[Example] Serve Jev's /v1/systemone typed-decision API on top of /v1/score

[luke-edward]

Motivation

Jev (TypeSafe, Sept 15) exposes a typed-decision API: a state plus choice / score / noul questions, answered with probabilities and a confidence, no generated text. vLLM merged an example server for it last week (vllm-project/vllm#57250). SGLang can serve it directly: /v1/score with label_token_ids already returns the next-token distribution over label tokens, and one call scores every question of a request as a batch.

Example only, under examples/runtime/: no core changes, no new dependencies.

cc /v1/score owners @sundar24295s @chanh @fortunecookiee.

Modifications

• examples/runtime/systemone/systemone_server.py: stdlib HTTP proxy. POST /v1/systemone validates the request (400 on schema errors, 502 upstream), builds one prompt per question ending where the answer label starts, sends one /v1/score call, and maps the distributions to Jev's answer shapes (choice argmax + probabilities; score = Σ i·pᵢ + legend; noul = P(Yes); confidence per TypeSafe's published formula). --temperature passes through for calibration.
• README.md: usage, mapping, and differences from the hosted model.
• examples/runtime/README.md: one line.

Accuracy Tests

Tested end to end on an M2 with Qwen/Qwen2.5-0.5B-Instruct (MLX, --mlx-enable-sampling): the request from the Jev docs returns the documented shape, probabilities sum to 1 at 3 and 26 options, 27 options → 400, schema errors → 400.

Speed Tests and Profiling

Not applicable.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests. (Example script; can add one under test/registered/ if wanted.)
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #37502511712
Latest PR Test (Extra): ❌ Run #37502511229
Latest PR Test (AMD ROCm 10): ❌ Run #37502511594

[luke-edward]

@sundar24295s @fortunecookiee, as the /v1/score owners: is an example like this wanted in examples/runtime/? It adds no core code and no dependencies; it only shows /v1/score with label_token_ids serving Jev's typed-decision API shape. If yes, could one of you /tag-and-rerun-ci? If you would rather see it as a cookbook page, or not at all, that is fine too — a one-line answer is enough.

[luke-edward]

@JustinTong0323, examples/ has no code owner and /v1/score lives in entrypoints/openai, which you own. Could you say whether this example is wanted, and /tag-and-rerun-ci if so? It adds no core code and no dependencies; if it belongs in the docs instead, I will move it.

[luke-edward]

@JustinTong0323, could you take a look when you have a chance?

[luke-edward]

Rebased onto upstream/main at 3ed6367; the new head is c867800. Changed-file pre-commit, py_compile, and a synthetic request/response smoke test pass.

[luke-edward]

On c867800, the remaining call-gate / pr-gate reds are only the maintainer-only run-ci label gate; every check-changes job passes.

[luke-edward]

Merged current upstream/main in 468075d without rewriting the reviewed commits. The SystemOne regression test and changed-file pre-commit pass on that head.