CommunityJev Practices & Patterns
Jev is the fish at the poker table
An independent poker eval against TexasSolver: Jev matches solver top actions on easy spots but shoves the nuts into traps and needs explicit hand-rank text to fold to shown flushes.
- Format
- Article
- Published by
- Community
- Author
- backnotprop
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Single solved Q♠9♦4♠ tree with Python-computed pot odds so Jev never does arithmetic.
- Undecomposed action choice hit 63% solver agreement; decomposed judgment pipelines scored 33–44% on contested spots.
- Revealing villain cards without a named hand_rank still produced confident shoves into nut flushes.
Watch out
One board, top-action agreement not EV loss; run-to-run probabilities swing a few points.
Questions about this entry
Short answers drawn from the linked project or page.
Mostly pattern-matching. It held the nuts, K-Q-J-T-9, on a Q spade 9 diamond 4 spade flop with a jack on the turn, and shoved sixteen runs out of sixteen where the solver checks one hundred percent of the time.
Weak on its own. It is one solved board, top-action agreement rather than expected-value loss, and probabilities swing a few points run to run. The 33 to 44 percent scores from decomposed pipelines are the sharper warning.
Only partly. Shown villain cards still produced confident shoves around sixty percent of the time unless the state also spelled out a hand_rank such as nut flush. The blind spot is named hand strength, not reasoning.
Pot odds, stack to pot, hand rank, and outs were computed in Python and passed in as a state object, so the model was never asked to do arithmetic. Worth copying: it isolates the failure surface to the choice itself.
One flop was solved with TexasSolver: heads-up, 100bb deep, LJ opens and BTN calls, flop Q spade 9 diamond 4 spade. The solve took eight minutes, 7.6 GB and reached 0.59% exploitability, giving the correct strategy for every hand on every turn card.
Not the top-action rate. Five phrasings over 30 spots, including value-laden wording and no option labels at all, matched the solver on the same 63% of spots. Value-laden verbs did push Jev toward bigger bets, which is why the final prompt carries no framing.
A few points. The trap-spot shove measured 59 to 65% across sixteen runs, and the flush-board shove came in around 60% across sixteen. That variance is why the author treats the 63% agreement as top-action on one board rather than expected-value loss.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Poker eval: Jev shoves the nuts into trap lines unless hand ranks are spelled out.
2 pointsTypesafe's Jev is the fish at the poker tableRead the thread on Hacker NewsRelated terms
Glossary definitions related to this entry.
More Jev resources like this
253GitHub stars
Independent cross-model benchmark for Jev-class decision models, running 534 frozen cases per complete entrant with scoring code and a four-axis score of accuracy, calibration, latency and cost.
Jev Practices & Patterns#community#python#benchmarks
Community
fstandhartinger/jevbenchPython
fstandhartinger/jevbenchPython10GitHub stars
An open, local-first alternative to Jev that reproduces the Choice, Score, and Noul interface on commodity zero-shot NLI models with temperature scaling and conformal abstention, and runs offline with no API key.
Jev Practices & Patterns#community#python#calibration
Community
rupeshpoojary9/poorjevPython
rupeshpoojary9/poorjevPython0GitHub stars
A pre-registered, independently rescored audit of Jev's calibration and type safety: calibrated on CLINC150 (ECE 0.0204), systematically overconfident on Banking77 (ECE 0.0936), and zero type errors across 8,576 responses.
Jev Practices & Patterns#community#python#calibration
Community
jourdanlabs/assay-001Python
jourdanlabs/assay-001PythonAn independent trial running Jev on 24 Norwegian resource-tax hearing responses with eleven typed questions each, then comparing agreement, calibration, latency and cost against DeepSeek V4.1 Flash.
Jev Practices & PatternsArticle#community#evaluation#calibration
Community
8.9kGitHub stars
Apache-licensed, locally runnable Jev-style decision models (0.8B, 4B, 9B on Qwen3.5) with released weights, training code, a System One-compatible server, frozen eval suites, and a playground.
Jev Practices & Patterns#community#python#open-models
Community
jaredpalmer/kevPython
jaredpalmer/kevPython1.1kGitHub stars
An independent open reproduction of the System One model class: Qwen3.5-based 2B and 35B mixture-of-experts models that return typed Choice, Score, and Noul probabilities in one forward pass, with nothing distilled from Jev.
Jev Practices & Patterns#community#python#open-models
Community
Mapika/deciderPython
Mapika/deciderPythonFrom the Jev guides
Original write-ups on what this entry is for.
Base decision checkpoints sit near chance on purpose. The typed-decisions loop, the corpora behind the students, and the temperature fit that turns probabilities into thresholds.
Jev Practices & Patterns#fine-tuning#calibration#open-models
Read article
TypeSafe publishes evals, but the numbers that matter are the ones measured on your task. A tour of the independent benchmarks, field reports, and the recurring finding that question design dominates.
Jev Practices & Patterns#benchmarks#evaluation#calibration
Read article
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
From the Jev community
Posts from builders shipping with Jev right now.