Page 02 · substitution matrix · checked 22 Sep 2026
Open-source alternatives to Jev
The useful question is not “which project looks like Jev?” It is “which layer can this project replace?” This dated audit separates the model, the System One-like interface, the runtime/privacy boundary, and the workflow job.
Negative finding, unchanged: no verified open-weight Jev model was found in this refresh. That is a verification-status finding, not proof that no such model exists. None of the six rows below is Jev-model parity.
Checked 22 September 2026: the six rows were checked against the direct repository or first-party documentation where available. A source-backed fact is kept separate from an inference about Jev equivalence. The dated research report contains the source ledger and the remaining unknowns.
Read the layers before choosing
Model
Weights are not an interface
The adapter, SemIf, jevlike, GLiClass, mini-jev, and a general LLM are different models or unknowns. A familiar name or output shape does not establish Jev weights, training, calibration, or parity.
Interface / SDK
Contract must be tested
Choice, Score, and Noul semantics, option probabilities, errors, retries, and changing options are separate checks. An adapter can preserve a surface while routing decisions to another provider.
Runtime / privacy
“Local” is not a blanket guarantee
Inspect model downloads, hardware, outbound calls, telemetry, storage, retention, licences, and provider boundaries. The matrix marks only what the checked source actually supports.
Workflow
Replace the job, not the label
A scorer or classifier may replace one bounded step; a structured-output LLM may add generation or explanation. Keep thresholds, permissions, side effects, and semantic checks in your code.
Six-row substitution matrix
| Alternative and evidence | Jev model | System One-like API / SDK | Local execution and privacy | Workflow role and trade-offs |
|---|---|---|---|---|
| system-one-adapter-python Official TypeSafe adapter · direct repository checked |
Different model. The README calls it a drop-in replacement backed by LLM APIs instead of TypeSafe. No Jev weights or Jev-model reproduction claim is accepted. | Closest interface substitute, not parity. It exposes Choice, Score, Noul, system_one, discrete/probability modes, validation, and retries. Native structured output or prompted JSON can be selected. A provider/model is required, so compatibility belongs in an application test. |
Python code can run locally, but the decision call goes to a selected OpenAI, Anthropic, or compatible provider. The README documents provider calls and request/debug history; it does not establish a local model, telemetry policy, retention period, or privacy guarantee. | Choose for: provider comparison, pre-waitlist development, or a fallback evaluation path. The repository has a LICENSE, tests, and provider extras; the TypeSafe organisation listing labels it MIT. Unknown: direct provider price, latency, reliability, calibration, and data terms. A usage latency field is not a benchmark. |
| SemIf / OpenJev Independent project · repository and method/demo pages checked |
Different open-model implementation. The README says it reproduces the interface pattern with open models and does not reproduce Jev’s undisclosed model or training. No Jev parity claim is accepted. | Partial / unknown. It reads typed option probabilities and supports runtime-defined decisions, but this refresh did not verify a complete Choice/Score/Noul contract, errors, retries, or SDK compatibility. |
The browser-only WebGPU demo documents no backend, and a llama.cpp/GGUF path is described. This supports a possible local data path, not a blanket privacy guarantee. Hardware/model downloads, outbound behaviour outside the demo, exact licence, and weight provenance are unknown. | Choose for: an experimental local decision readout when the team can pin and measure an open model. Method timings on a warm RTX 3090 exclude model load and result writes; they are project measurements, not Jev comparisons. Hosted price, maintenance, calibration, and production readiness are unknown. |
| jevlike Research starter · direct README checked |
Different trainable scorer. Each option is represented and scored against context. The README calls it a research starter and says it is not a copy of Jev. | Drops the Jev contract. The input is JSONL with context, options, and a zero-based label; output is option scores/probabilities. No Choice, Score, or Noul SDK is established. |
Local path stated. Quickstart supports CPU, Apple MPS, and CUDA. The default byte encoder trains from scratch; the optional frozen Hugging Face encoder needs the same model at load time. The code is MIT; downloaded data and pretrained models retain their own terms. Network/telemetry behaviour is not stated. | Choose for: a bounded multi-option scorer when you own training data and deployment. The README reports accuracy, calibration error, controls, and local experiments, but says these do not show equal quality with Jev or reproduce TypeSafe’s private training. Hosted price, production support, and independent replication are unknown. |
| mini-jev Retained lead · current direct facts unavailable |
