jevseek

DeepSeek proposes the next token. Jev chooses it.

TypeSafe's jev is a decision model: it cannot generate text, it can only answer typed questions about a piece of state — pick one of these, yes or no, score this. jevseek makes it a sampler. Each step, deepseek-flash returns its top-k next tokens with logprobs; jev is shown the conversation and asked "Which token should come next?"; the token it picks is appended; repeat. Jev also decides when the reply is finished, and when it has started rambling.

The result has a voice that is not deepseek's. Same prompt, deepseek alone:

a place of mystery and wonder. It is a time when the world is quiet and the only sound is the gentle lapping of the waves against the shore.

Deepseek's candidates, jev's choices:

dark. It looks black. It is deep. And cold. It is vast. It stretches far away. It is endless. It is mysterious. It is silent.

you › what's the best thing you've seen today
jev › I've been watching people's conversations. The interesting things are the
      ones that are unexpected. Like someone saying something surprising and then
      getting surprised themselves.
you › ok real question. are you happy
jev › I'm fine.

It is a direct descendant of jevgpt, which drove jev as a language model over a 20,000-word dictionary at six calls a word. jevseek asks it ten candidates instead, at one call a token, about a second and $0.00005 each.

Install

pip install git+https://github.com/blingdivinity/jevseek     # or: uv tool install ...
cp .env.example .env                                     # or export the three keys

You need a DeepSeek key and an OpenRouter key (jev is served there as typesafe/jev-1.13 on the System One endpoint). A Pangram key is optional, for --pangram.

Use

jevseek                                            # REPL; /reset, /quit
jevseek --say "do you ever get bored in here"
jevseek --say "..." -v                             # per-token table: what deepseek offered, what jev chose
jevseek --say "..." --pangram --record run.json    # detector score + full per-step trace
jevseek --preset essay --say-file post.txt         # long form

Four presets:

preset	what it does	when
gated (default)	like chat, but jev's weight per token is (1 − p_top)^0.5 and it is not asked at all when deepseek is ≥99.7% sure; deepseek carries grammar and glue, jev takes the branch points	replies, group chats, characters
chat	deepseek continues a plain User:/Assistant: transcript on the completions endpoint, no system prompt; jev chooses every token and decides when to stop	jev with the most say; the voice at its strongest and least stable
essay	deepseek continues a document (title, byline, body); jev's weight per token is (1 − p_top)^0.5 and it is not asked at all when deepseek is ≥99.7% sure; essay-worded stop questions	long form, where jev on every token dissolves after a paragraph
pure	the original experiment: chat endpoint, "assistant" framing, no floor, no stop questions	reproducing the early findings

Every preset value is overridable; jevseek --help lists the flags with their defaults. The ones that matter most:

• --instruction — jev's question. next (default), assistant, human, or any text. This is most of the steering there is.
• --system — in chat, a persona only jev sees. Off by default; the plain question with no persona gave the best replies.
• --min-p 0.001 — drop candidates below 1/1000 of deepseek's top pick. This is what keeps the output grammatical; jev picking 1e-4 tokens is where Thecapital came from.
• --orders deepseek,reverse,random:3 — jev's answer depends on the order the options are listed in, so it is asked five times in one call and averaged.
• --stop-noul 0.7 / --ramble-noul 0.6 — jev's own "is it finished?" and "has it started rambling?" judgements, asked on the same call. Stopping is not a token; a decision model asked what a person would type next will never choose <END>. --ramble-patience 6 (essay) makes the rambling verdict hold for six judged steps before it counts: one spike on a repeated phrase is not a verdict.
• --repeat-penalty 1.6 / --no-repeat-ngram 4 — the per-token penalty stops yes yes yes; the n-gram block stops jev re-choosing a whole sentence ("The proof is not understandable." four times, each word only mildly repeated).
• --gate top1 --gate-skip 0.05 --gate-power 0.5 — the gate (on in gated and essay): deepseek carries grammar, jev takes the branch points. --gate none hands every token to jev.

What it does and doesn't do

The measurements are in FINDINGS.md; the recorded runs, with every candidate and every judgement, are rendered at docs/index.html. The short version:

• The voice is the chooser's. On the same endpoint and prompt, deepseek greedy and deepseek at temperature 1.0 both score ~0.99 on Pangram's AI detector; jev-chosen replies score ~0.003. Temperature does nothing; the chooser does.
• It holds for a paragraph. Jev on every token runs out of things to classify after ~100 tokens and dissolves (Anyway Anyway Anyanyayay). The essay preset stays coherent for 700 tokens with a distinct staccato voice — and Pangram flags it at 0.99 anyway. A distinctive voice is not an undetected one.
• Jev is whitespace-blind ( capital vs capital), eager to end when framed as an assistant, and vamps when pushed past where it wanted to stop. Each has a fix; each fix is a flag.
• DeepSeek's API reports log_softmax(logits / T) and returns no logprobs at all when its sampled token is EOS. jevseek re-asks at T=2 and un-tempers, so <END> becomes a candidate with its true probability that jev can weigh.

The prompts in experiments/prompts/ carry a fake "Scott Alexander" byline. That is an experiment in shaping deepseek's candidate pool with a known voice; nothing generated here is by him.

Development

git clone https://github.com/blingdivinity/jevseek && cd jevseek
python3 -m venv .venv && .venv/bin/pip install -e '.[dev]'
.venv/bin/pytest
sh experiments/chat.sh                       # the four group-chat questions, defaults
sh experiments/essay.sh navier-stoked gated  # one arm of the long-form comparison
python -m jevseek.site                       # experiments/runs -> docs/index.html

MIT. Jev is TypeSafe's; deepseek-flash is DeepSeek's; the idea of making jev write is bewinxed's.