particular drive will return a bad sector, misuses the quantity. Their proper use is at system level: sizing redundancy, verification strength and inspection density — not triggering action on an individual drive.
Write Criteria as Rates and Inflections
Actionable criteria live in rates of change: whether spare block consumption departs from the curve implied by capacity and written volume; whether uncorrectable error counts move from flat to non-linear growth; whether temperature runs against the control threshold for extended periods. An inflection in any of these matters more than a single crossing of an absolute threshold — a threshold says arrived, a rate says how long is left. This also requires a fixed sampling period, since a rate is a differential and drifting intervals contaminate the result directly.
Baselines Bind to PN and Revision
Model names were frozen on 2026-07-22, while capacity, NAND, controller, firmware and BOM changes are managed through SKU / PN / Revision / ECN. So the normal curve for health data is not one product-level curve but a family of curves bound to a specific PN and firmware version. J750/J770 are in product definition/EVT and the figures above are planned values; the collection semantics of the health fields and their baseline curves are being fixed alongside engineering validation. Monitoring should record PN and revision with every sample, or comparisons across build lots have nothing to align on.