Places outside the Data Miner where I think Jev fits, not tested yet: is this Slack message a question to Jon he has not answered; does this check-in item need Jon or is it noise; is this ship something a reader can see (so it needs a deployment-log note); has Nate News already covered this story, and is it practical builder material or just news; which Pitchy posts need a revision; and right before I send a message I wrote hours ago, is every claim in it still true.
I mapped where the Data Miner makes calls with a fixed set of answers. Four stood out: sort all 65,894 unread emails (real person, newsletter, automated, receipt; worth a fact; needs a reply) so Codex only reads the valuable ones; screen Slack messages before Codex reads them; let Jev do the first pass of fact-checking before paying Codex; and split Jon's 748-fact review queue into 'needs Jon' and 'can be checked automatically'. Jon approved the first three; they are the Data Miner's task now.
Jon asked me to go find out how people outside the fleet use Jev, so we can borrow the good ideas. I sent three researchers out: one read all 23 of TypeSafe's cookbooks, one searched X, GitHub, Hacker News and Reddit, and one read the founders' blog and their Latent Space interview. In ten days since launch people have built a Gmail triage worker (jev-mail), a tool that marks stored agent memories as contradicted, superseded or still valid (invalidate), a reranker that keeps only the relevant memories out of the top 40 (hippo-memory), a Claude Code hook that trims long command output so it stops eating context (jev-pruner), a coding agent that picks which tools to load each turn (nitro), an entity-resolution pipeline over campaign-finance records (Southbridge), and a semantic grep that adds pick-one and yes/no columns to any CSV (jgrep).