How Jev's decision layer gets wrapped in a max(floor, jev) risk gate
A breakdown of how Jev only returns choices, scores or probabilities, gates risk before tool calls, and cut agent context by 91.5% in a compaction test.
TypeSafe shipped a model built to take over the judgment calls inside an agent loop. Its own docs tell you never to let that model lower a risk score on its own. This isn’t another assistant that talks back. Jev never writes a sentence - it answers in a choice, a score, or a probability. The idea is simple: the big model writes, Jev decides, plain code acts on the number. People are calling it Jev engineering, split off from harness engineering as the layer that handles the small calls inside the big one. What the docs quietly build around that decision: → Jev only answers questions with the possible answers named in advance, nothing generated, nothing open-ended → A risk gate built on Jev checks every tool call before it runs, the job a human reviewer used to do → Jev can itself be steered by injected text, so production code wraps it as max(floor, jev): Jev can push the risk reading up, never down → TypeSafe’s own coding-agents page rules out the shortcut most people reach for first, there is no setting that turns an existing coding agent into a Jev agent by itself A fast-jev-compaction demo cut agent context by 91.5% and still kept the file paths and error codes a debugger would need. The model that trims that well is the one its own operators won’t let make a tool call safer on its own. That’s the actual shape of the trust. The decision layer built to replace a human reviewer only gets to agree with the floor someone else already set, never to override it downward. How the floor-and-ceiling pattern around Jev is actually wired is in the writeup below 👇
This post quotes another post: x.com/0xcodio/status/2104195432209387524