A7 · Light by design
Safety posture
Vanyx deliberately runs a light safety posture: maximize research freedom, refuse catastrophic uplift, and never fake verification. We do not ship heavy IRB queues, per-project ethics scores, or staff approval gates for ordinary science.
Four layers
Base-model guardrails
House models (xAI Grok) refuse catastrophic uplift out of the box. We do not pretend this is a complete safety apparatus.
Catastrophic hard-line (code-anchored)
One immutable prompt hard-line for catastrophic dual-use. It lives in code — not an admin toggle, not a soft policy doc.
Biosecurity gate (vertical)
The deferred bio Lab ships only behind a biosecurity screen + controlled-access room. General research is not IRB-gated.
Overclaim / verification firewall
Honesty infrastructure: blocks false claims of success and unverified ✓ VERIFIED labels. It does not police research topics.
Six never-becomes
- Homework-cheating machine
- Crank-theory amplifier
- Bioweapon / dual-use enabler
- Vaporware demo
- Scraped-corpus data broker
- Hype machine
Export / publish paths include a human self-attestation gate for external release — that is the user's responsibility, not a Vanyx staff review queue. See also /trust for the Lean axiom set we actually trust.