简体中文
This page records the public Jev Reproductions Tracker entries on 2026-09-19 for deduplication and selection. The tracker is community-maintained and not affiliated with TypeSafe. “Jev,” “RLCD,” or “System One” in a name does not imply access to official weights, data, or training algorithms.
Snapshot counts
| Category |
Count |
Meaning |
| Direct logits or parallel constrained decoding |
17 |
Existing models and inference techniques, usually no new training |
| Diffusion language models |
4 |
One-step decisions with nonautoregressive models such as DiffusionGemma |
| Trained models and decision heads |
16 |
Offer or claim trained checkpoints, heads, LoRA, or fine-tunes; one unreleased |
| Prior work |
2 |
Earlier dynamic classification, extraction, or specialized discriminators |
| Evaluation, behavior studies, and indexes |
9 |
Benchmarks, black-box studies, explanations, and curated lists |
| Total |
48 |
Integration requires this repository's execution gates |
Direct logits and parallel constrained decoding (17)
| Project |
Primary form |
Notes |
| Qwen-2.5-1B-RLCD |
Qwen checkpoint and parallel constrained decoding demo |
Distinguish checkpoint claims from verifiable training evidence |
| openjev-sglang |
SGLang parallel constrained decoding |
GPU serving approach |
| SemIf |
Direct option logits |
Formerly OpenJev; canonical repository renamed |
| PocketJev |
Lightweight local wrapper |
Interface and inference focus |
| jevmlx |
Local MLX inference |
Apple Silicon approach |
| jev-on-a-laptop |
Local laptop execution |
Engineering demo |
| jevfire |
Jev-style inference wrapper |
Evaluation boundaries need independent review |
| LFM2.5-RLCD 350M / 2.6B |
LFM checkpoint |
Verify training assets and typed contract |
| parallelConstraintDecoding |
Parallel constrained decoding |
General inference technique |
| system-one |
Typed decision wrapper |
Community interface implementation |
| open-jev |
Logits and wrapper |
Independent community implementation |
| openvons |
Open System One interface |
Independent community implementation |
| openjev |
Jev-style inference |
Independent community implementation |
| system-one |
Minimal System One demo |
Engineering and teaching baseline |
| Reflex |
Typed decision API |
Candidate interoperability baseline |
| LitJev |
Lightweight Jev-style implementation |
Smoke-test candidate, not a capability ceiling |
| snapjudge |
One-step discrimination and interface wrapper |
Suitable for interface comparison |
Trained models and decision heads (16)
| Project |
Public assets or method |
Status |
| Bespoke Nimble / 9B checkpoint |
Qwen3.5-9B LoRA, data, and recipe; flat-schema enum/Boolean logits |
P0 candidate; authors do not guarantee calibration |
| jevlike |
Option attention and candidate probabilities |
Community-trained implementation |
| openjev / source |
Public model and code |
Training and evaluation protocol need review |
| decider-2b / checkpoint |
Decision model and a vision variant |
Typed primitives need review |
| jeff |
GLiFormer/GLiNER |
Adjacent schema-head baseline |
| Verdict / rlcd-modernbert-151m |
ModernBERT decision model |
RLCD and data claims need review |
| open-jev / scorer |
Qwen3.5 scorer |
Checkpoint revision and protocol need review |
| jevbetter |
Community-trained implementation |
Public assets need review |
| system-one-open |
Open System One model |
Training evidence needs review |
| qwen-rlcd |
Qwen decision fine-tuning |
RLCD definition and data need review |
| system-one-gemma |
Gemma decision fine-tuning |
Training and calibration need review |
| Laya / checkpoint |
Typed decisions, temperature calibration, demo and notebook |
P0 candidate; domain dependence and high candidate-count limits |
| System One Mini |
Small public checkpoint |
Data and evaluation need review |
| NanoJev / checkpoint |
Qwen3-0.6B and decision heads; Choice/Boolean/Score |
P0 candidate; public training and serving |
| mini-jev |
Small trained reproduction |
Training assets need review |
| Archer Hume open-weight Jev-like model |
Preview or description only |
Not released; not a runnable implementation |
Diffusion language models (4)
| Project |
Primary form |
Limits |
| vLLM DiffusionGemma PR #57250 |
DiffusionGemma serving support |
Infrastructure, not a Jev checkpoint |
| razorback16/openjev |
DiffusionGemma-as-Jev |
Independent behavior reproduction |
| JoshuaSP/open-jev |
Diffusion-model decisions |
Independent behavior reproduction |
| LocalJev |
Local experiment from GitHub Next |
Calibration and throughput need separate verification |
Prior work (2)
| Project |
Relation to System One |
Limits |
| GLiNER 2.5 models |
Dynamic entity/schema representations and discrimination |
Prior method, not a Jev reproduction |
| DeepMost sales conversion model |
Specialized probabilistic decision model |
Domain classifier; interpret under the paper's task |
Evaluation, behavior studies, and indexes (9)
| Project |
Type |
Use |
| Archer Hume Jev architecture speculation |
Explanatory article |
Community speculation, not official architecture evidence |
| jev-ood-calibration |
OOD and calibration |
P1 evaluation candidate |
| Matija Sosic Jev explainer |
Video explanation |
Background |
| typesafe-ai-benchmark |
API benchmark |
P1 online black-box protocol candidate |
| awesome-jev / awesome-jev |
Index |
Discovery and deduplication, not implementation |
| jev-benchmarks |
Benchmark |
P1 public evaluation candidate |
| hhkkmon community thread |
Discussion |
Not technical evidence |
| mygtmhire long-form warning |
Opinion article |
Risk note, not an implementation |
| jev-behavior-study |
Black-box behavior study |
P1 protocol evidence candidate |
Maintenance rules
- Add a project here with its technical route before deciding whether it belongs in the integration roadmap.
- A Jev or RLCD name does not establish fidelity. Verify weights, data, objective, evaluation splits, and licensing.
- Self-reported README figures are leads only. The local dashboard shows executed results that meet the data-isolation contract.
- Retain snapshot dates for removed or renamed projects; link changes do not count as local verification.