The numbers we
can actually prove.
Most of this page is zero on purpose. CIRVIX runs locally and keeps no telemetry, so the only numbers we publish are the ones we can measure and defend — each one carries its date and how it was obtained.
Verified engineering facts
These are measured properties of the released codebase, re-runnable by anyone:
| Fact | Value | How you check |
|---|---|---|
| Deterministic policy decisions in the adversarial suite | 11,629 cases · 0 bypasses | npm run verify:adversarial in the public repo |
| Engine test suite (Node) | 793 passing | npm test, packages/agent-control |
| Engine test suite (Python) | 108 passing | python -m unittest discover -s tests, packages/cirvix-python |
| Audit chain integrity | hash-linked records, verifiable offline | cirvix audit verify after any run |
| Decision overhead (demo workload, P99) | <15 ms per decision | npx @cirvix_ai/agent-control demo --fast locally |
Adoption numbers
installs reported. The runtime phones home to no one — including us — so we cannot see them until you tell us.
decisions protected that we know of. Your audit chain is yours; it never leaves your machine.
teams onboarded. This page updates when a number is real, not before.
Last updated: 2026-08-25. When these change, the date changes with them.
Why publish zeros
Because every inflated launch number gets debunked in the comments within an hour, and a company that pads its metrics once is never believed again. If honesty about scale costs us distribution early, that is a price we pay knowingly. What we can promise instead: the engine itself is deterministic, auditable, and free to try on your own worst agent — right now.
Install Free Benchmarks Escape Benchmark Sandbox Founding 100 Book a walkthrough