Rivet split one Jev severity question into three yes/no ones
Nathan Flurry says the single SEV1 to SEV4 prompt gave mediocre results, and plain if-statements fixed it.
Rivet's first attempt at ticket triage with Jev was one question. Nathan Flurry says they asked for the severity of an issue in a single prompt, with SEV1 through SEV4 as the possible outputs and examples of what each level meant, then returned whatever came back. His summary of the result is that it gave "meh results", and that the approach "defeats what makes jev awesome".
The fix was to stop asking for a judgement and start asking for facts. Flurry's write-up shows the rewrite as three separate calls, is_down for whether a production service went down or became unusable, is_blocked for whether the customer is blocked with no workaround, and is_broken for whether something in Rivet is broken or failing. The ranking then happens in ordinary code. Blocked returns SEV1, down returns SEV2, broken returns SEV3, everything else is SEV4. He says it "works like a charm".
jev does not reason
this is a feature of jev, not a limitation
Why the shape of the question matters
Flurry's framing is that Jev does not reason, by design, and that Rivet should have read the manual first. A normal LLM would burn reasoning tokens arriving at a severity label. Jev does not, so a question that requires the model to weigh several factors against each other and collapse them into one of four buckets is the wrong question to hand it. He argues this is a feature rather than a limitation, because the weighting lives in code you can read, monitor and backtest.
That backtesting is part of the loop. When a ticket comes out weighted wrong, Rivet hands the ticket and the expected answer to an LLM, which adds it to a test suite, adjusts the decision logic and reruns it against old issues. Flurry says the classifier logic was solid within a few tickets.
The question counts give a sense of how far the decomposition goes in Rivet's internal support tool. Support classification and customer status takes 13 Jev questions. Severity takes 6. Team replies and problem identification takes 6, follow-up timing 6, linked dependencies and later conversations 4, addressed teammate 2, conversation grouping 2.
For people building on it, this is one team's experience on one triage workload, not a benchmark. But if you ported a reasoning prompt straight across and it underperformed, the number of judgements packed into a single ask is the first thing to look at.
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.
