Post
CommunityJev Media & Talks
Can Jev save my inbox from the Democrats?
Stephen Solka's Hugging Face article tests whether Jev can filter political fundraising from an inbox: a frozen 100-email eval, one noul question per email, then follow-on experiments that shrink the classifier to CPU-only inference with SetFit and ONNX Runtime.
- Format
- Post
- Reading time
- 2 min
- Published by
- Community
- Author
- Stephen Solka
- Last verified
- Oct 2026
- Languages
- Python
- Platforms
- OpenRouterAPI
- Capabilities
- Requires API key
Highlights
- Frozen 100-email eval: 50 political and 50 nonpolitical, half from the author's Fastmail Feed folder and half from public sources.
- First run: 99 of 100 labels correct, 0 of 50 political emails missed, 1 of 50 nonpolitical flagged, across seven bulk calls.
- Requests use model typesafe/jev-1.13 with one noul question per email over shared state on OpenRouter's Decisions API.
- Follow-on experiments move to CPU-only inference with Python, Hugging Face Datasets, SetFit and ONNX Runtime.
Quickstart
python
response = requests.post(
"https://openrouter.ai/api/alpha/decisions",
headers={"Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}"},
json=payload,
timeout=180,
)
answers = response.json()["answers"]
probabilities = {email_id: answer["noul"] for email_id, answer in answers.items()}Watch out
Page copy reviewed 2026-10-15 from the Hugging Face article, which was truncated mid-eval. The 99/100 result is the author's own small curated pilot with assistant-reviewed labels, not a human-adjudicated error-rate estimate for the whole inbox.
Questions about this entry
Short answers drawn from the linked project or page.
Filtering political fundraising email. The author writes a policy, labels messages political or nonpolitical, then asks Jev one noul question per email and reads the probability that the message is political outreach.
From 100 frozen emails: 50 political and 50 nonpolitical, half from the author's Fastmail Feed folder and half from public sources including SetFit/enron_spam, balanced so the test cannot just reward recognizing a sender, footer or writing style.
99 of 100 labels correct: zero of the 50 political emails were missed and one of the 50 nonpolitical emails was flagged, across seven bulk calls. It is a small curated pilot, not an estimate of the whole inbox error rate.
One request carries several emails in shared state, with a keyed noul question per email using model typesafe/jev-1.13 through OpenRouter's Decisions API. Labels and annotation rationales never entered the request, and the runner validated returned IDs and saved raw responses.
Two follow-on experiments about privacy, local inference and size: the author works toward a classifier cheap enough to run on a CPU, using Python, Hugging Face Datasets, SetFit and ONNX Runtime.
It does not estimate prevalence or error rate for the whole inbox, and labels were assistant-reviewed rather than human-adjudicated. New discoveries went into separate training and challenge pools instead of the frozen test set.
Stephen Solka published it as a Hugging Face community article on October 6, 2026. The digest copy was reviewed on 2026-10-15 and was truncated partway through the evaluation results.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: 1,565 emails, six decision models
X: Fifty emails in under two seconds
More Jev resources like this
Toolkit, knowledge base and local dashboard for building with Jev via OpenRouter Alpha Decisions or the TypeSafe Direct API: a zero-dependency SQLite request log, a localhost REST server and Python CLI demos.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
rileycarney/jevtoolsHTML
rileycarney/jevtoolsHTMLProbabilistic control flow for TypeScript: Hunch asks a System One model such as Jev yes/no questions, thresholded booleans, picks and ratings, and returns typed answers you can branch on in ordinary code.
Jev Repos & SDKs#community#catalog-import#cobanov
Community
carldaws/hunch-tsTypeScript
carldaws/hunch-tsTypeScriptMicrosoft's announcement of Microsoft-Decision-1, a model for fast decision scoring built for routing, classification, prioritization, verification and workflow control, available in Microsoft Foundry and through OpenRouter. Published 2026.10.09.
Jev Media & TalksPost#community#catalog-import#madewithjev
Community
A 144.3M-parameter bilingual typed decision model fine-tuned from jhu-clsp/mmBERT-small: it answers noul, choice, and score questions in one forward pass with calibrated confidence and zero output tokens, and it runs on CPU.
Jev Models & Reimplementations#community#catalog-import#cobanov
A dependency-free Python pipeline that rebuilds a site's internal link map with Jev: code picks candidate targets and in-copy anchor phrases, Jev judges them, and only links that clear the threshold are placed.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
stas4000/jev-linkmapHTML
stas4000/jev-linkmapHTMLA small, idiomatic Elixir client for Jev: Jev.ask?/2 returns a real boolean, choose/3 picks among supplied options, score/3 rates ordered levels, and analyze/3 sends several typed questions in one provider request.
Jev Repos & SDKs#community#catalog-import#madewithjev
Community
elue-dev/jev_elixirElixir
elue-dev/jev_elixirElixirFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.