CAIN-42 CAIN StudioWhat actually ran
Measured numbers
Latency and correctness from a real run on the host that built this page. Not a production benchmark.
The run
The numbers below come from python3 -m aether33.stress, which ingests a seeded corpus, runs recall queries, performs cross-tenant isolation probes, verifies every tenant's chain and executes all 33 invariants. The JSON report is stored at artifacts/aether33/stress_report.json and is what this page renders — there is no hand-entered figure.
Report
6.139 mswrite p50 (2,000 writes)
11.518 mswrite p95
6.623 msrecall p50 (500 recalls)
8.76 msrecall p99
33/33invariants
0isolation failures of 50
62.053 msconcurrent write p50 (8 threads)
0concurrent errors · 8/8 chains intact
| Measured at | 2026-11-02T15:40:06Z |
|---|---|
| Host | Linux-7.0.0-38-generic-x86_64-with-glibc2.43 · Python 3.14.4 · 8 CPUs |
| Workload | 2,000 documents · 4 tenants · 500 queries · seed 7 |
| Throughput | 145.6 writes/s (write phase, 13.741 s) · 31.428 s total including the concurrency phase |
| Database | 2,224,128 bytes on disk |
| Concurrency | 8 threads · 800 writes · 0 errors · p50 62.053 ms / p99 205.313 ms · 8/8 chains ok · 0 isolation failures of 8 SQLite is single-writer; concurrent tenants serialize on the write lock. Higher latency here is expected, not a defect. Zero errors and intact chains are the pass condition. |
| Chains | stress_t0_7: 500 entries, ok; stress_t1_7: 500 entries, ok; stress_t2_7: 500 entries, ok; stress_t3_7: 500 entries, ok |
| Determinism | 0 ordering mismatches across repeated identical queries |
Measured on this host, single process, one SQLite file in WAL mode, no network and no model. The concurrency phase runs 8 threads against that same file; SQLite is single-writer, so higher latency there is expected. Not a production benchmark and not a multi-host test.
How to reproduce
cd /root/cain-studio-platform
PYTHONPATH=cain-microservices:. \
python3 -m aether33.stress --documents 2000 --tenants 4 --queries 500
AETHER-33 is a product over CAIN-42's governed-memory stack, not a new architecture moat.
Memory never carries authority: every response is annotated
grants_authority: false.
No third-party audit has been performed. Limits ·
FAQ · what is live now