84 of 84: correctness and isolation under a hostile harness
Empty strings, 100 KB values, null bytes, emoji, 16 threads hammering across tenants. The stress harness throws 84 nasty checks at Crowkis and counts the cross-tenant leaks. The leak count is zero.
Speed is negotiable; correctness is not. A cache that occasionally serves tenant A's answer to tenant B isn't a fast cache, it's a data breach with good latency. So the harshest part of our harness isn't about performance at all, it's 84 correctness and robustness checks designed to make Crowkis misbehave.
Robustness, correctness, concurrency, and benchmark sections, all green.
The robustness checks feed Crowkis the inputs that break naive servers: empty and whitespace-only queries, 100 KB values, Unicode and emoji, embedded CRLF and null bytes, absurd thresholds, extreme TTLs. None crash it, none corrupt a neighbouring entry. The correctness checks verify the invariants that actually matter: exact round-trips, negative-cache behaviour, pinning, and agent-memory isolation.
The concurrency check is the one we lose sleep over: 16 threads running 60 operations each, deliberately interleaving reads and writes across tenant boundaries, then auditing whether anything crossed. Zero leaks. That result isn't an accident of timing, it's the single-writer actor making races structurally impossible, validated under exactly the load that would expose them.
We score ourselves 9 out of 10 on correctness and isolation, and 4 out of 10 on hot-path latency. The point of a brutal scorecard is that you can trust the high numbers because you can see the low ones.
That honesty is the product. The same harness that returns 84/84 on correctness flags CDEDUP's latency stall and the throughput ceiling without flinching, because a scorecard you can only pass isn't measuring anything. The cache earns the critical path by being boring exactly where boring is the whole job: never the wrong answer, never the wrong tenant.