Repurposing a Commodity Hard-Disk Drive as a Physical Reservoir: A Preliminary, Single-Device Case Study

Mechanical hard-disk drives exhibit reproducible, non-random timing structure (actuator seek dynamics, spindle-phase modulation, drive-cache state) that is normally treated as noise to be hidden from applications. We ask whether this structure can instead be used as a computational substrate, in the spirit of physical reservoir computing. On a single commodity 5400 RPM laptop drive (Seagate ST1000LM035), we report four preliminary findings, each against an explicit baseline or control: (1) a physical-reservoir readout on the Mackey-Glass task outperforms a linear AR baseline (NMSE 0.384 vs. 0.555) but remains far behind a software echo-state network (0.0037); (2) a dense-associative-memory readout implemented as an unbuffered sector scan recovers 128 stored patterns at 20% noise with 100% accuracy, at throughput low enough to reveal a genuine host-side memory wall at the 10 GB scale; (3) a spike-timing-dependent relocation scheme reduces average access latency for causally paired addresses by roughly two orders of magnitude after training; (4) a Rosenstein-style largest-Lyapunov-exponent estimate on drive-timing traces under a slow chaotic drive signal is significantly positive against a surrogate-data null, but we also show this positive result is not exclusive to HDD mechanics — an SSD control shows the same qualitative effect at roughly a third of the magnitude, so we retract our earlier stronger claim and report the weaker, corrected one. We view every result here as a hardware-timing case study, not a validated computing paradigm. Code and data: https://github.com/DeptCreator/hdd-physical-computing (MIT).