Sydney Runkle posts four open questions on Jev context
A short public checklist of what nobody has settled yet about feeding state and questions to Jev.
Sydney Runkle posted four questions she has about context engineering for Jev, and no answers. The list is what should Jev do and what should it not do, how to represent a problem as state and questions, how to present that state and those questions optimally, and how to calibrate next steps on Jev's confidence and probabilities. She closed with "what else?", which is the part that makes it a checklist rather than a claim.
how can i calibrate next steps on jev's confidence / probabilities?
Nothing here is a benchmark and nothing here is a result. It is one practitioner writing down the open problems in public, which is worth reporting mostly because of which problems she picked.
The confidence question is the live one
The fourth item is the one that costs people money. Jev returns probabilities, and a probability is only useful if you have decided in advance what to do at each band. Runkle is asking how to calibrate that mapping, not asserting a rule for it. If you are building on Jev today, that decision is yours to make and defend with your own data, because there is no published answer attached to this post.
The first two questions are scoping questions. Deciding what Jev should not do is the same call as deciding which parts of a problem you hand to something else, and turning a problem into state plus questions is the work that precedes any of that. The third, how to present state and questions optimally, is the one most likely to have a boring empirical answer that somebody will eventually measure.
For people building on it
Treat this as a list of things to test rather than a set of instructions. The value of the post is that a builder is willing to say in public that these four are unsettled, which is more honest than most integration write ups. If you have run experiments on any of them, Runkle explicitly asked for additions to the list.