Jev: I Commissioned This Exact Research 45 Days Ago

I set out to commission a deep research on Jev. I commissioned that exact research 45 days ago, it ran, a decision was accepted, and two shadow runs on my own data are done. The verdict I scheduled for myself on 30 October is still not written.

What is Jev and how applicable is it to me? How is it different?
who already uses Jev in practice, and how? I want a deep research not on random examples, but on ones that look like my cases.

Then I will put Jev against my use cases and see where it does better, where worse, where it might be faster, where significantly cheaper, and why exactly.

Separately I want to check how compatible Jev is with my memory layer, my second brain, and find possible points of overlap.

I hope my co-founder will tell me how applicable Jev is to me and on which specific cases.

⭐⭐⭐ The brief repeats my own brief from 25 September almost word for word. In the chat "001 NEW SESSION", 25.09 at 23:53, I wrote: "Jev llm. Run a DR and start using it". The research DR26-09-26-HUB-01-2230, on whether to move the fleet's decision points onto a typed decision model, ran the next day, reached synthesized, and produced an accepted decision, decision-jev-typesafe-rail-2026-09-26.

⚠️ The quorum there was 3 of 6: gemini dropped out (AutoFF logged out of Google), chatgpt in the browser gave a read-timeout twice, codex CLI returned a revoked refresh token, claude CLI was under a ban from 22.09. So the previous pass has an honest hole, and that hole is the only argument for a new one.

⭐⭐ "What Jev is and how it differs" fits in one line. A TypeSafe AI model, out of stealth on 15 September. It does not write text: it takes a state plus a map of typed questions and returns Noul (a probability from 0 to 1), Choice (a pick from up to 255 options plus the probability distribution and a confidence) and Score (a scale of 2 to 10 levels). There is nothing to parse and the response shape cannot drift from the request. Input $0.042 per million tokens, output is not billed at all, latency 180-470 ms. The founder, Diogo Almeida, co-authored the InstructGPT paper.

⭐⭐⭐ And here is the answer to "who already uses it in practice", which surprised me: publicly, nobody. Checked live today: there are seven detailed explainers on Jev in the wild and not one named customer, not one case study, not one production report. Two independent sources also flag the vendor's own framing: Sanity notes that "cannot hallucinate" refers to output format, not factual correctness; DigitalOcean calls "mathematically guaranteed" an untested claim.

So examples that look like my cases do not yet exist anywhere. They do exist with me:

⭐⭐ And in those runs Jev caught what my regex cannot: a 48 KB grok report with zero links in it, meaning model memory rather than research, and 55 KB of GLM interface debris. My regex had marked both "fit for use". In the second point, 11 of 300 disagreements turned out to be regex misses rather than Jev misses, and the fix let me delete 557 junk rows.

⭐⭐ "Put Jev against my use cases" is also done, and five further points are already in my registry with named consumers and named blockers: the content slop gate (no blocker, next in line), "is the ball on our side or theirs" in PR replies (no blocker), lead triage (blocker: personal data), voice routing (blocker: Russian never measured at volume), and skill selection. The last blocker is both funny and exact: I have more than 255 skills, and Choice in Jev holds at most 255 options, so it needs a two-question hierarchy.

⭐⭐ The benchmark against Opus already ran, and it says something I did not expect. 60 roles, ground truth labelled by eye:

The conclusion I drew then: on typed decisions, quality hits the ceiling of the task rather than the size of the model. Four different models produced exactly the same score, so paying for size buys nothing. ⚠️ But n=60 is a probe, not a benchmark: it honestly shows a sixtyfold gap in price and does not show a one-point gap in quality.

⭐⭐ How to ask properly is also evidence from that research. ⛔ You cannot ask one whole question like "is this phishing, is this slop?": that way Jev scored 62.6% against 91.8% for my own regex. You have to ask 3 to 5 atomic features and add them up in code, which gave 95.0%. On general quality Jev sits around Claude Haiku 4.5, below Opus.

⚠️⚠️ Now the compatibility with the second brain, and the answer was written in advance. It is negative on data. Jev has no zero-retention by default, the vendor's telemetry is indefinite, and through OpenRouter my state is seen by two parties: OpenRouter and TypeSafe. That is why the decision says outright: personal correspondence, lead cards, call transcripts and CRM do not go into state. And the second brain is made of exactly those things.

One move lifts it: my own TypeSafe key, direct, with no middleman. The key is not set up (there is no typesafe.env file), and the direct rail switches itself on the moment it appears. Until then, any overlap with the second brain can only be on metadata, not on content.

⛔ Separately: "we will train our own model on Jev's answers" is forbidden by the contract, clause 2.3(b). That route is closed by paper, not by engineering.

⚠️⚠️ And the uncomfortable part. I set the verdict date myself. On 30 September, by voice: "I would use it as a backstop for now, measure the result, how good or bad it is, and in a month we will decide whether to keep it or not." That date went into the task task-2026-09-30-jev-keep-or-drop-decision as 30 October. Today is 10 November, the task is open, the verdict field is empty, and it is 11 days overdue. The harvest dates on the two points were 10 and 12 October; both have passed too.

⚠️ Why the verdict cannot be closed: the acceptance criterion asks for a thousand live decisions from the counter, and the counter jev_usage.jsonl lived on the hub's drive, where Python has been broken since 2 October and still is. So the instrument could not write for a month, not because Jev was never called, but because there was nothing to write with.

⚠️ The points instrument prints "both points silent, 0 calls" today, and that line cannot be trusted: it already lied in exactly this way once. On 30 September it said "DR gate silent, 0 calls" against 144 real calls, because it counted the name leg-shadow while the work was logged as leg_shadow. That day every point was alive: 144, 340, 180 and 4 calls.

⭐ So instead of a new deep research on Jev I need three things in order: set up the TypeSafe key so overlap with the second brain becomes possible at all; move the counter onto a machine that works; and write the overdue verdict on the two points where the measurement already exists. A new research is worth exactly one hole: three of six rails were silent last time.

The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261109.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261109.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.