Unknown. No current direct source was successfully checked, so the historical “teaching bench” label is not treated as fact. No Jev parity or open-weight claim is accepted. | Unknown. Choice/Score/Noul behaviour, input/output, logits, retries, and compatibility were not verified on 22 September 2026. | Unknown. Local requirements, weights, licence, model provenance, outbound network behaviour, and data handling were not verified. | Do not choose for production on this evidence. Inspect the repository and runtime before selecting it. Price, latency, evaluation, calibration, maintenance, and operational status are all unknown; the direct URL is retained for follow-up only. |
| GLiClass Open-source classifier · direct repository checked |
Different classifier family. The repository describes a generalist/lightweight text-classification model and names downloadable IDs such as knowledgator/gliclass-edge-v3.0. No Jev model or parity claim is accepted. |
Drops the Jev contract. The documented client takes text, labels, and a threshold and returns label/score objects. It does not establish Choice, Score, Noul, or Jev state semantics. |
Local/self-hostable path stated. The README documents Python use, model loading, Ray Serve, and CPU/GPU-oriented configurations, including CUDA streaming examples. Actual outbound calls, telemetry, storage, and retention are unknown. A LICENSE file is present; its exact licence text was not relied on for this refresh. |
Choose for: zero-shot or multi-label classification, routing, sentiment, reranking, or fact-check triage when a local classifier API is enough. No direct cost, latency, calibration guarantee, Jev comparison, or privacy commitment is established. |
| General LLM + structured output Official OpenAI structured-output baseline |
Different hosted generative model. The official source describes schema-constrained output, not Jev or System One model parity. | Schema contract only. Structured Outputs can make returned objects conform to a developer-supplied JSON Schema; this does not provide Choice/Score/Noul semantics, Jev state handling, calibrated decision probabilities, or semantic correctness. JSON mode alone does not guarantee a schema. |
The cited baseline is a hosted provider API. Local execution, current retention/training settings, and a specific model/price are unknown because no deployment was selected for this row. | Choose for: extraction, generation, explanation, or a bounded decision step when your code owns schema validation, thresholds, retries, and semantic safety checks. The provider-reported schema-following evaluation is not a Jev benchmark. Current price, latency, failure rate, data terms, and semantic accuracy are unknown for the unspecified deployment. |
Choose / do not choose
Choose
Compare providers
Use system-one-adapter-python when you need the typed question set against ordinary LLM providers or a development fallback before Jev access. Confirm provider terms, cost, and behaviour on your workload.
Choose
Experiment locally
Use SemIf for an experimental local probability readout, or jevlike when you own labelled data and want a trainable scorer. Pin the model and licence; test calibration, hardware, and network behaviour.
Choose
Classify, not emulate Jev
Use GLiClass when zero-shot or multi-label classification is the job and a local/self-hosted classifier API is enough. Call its score a classifier score, not a calibrated Jev probability.
Choose
Generate around a schema
Use a structured-output LLM when the task needs explanation or flexible generation and your application can validate meaning, retries, thresholds, and side effects.
Do not choose by resemblance: do not choose the adapter, SemIf, jevlike, GLiClass, or a generic LLM because its interface looks similar to Jev; none is verified as the Jev model. Do not use any local-looking row for sensitive production data until the full data path, outbound calls, storage, licence, and operational controls are checked. Do not promote mini-jev beyond an unverified lead.
Keep the application boundary: ordinary code owns thresholds, permissions, side effects, explanations, and human fallback. A score or schema-valid object can still be a wrong or over-confident decision.
Source ledger and next check
Direct sources checked or attempted on 22 September 2026
- system-one-adapter-python repository — README, file tree, provider examples, API/response description; the TypeSafe organisation listing supplies the organisation-level MIT/update observation.
- SemIf / OpenJev repository, its method, and local WebGPU demo — independence/non-parity statement, local/browser/GGUF descriptions, and timing boundary.
- jevlike repository and README — architecture, JSONL contract, device options, evaluation commands, local results, and MIT statement.
- mini-jev repository — direct access was not completed; no current substantive claim is accepted.
- GLiClass repository and training source — model IDs, client/serving examples, architecture and training descriptions.
- OpenAI Structured Outputs announcement and official Node guide — schema-constrained output, JSON-mode boundary, Responses parsing, and schema restrictions.
The direct audit improves the boundary between an interface adapter, an open local experiment, a trainable scorer, a classifier, and a generic structured-output provider. It does not establish model equivalence, production adoption, or a verified open-weight Jev model. Report: alternatives substitution refresh; run run:12eff83e-6150-42af-a3bc-96f68a5c95ce.