Research report · alternatives substitution matrix · 22 Sep 2026
Alternatives substitution matrix refresh
Run run:12eff83e-6150-42af-a3bc-96f68a5c95ce checks the six rows already named on the catalogue page. It separates model identity, interface, local/privacy boundary, workflow role, and trade-offs.
Executive finding: no verified open-weight Jev model was found in this refresh. This is a verification-status finding, not proof that no such model exists. None of the six rows is accepted as Jev-model parity.
Checked 22 September 2026: direct repositories or first-party documentation were checked where available. Unknown means the checked source did not establish the claim; no price, licence, privacy guarantee, or benchmark is inferred.
What each row actually substitutes for
| Row | Model and interface | Runtime / privacy | Workflow fit and limits |
|---|---|---|---|
| system-one-adapter-python Official adapter | Another provider model; preserves much of the Choice/Score/Noul surface, but not Jev weights or parity. Provider/model is required. | Adapter code runs locally; decisions go to OpenAI, Anthropic, or a compatible provider. Retention, telemetry, and direct price are unknown. | Best for provider comparison, pre-waitlist work, or a fallback. MIT is shown by the TypeSafe organisation listing; latency usage fields are not a benchmark. |
| SemIf / OpenJev Independent local experiment | Open-model reproduction of a decision-interface pattern; it explicitly does not reproduce Jev's undisclosed model/training. Full Choice/Score/Noul parity is unknown. | Browser WebGPU demo says no backend; a llama.cpp/GGUF route is documented. Hardware, weights, licence, outbound behaviour, and privacy beyond the demo are unknown. | Use for an experiment with pinned local models. Method timings on a warm RTX 3090 exclude model load and writes; they are not a Jev benchmark. |
| jevlike Research starter | Trainable option scorer with JSONL context/options/label input; drops the Jev SDK contract. | Local CPU, Apple MPS, or CUDA paths are documented. MIT code; downloaded data/models retain their own terms. Network and telemetry are unknown. | Use when you own labelled data and need a bounded scorer. Reported accuracy/calibration/speed experiments are local and explicitly not Jev parity. |
| mini-jev Unverified lead | Current model identity, weights, contract, and Jev parity are unknown. | Runtime, licence, data path, and hardware are unknown because the current direct check was not completed. | Do not select for production on this evidence. Price, latency, evaluation, and maintenance remain unknown. |
| GLiClass Classifier | Different classification family; client takes text, labels, and threshold and returns label/score objects. No Jev primitives or parity. | Local/self-hosted Python and Ray Serve paths are documented, including CUDA examples. Data path, telemetry, storage, and exact licence are unknown in this refresh. | Use for zero-shot/multi-label classification, routing, sentiment, reranking, or triage—not as a calibrated Jev probability service. |
| General LLM + structured output Official provider baseline | Hosted generative model with a JSON Schema contract, not Choice/Score/Noul or Jev state semantics. Schema validity is not semantic correctness. | Provider-hosted baseline; selected model, price, local option, and current data terms are unknown. | Use for generation, extraction, or explanation around a schema when application code owns validation, thresholds, retries, and side effects. JSON mode alone is weaker than Structured Outputs. |
Builder guidance
Choose
Compare providers
Use the adapter for the same typed question set against ordinary LLM providers or as a development fallback. Confirm provider terms and behavior on your workload.
Choose
Run a local experiment
Use SemIf for local probability readouts or jevlike for a trainable scorer when you can pin models, supply data, and measure calibration.
Choose
Classify or generate
Use GLiClass for classifier work and a structured-output LLM when you need explanation or flexible generation around a schema.
Do not infer
Similarity is not parity
Do not treat an adapter, classifier, open local model, or schema-valid response as the Jev model. Keep thresholds, permissions, side effects, and human fallback in code.
Source ledger
- system-one-adapter-python — README, provider examples, typed surface, retries, structured-output modes, and usage/debug fields; the TypeSafe organisation listing supplies the MIT/update observation.
- SemIf, its method, and WebGPU demo — independence/non-parity statement, local/browser/GGUF paths, and timing boundary.
- jevlike — architecture, JSONL contract, CPU/MPS/CUDA paths, evaluation commands, source-reported local experiments, and MIT statement.
- mini-jev — direct check incomplete; no current substantive claim accepted.
- GLiClass and its training source — model IDs, serving/client examples, architecture, and training descriptions.
- OpenAI Structured Outputs and the official Node guide — schema-constrained output, JSON-mode boundary, Responses parsing, and schema restrictions.
The audit separates interface, model, runtime/privacy, and workflow substitution. It does not establish model equivalence, production adoption, or a verified open-weight Jev model.