User repository on GitHub
Jev-style calibrated decision model (Choice/Score/Noul) on Qwen3.5-0.8B
- Forks
- 1
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Practices & Patterns
Qwen Rlcd
An inference-only prototype of a Jev-style decision model on a sub-billion-parameter hybrid backbone, returning typed probability distributions in one forward pass, with its architecture treated as an explicit guess.
- Stars
- 5
- Language
- Python
- Published by
- Community
- Author
- shamazharikh
- Last verified
- Oct 2026
About this project
A prototype of a "System One" decision model, inspired by TypeSafe AI's Jev. The model takes a state (context) and a set of typed questions. In a single forward pass it returns calibrated probability distributions over a fixed set of answers, with no text generation. The implementation targets Qwen/Qwen3.5-0.8B-Base. See PLAN.md for milestones. Current scope is inference only; training and fine-tuning (PLAN.md M2) are on hold.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Runs every question and answer branch through one prefix-forked cache, measured at 8 to 19 times faster than a full forward per branch.
- A tree attention mask cannot isolate sibling branches inside the hybrid model's recurrent layers, which is why prefix-forking is used instead.
- Choice options are always rendered in sorted key order, and tests assert question and option order invariance below 1e-5.
- Three scorers are implemented: letter logits, answer likelihood with optional PMI, and linear heads at a read token that stay untrained.
Watch out
This is not a Jev reimplementation and the README does not pretend otherwise: TypeSafe has not published Jev's architecture, so only the clearly labeled public-facts section comes from TypeSafe's documentation and everything else is the author's hypothesis about how such a model could work. Scope is inference only. Training with adapters and heads is on hold, the read-token heads are untrained until then, and the zero-shot baselines are CPU runs at n=100 where the README puts the confidence interval on accuracy at about plus or minus nine points. Three of the five milestones are unstarted, including temperature fitting, the calibration pass and serving. The measured speedups come from single-GPU runs on Turing cards, where the linear-attention kernel dominates rather than the matrix multiplications, and the author's own hardware notes include a GPU dropping off the bus mid-benchmark. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
A model that resembles one, by the author's own framing. The README separates public facts, which come from TypeSafe's documentation and launch coverage, from everything else, which it labels as its own hypothesis about how such a model could work, and states plainly that TypeSafe has not published Jev's architecture. What it does share is the contract: a state plus typed questions in, calibrated probability distributions over a fixed answer set out, in a single forward pass with no text generation.
Branch count is handled by prefix-forking: the state is prefilled once, its cache is copied along the batch dimension with one row per branch, and every question and answer branch runs as a single right-padded batch at the same position offset. That is what isolates branches in every layer type by construction and why order cannot leak between questions. Without chunking, memory grows with branches times state length and long states run out of memory on an 11 GB card, so a chunk size caps rows per forward and makes peak memory flat in branch count.
Not yet, and the README's own tables make that clear. The baselines are CPU runs at n=100 where the stated confidence interval on accuracy is about plus or minus nine points. Choice scoring reads 0.60 on ARC-Challenge with the letter scorer, and the five-level score task is near chance with the best of the three scorers at 0.33. The linear heads are untrained until training resumes, temperature fitting and the stress suite have not started, and serving is not started either.
Because this backbone is hybrid: 24 layers split into 18 linear-attention layers and six gated full-attention layers, and a tree mask cannot isolate sibling branches inside recurrent layers. Prefix-forking instead keeps the KV cache for the attention layers plus the convolution and recurrent state for the linear-attention layers, so isolation holds in every layer type. Equivalence against uncached sequential forwards is pinned by tests, and negative controls that zero the forked recurrent state are expected to move the reads substantially.
Prefill the state once and keep its cache, copy that cache along the batch dimension one row per branch, then run every question-and-answer branch as one right-padded batch at the same position offset and read the hidden state at each branch's last real token. Branches are isolated by construction, so results cannot depend on question or option order.
The current scope is inference only; training and fine-tuning are on hold. One reason is technical: the cache updates the DeltaNet state in place, so backward through a forked cache fails, which matters for training but not for the inference work that has been done.
The dev box is two RTX 2080 Ti cards with 11 GB, and only one GPU job at a time should run because a dual-GPU benchmark made one card fall off the bus. Memory grows with branch count times state length, so 2048 tokens by 50 branches runs out of memory without chunking; forward_branches chunk_size caps rows per forward and keeps peak memory flat in the branch count.
More Jev resources like this
16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
erendikmenn/jev-rag-benchmarkPython
erendikmenn/jev-rag-benchmarkPython24GitHub stars
Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPython22GitHub stars
Offline evaluator that checks whether classifier confidences match reality and converts mistake costs into a deployable threshold, emitting an HTML report plus machine-readable threshold files.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rlaope/jevalPython
rlaope/jevalPython37GitHub stars
Minimal Python recipe that turns any causal LLM into a Jev-style classifier by prefilling choice_index and constraining sampling to legal indexes in one forward pass.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
sgoedecke/system-onePython
sgoedecke/system-onePython39GitHub stars
Open replica of the typed decision API on openly available base models, answering choice, score, and true-probability questions in one forward pass with no decoding.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
mithalouni/system-one-openPython
mithalouni/system-one-openPython34GitHub stars
Curated catalog of community Jev applications grouped by use, where each case explains the mechanism and states what the public evidence does and does not prove.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
JackZeng/Jev_appsPython
JackZeng/Jev_appsPythonFrom the Jev guides
Original write-ups on what this entry is for.
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
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.