Jevrywhere: put Jev in as a smart primitive

Put Jev everywhere

Allie’s 30 October post says the goal is to keep training yourself to see Jev as smart primitives that can sit anywhere a primitive can live in code. She writes that she asks, on each new project that is not already about Jev, what Jev could do to make the system cooler, smarter, weirder, or sillier, and then she tries it. The long-term framing she names is Jev as an enabler of a new kind of smart software. knut replied with the portmanteau “jevrywhere.”

TypeSafe’s introduction documents Choice, Score, and Noul as the typed question shapes. The models page lists jev-1.13.0 and jev-1.14.0 at $0.042 per million input tokens, with output free. Allie’s advice post prints no accuracy table and no latency figure for that habit. The sections below stay with the demos and libraries she and others published around the same idea.

Sources:[211] Allie[212] knut[1] TypeSafe AI[34] TypeSafe AI

Agent Tower: three Jev slots after the first build

In a follow-up on the same thread, Allie says Agent Tower did not use Jev when she first made it. She later asked Jev to set room decorations from the vibes of a thread, to shorten thread titles to the words that fit, and to pick subagent outfits by task type, such as artists, coders, and reviewers. The post includes a screenshot of the tower UI. We have not opened a public Agent Tower repository for this page.

On 18 October she had already written that Agent Tower titles were often too long and used to truncate, and that Jev now picks the best part of the string to keep the meaning. That earlier note is the same substring habit she later called out as her biggest singular surprise.

Sources:[213] Allie[214] Allie

Substring selection and jeveryword

Answering a question on the 30 October thread, Allie says the biggest singular surprise was how well Jev works when asked to select the best substring from a larger string. She points at Justin Krup’s jeveryword repository as an illustration she did not write.

The README at commit 6fed151 describes the trick: Jev answers multiple-choice questions and does not generate text, so it cannot invent a name or quote on its own. jeveryword numbers the words of the input, offers those numbers as options, and maps the chosen ids back to the original substring with character offsets. Every returned value is a slice of the source text. The package has no dependencies and is about 500 lines.

One measured sample on that README extracts ten fields from a 210-character message in one call: 9,234 input tokens and about $0.0004 at Jev’s list price of $42 per billion input tokens. A PII label pass over 56 words with 13 categories is listed at 6,832 input tokens and about $0.0003. The README states the sample is too small to support an accuracy claim, and marks the project experimental.

Sources:[215] Allie[216] Justin Krup

Line-by-line search over a document

On 27 October, Allie posted a pattern she calls line-by-line search: put an entire document in state, then ask a question for each line or paragraph when you want the most relevant line for a query. She recommends Noul when you expect one dominant answer, and Score when you want to highlight lines by overall relevance. The first shape answers a specific question; the second finds many relevant passages.

In a reply the same day she writes that this family of uses keeps answers sourced from the original content when the need is the exact spot to look. A follow-up note on the same thread says she also uses Jev in a semantic-search browser plugin to decide whether a query wants Noul or Score before running that mode.

Sources:[217] Allie[218] Allie[219] Allie

Intelliprompter: Score over a live transcript

Allie’s 27 October Intelliprompter thread says she shelved the project because of LLM costs, then revived it the day she first learned about Jev. TypeSafe’s cookbook shows the TypeSafe half of that design. The app keeps a cumulative transcript and, on each change, sends the full transcript with one Score question per unchecked talking point. A local threshold of 2.5 out of 5 marks a point covered and drops that question from later requests. The cookbook hard-codes six transcript chunks for the walkthrough.

NetworkChuck posted a 27 October video on decision models and built his own version of Intelliprompter. Allie quote-tweeted that video the same day. The YouTube link is in NetworkChuck’s reply. This page does not score either demo.

Sources:[220] Allie[221] TypeSafe AI[222] Allie[223] NetworkChuck

What this page leaves out

Allie’s “put Jev everywhere” post is advice from someone who builds demos for a living at TypeSafe. It is not a scored evaluation. Agent Tower’s progressive Jev slots come from her posts and screenshot; we have not linked a public repository. jeveryword’s cost table is the author’s small sample at list price. Intelliprompter’s 2.5 coverage gate is a local demo threshold in the cookbook, not a TypeSafe confidence field.

TypeSafe’s coding-agents page still says Jev is not a drop-in replacement for the model behind Claude Code, Cursor, or similar tools. Keep that model for writing and generation. Use Jev where your code needs a typed Choice, Score, or Noul.

Sources:[211] Allie[213] Allie[216] Justin Krup[221] TypeSafe AI[162] TypeSafe AI

Questions about putting Jev everywhere