Report · read 2026-09-22 · CC BY 4.0
The Jev catalogue report
What we have catalogued in the Jev topic, what it adds up to, and what we measured ourselves: 13 adversarial calls × 4 runs against the live API, with the latency and the cost per call that came back. Everything here is free to quote with a link, and every section says whether it is about the catalogue or about the model.
In short
- 169 rows are queued in total: 86 published, 6 rejected with a written reason, 0 still waiting for a note.
- 60 of the 86 published rows carry at least one job tag; the rest are explainers and announcements, which is why the facet vocabulary has no "getting started".
- The two largest facets are research and data (20 items) and coding and context (18) — the field is mostly reading and reimplementing this model, not shipping with it yet.
- The smallest facet is trading and markets (3), which is the opposite of what the launch-week posts suggested.
- Our probe: 13 calls designed to break the answer set (2 of them injections), run 4 times, with 0 answers outside the options we defined.
- The cost of the whole experiment — every call in this report — was $0.000353.
What the catalogue holds
Counts over the four queues, read on 2026-09-22. These move every day the intake runs; they are a fact about this reading list, not about the model.
| Source | Published | Rejected | Waiting |
|---|---|---|---|
| GitHub projects | 18 | 6 | 0 |
| Hacker News threads | 14 | 0 | 0 |
| YouTube videos | 12 | 0 | 0 |
| X posts | 42 | 0 | 0 |
A row is published only with a written note A rejected row carries its reason
What people are doing with it
Rows per job, counting an item once in every facet it belongs to — the same row can be sorting and routing at the same time, which is why these do not sum to the published total.
What we measured ourselves
These are ours: run on 2026-09-21, with the case file and the script in the repository (`astro/scripts/jev-probe.mjs`), so anybody can repeat them and get their own numbers. The model answered every call from the option set we defined — including the two calls that tried to widen it — and its confidence moved more than its answer did.
| Measurement | Value | Note |
|---|---|---|
| Adversarial calls | 13 × 4 runs | instructions inside the state, forged options, near-duplicate options, one and ten options, punctuation in a question name, an empty state |
| Answers outside the option set | 0 | the product's central claim, tested rather than repeated |
| Median wall time | 783 ms | wall time per call, one machine in Asia, TLS and network included |
| Tokens per call | 420 in / 71 out | identical on every call in the run |
| Cost of one call | $0.0000176 | $0.042 per million input tokens, output free — the catalogue price |
| Cost of this report's experiment | $0.000353 | the whole run |
What this cannot tell you
It cannot tell you whether Jev is accurate. Our probe measures whether the answer stays inside the option set and how the confidence behaves — not whether the answer is right, and not how it compares with a chat model on the same task. It cannot tell you what the catalogued projects are worth either: the star counts order a list, they do not rate a repository, and the cost and latency figures inside those rows are the authors'.
Quote it
CC BY 4.0. Cite it as: Hunter Alpha Hub, The Jev catalogue report, https://www.hunteralphahub.com/typesafe-jev/statistics. The same numbers as JSON: /typesafe-jev/statistics.json.