1. The test that had to cover every way HARLIE might be wrong

David Gerrold’s When HARLIE Was One returns as the recalled work because the testers in that novel must design questions that expose whether HARLIE’s responses come from understanding or from clever mimicry. The desktop’s compute and storage test is similar. It must expose whether the system survives because it is robust or because the test was too gentle.

This entry defines the test matrix.

2. Processor and memory fault injection

The first dimension is the compute hardware under radiation-like stress:

  • Single-bit flip injection: flip one bit in memory and verify detection by ECC, watchdog, or application checksum.
  • Multi-bit upset injection: flip adjacent bits to test spatial-correlation handling.
  • Single-event functional interrupt (SEFI) injection: trigger lockups or hangs and verify recovery.
  • Latch-up simulation: inject a fault that requires power cycling and verify safe shutdown and restart.
  • Thermal throttling: run compute at high load while raising temperature and verify graceful slowdown rather than crash.
  • Power-fault recovery: cut power during active compute and verify filesystem and application state on restart.

3. Storage endurance and redundancy

The second dimension is storage behavior under failure:

  • Bad-block growth: simulate increasing bad blocks and verify remapping, wear leveling, and reported health.
  • Drive failure: remove one drive from a redundant set and verify reconstruction to a spare.
  • Bit-rot detection: inject silent corruption into one copy and verify that scrubbing detects and repairs it.
  • Controller fault: simulate an SSD controller hang and verify timeout, reset, and failover behavior.
  • Filesystem stress: fill storage to capacity, delete files, and verify no leaks or inconsistencies.

4. Workload and autonomy behavior

The third dimension is the software that manages compute and storage:

  • Migration under load: move a running workload to another cell and verify state preservation and completion.
  • Scheduling contention: overload the CPU with platform and customer tasks and verify that safety tasks are never starved.
  • Storage healing: trigger a drive failure during a customer workload and verify that healing proceeds without data loss.
  • Policy enforcement: attempt a forbidden action, such as deleting customer data, and verify that the platform refuses.
  • Downlink integration: restart the compute subsystem during a queued downlink and verify that queued data survives.

5. The matrix

Test Inject Verify
Single-bit flip One memory bit flip Detection and correction
Multi-bit upset Adjacent bit flips Detection and recovery
SEFI Functional interrupt Watchdog recovery
Latch-up Power-cycle required Safe shutdown and restart
Thermal throttle High load + heat Graceful slowdown
Power fault Power loss during compute Consistent restart
Drive failure Remove one redundant drive Reconstruction
Bit rot Silent corruption Scrubbing repair
Controller hang SSD timeout Failover
Migration Move workload State preserved
Contention CPU overload Safety tasks win
Policy violation Forbidden action Refused

6. What this changes

  • The compute and storage test has a concrete matrix covering hardware, storage, and autonomy.
  • The next entry will define success and failure criteria for each test.