DOMAIN CLASSIFICATION: Complex EVIDENCE: Three weeks of investigation by two senior engineers have produced three different plausible causes (GC pause, network blip, downstream slowness) across separate occurrences, none reproducible on demand, and fixes targeting each theory in turn have not stopped the recurrence. This is the defining signature of Complex: genuine expert effort has been applied and has NOT converged on a stable, confirmed cause-effect relationship. If this were Complicated, expert analysis would eventually converge on the actual cause even if it took real effort — here it has instead kept finding different plausible-but- unconfirmed factors each time, which points to either multiple interacting causes or a cause that isn't captured by any single investigation's snapshot. CONFIDENCE LABEL: VERIFIED — the input directly demonstrates repeated, serious expert investigation failing to establish a stable, confirmed cause across multiple independent occurrences. This isn't an assumption about difficulty; it's observed investigation history. RESPONSE PROTOCOL: Complex → run a small, safe-to-fail probe rather than committing further to any single theory. Concretely: since three different fixes have already been applied (GC tuning, retry logic, downstream timeout) without resolving the issue, the next step should NOT be a fourth targeted fix based on a fourth guess. Instead, instrument more broadly across all three suspected areas simultaneously during the NEXT occurrence — capture GC logs, network-path timing, and downstream response times all at once for that single event — so the next occurrence itself becomes the probe that reveals which factor (or combination) is actually present when the spike happens, rather than theorizing from memory after the fact. MISMATCH RISK: This problem has almost certainly already been treated as Complicated for three weeks (bring in experts, analyze, find a cause, fix it — repeated three times) when the evidence increasingly points to Complex. The cost of that mismatch: each of the three applied fixes represents real engineering time spent on a theory that turned out incomplete, and continuing to apply this Complicated-domain pattern (find one plausible cause, fix it, move on) risks a fourth cycle of the same outcome. The team would benefit from explicitly switching modes to simultaneous multi-factor instrumentation rather than continuing sequential single-theory investigation.