Should you adopt it or move to SGLang?
The openjev-sglang maintainer recommends SGLang's native decisions API instead of this experiment. That is the strongest reason to evaluate SGLang first for a new service. The supplied GitHub metadata also marks openjev-sglang as archived. Existing users should check the native API's compatibility details before changing their endpoint.
This repository still has a useful, narrower role: explaining how a Jev-compatible API can turn model scores into bounded decisions. Its documented implementation uses Qwen3.6-35B-A3B with SGLang 0.5.19. Noul returns a yes probability, Choice returns the most probable option and its distribution, and Score returns a probability-weighted level index.
What our sandbox established
Our run on October 5, 2026 used commit bf6a53b. The matching sandbox record specifies 3 CPUs, 8 GB of memory, and the ghcr.io/astral-sh/uv:python3.12-bookworm image. Its install tail identifies Python 3.12.12.
Our harness records each command's success flag, exit code, elapsed time, timeout status, and available log tail. Here, install and build returned exit code 0; the test command returned exit code 1. The attached harness durations are 34 seconds for installation, 6 seconds for build, and 13 seconds for tests. These are command outcomes, not evidence that a deployed decision service worked.
The stored parser summaries report 48 passed, 3 failed, and 26 errors, with a total field of 77 and zero skipped tests. They are not verified exhaustive counts of unique tests. We have not performed a new execution or reconciled a complete log. The raw pytest footer separately reports one deselected test and a pytest duration of 5.40 seconds; that footer duration differs from the harness's elapsed command time.
The three named backend failures concern sibling cancellation, the global branch concurrency limit, and sharing a selected-logprob batch between warmup and scoring. Each says async functions are not natively supported. The tail also warns that asyncio_mode is an unknown configuration option. An API test reports an async fixture without a handling plugin; other API entries show AssertionError without enough traceback to diagnose their causes. These observations warrant checking the test environment before interpreting the result as broken inference behavior.
The build tail is empty. Its zero exit code does not establish packaging correctness, CLI behavior, GPU execution, or Modal deployment. No live inference result is attached.
The recorded pip-audit advisory total is zero. Scan coverage is unknown because no scannedPackages value is supplied; the separate packages field is zero and does not establish dependency coverage. This is not a security assessment.
The supplied testing methodology explains the publisher's command recording and sandbox limits. That page was fetched on October 31, 2026; it does not establish which methodology version governed our October 5 run.
What the decision probabilities mean
The server warms a common prefix, then requests one token for each question and reads the answer labels' log probabilities. It ignores the sampled token and normalizes the requested label scores. The documented implementation makes N+1 one-token calls for N questions, including the warmup.
That mechanism avoids generating and parsing a prose answer, but it does not establish decision accuracy or a cost advantage for your workload. The README says option order and prompt wording affect the probabilities. Its confidence value describes the distribution's concentration rather than calibrated correctness.
The documented chat state accepts text, not image, audio, or video content. Cache reuse is opportunistic. SGLang 0.5.19's Rust frontend omits cache counts, so the API leaves its cache-count header absent rather than reporting zero. Neither the README's live-deployment observations nor its smoke-test description are measurements from our sandbox.
What deployment still requires
Local dependency installation covers the API, deployment tools, and tests. Real inference requires a suitable SGLang backend with matching model and tokenizer revisions, selected-token log probabilities, radix caching, and enough context capacity.
The included Modal profile assigns one B200 per container. Initial startup downloads model weights and compiles or captures kernels; persistent volumes retain weights and compilation caches, while CUDA graph capture still runs at startup. A server waking from zero returns 503 while starting.
The sample Modal endpoint is intentionally unauthenticated and has no explicit container cap. API and backend bearer keys are optional. Before exposing that configuration, choose authentication and capacity settings appropriate to your service. Our recorded scan found a tests directory, no CI workflow files, and no Dockerfile; that scan does not establish the absence of checks outside the checkout.
Which alternative answers your requirement?
SGLang's decision-model documentation describes both a native decisions route and a System One compatible route. This makes it the relevant replacement candidate for this repository. The documentation says to use a nightly build until a release contains the endpoint, so verify availability in the version you intend to deploy.
Compatibility still needs testing. The documented System One route has exceptions, including unsupported SDK model listing and response request IDs, different usage accounting, and confidence thresholds that may not transfer between providers. Validate representative requests and thresholds before moving an existing client.
vLLM documents an OpenAI-compatible API server. Consider it when your requirement is general model inference and serving. The supplied vLLM extract does not establish a Jev-compatible decisions endpoint, so it does not support treating vLLM as a direct replacement for this API. Neither alternative was executed in our recorded run.
How to read the ratings
These editorial ratings use a five-point scale: 1 indicates a substantial constraint, 2 requires significant operator work or has limited supporting evidence, 3 indicates adequate evidence for ordinary use, 4 indicates strong evidence with minor gaps, and 5 indicates unusually complete evidence for the dimension assessed. They are judgments, not benchmark results.
Setup ease remains 2 because installation completed but inference needs a separate backend and the supplied deployment requires GPU infrastructure. Documentation remains 5 for its detailed explanation of requests, probability semantics, caching, configuration, and deployment limits; that rating does not verify every documented behavior.
Community remains 2 because the supplied snapshot provides limited evidence of sustained support. Stars and issue counts cannot establish adoption or response quality. Maturity remains 1 because the project calls itself an early experiment, recommends another implementation, is marked archived, and did not complete the recorded test command successfully.

