Jack Cheng runs Jev on both ends of email triage
The second classifier takes how well he slept as an input, so the same inbox sorts differently day to day.
Jack Cheng wired Jev into two stages of an email classification system, a shape he floats as a Jev sandwich. The first pass classifies the mail. The second pass decides what is worth his attention based on how well he slept and, in his words, his current vibes. Same inbox, different output depending on the day.
to figure out what’s worth my attention based on how well I slept and my current vibes
He calls it an experiment. There are no accuracy numbers, no latency figures and no cost figures attached to the post, and nothing about how sleep and mood get into the prompt in the first place. So this is a shape, not a result.
Why the second pass is the interesting half
The usual way to build this is one classifier followed by rules. Mail gets a label, then a filter decides what surfaces, and the filter is code somebody has to maintain. Cheng's version moves the decision into a second classification call and makes his own state an input to it rather than a condition applied afterwards. The triage rule is not written down anywhere. It is inferred per message, per day.
That is a different failure mode than a rules engine has. A filter that drops something important is debuggable, you read the rule. A second classifier that drops something important on a bad sleep night is debuggable only if you logged the state you fed it. Anyone copying this should keep that input in the log next to the decision, or a missed email is unreproducible.
What this does not show
For people building on it, one person's inbox is not an eval. There is no baseline here, no comparison against single pass classification, and no count of how often the second pass changed the answer. The claim on the table is that the shape works well enough for Cheng to keep running it, and nothing stronger. If you want the pattern, take the part that is cheap to test, user state as a classifier input instead of a post hoc filter, and measure it on your own mail before you trust the sandwich.