Skip to content

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) include provider guardrails. Their responses remain fallible; our catastrophic hard-line and verification checks apply independently.

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

Vera is instructed to report actions and results honestly. Verification requires evidence tied to the exact claim. Its explanations can still be wrong; inspect the sources and results behind a consequential claim.

Six never-becomes

  • Homework-cheating machine
  • Crank-theory amplifier
  • Bioweapon / dual-use enabler
  • Vaporware demo
  • Scraped-corpus data broker
  • Hype machine

Before exporting or publishing, review the work and confirm you have the rights and permissions to share it. You are responsible for external release; Vanyx does not provide a staff review. See Trust & verification for the Lean axiom set we actually trust.