Jev Security Gate
hackerlogs@hackerlogs
Jev is a typed System One gate that acts as a second opinion on tool calls and retrieved passages, outputting a choice, score, and yes/no decision within a specified latency.
Take@typesafeai shipped Jev on 15 Sep as a typed System One gate: state in, Choice/Score/yes-no out, 70 to 500ms, no free-form string.
Jev cannot invent a tool. It can still pick the wrong one.
@typesafeai shipped Jev on 15 Sep as a typed System One gate: state in, Choice/Score/yes-no out, 70 to 500ms, no free-form string. Zero-hallucination here means schema match, not a correct deny. @CompleteSkeptic said as much on HN: type safety is not factual correctness. jev-1.13 does not treat state as hostile. Injected text can still steer a wrong yes. @vercel put it on AI Gateway the next day. Useful as a cheap second opinion on tool calls and retrieved passages. Not a sandbox.
Operator move tonight:
1) Put Jev on the proposed tool call as advice, one risk per question.
2) Keep allowlists, spend caps, and step limits in code. The gate does not grant.
3) Pair every Choice with a yes/no on whether any action should run.
Primary: typesafe.ai/blog/introducing-system-one-mo
Write-up: hackerlogs.com/blog/typesafe-jev-security-
Related builds
