Organization repository on GitHub
Production OpenJev-compatible API for SystemOne deterministic classification
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Systemone — superagent-ai
Self-hostable SystemOne-compatible API serving the frozen AutoJev Security V2.3 release as security-one, with shared-state parallel questions and bounded admission.
- Stars
- 1
- Language
- Python
- Published by
- Community
- Author
- superagent-ai
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Systemone — codesoda.
About this project
Production API for the Security-One deterministic structured-classification model. Its public wire contract is compatible with OpenJev's POST /v1/systemone: one shared state, any mix of noul, choice, and score questions, and answers returned under the caller's question IDs. The production deployment serves the verified AutoJev Security V2.3 release as security-one: frozen Qwen3.8-27B BF16 checkpoint with an exact 255-way AutoJev readout export; zero class flips and zero logit drift in the 64-example export gate;
Quoted from the project's README. Read the full README.
Highlights
- POST /v1/systemone with string, object, or array state evaluated in parallel per question.
- Noul probabilities, Choice distributions with entropy confidence, and Score levels with legend.
- Extra endpoints for models, limits, health, liveness, docs, and OpenAPI reference.
- Smoke command covers all three answer types before moving customer traffic.
Watch out
Paused service redeploys to Modal; local runs need a running SGLang backend and research packs are evaluations only. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Yes for the wire shape: one shared state with any mix of noul, choice, and score questions, answers keyed by caller IDs. Model aliases such as security-one-latest, systemone, and jev-latest select the same loaded release, defaulting to security-one when omitted.
A frozen Qwen3.8-27B BF16 checkpoint with an exact 255-way readout export, zero class flips in the 64-example gate, fixed 0.14527332485151376 temperature, SGLang logprobs, prefix reuse, and 65,536-token branch context with no silent truncation.
Set SYSTEMONE_MODEL_PATH and SYSTEMONE_BACKEND_URL, start with uv run systemone serve, and keep SYSTEMONE_API_KEYS non-empty. Local auth is disabled only when the key list is empty, while the Modal deployment refuses to start without one.
One warm B200 container capped at one for spend, keys compared in constant time and never logged, state never logged, overload as 503 with Retry-After 1, plus request IDs and cache diagnostics without customer data.
noul is binary and returns the probability of true as noul. choice returns the highest-probability option, the full distribution and a normalized-entropy confidence. score returns the expected zero-based level, a legend, the full distribution and confidence. Every question sees the same shared state and is evaluated in parallel.
Overload, timeout and backend failures return 503 with Retry-After: 1. Request bodies, question counts, option counts, per-branch context, total tokens, concurrency and deadlines are bounded before or during admission, and each response includes x-systemone-request-id, x-systemone-model, cache diagnostics and Server-Timing.
Keys are compared in constant time and never logged, and state, questions and prompts are never written to application logs. Multiple early-user keys can be supplied comma-separated, the Modal deployment refuses to start without at least one key, and the project recommends issuing one key per customer so individual keys can be revoked.
More Jev resources like this
6GitHub stars
Open-weight 4B decision model on Qwen3.5-4B-Base serving a Jev-compatible /v1/systemone API, returning per-option probabilities with adjustable thinking.
Jev Agents & MCP#community#github
Community
org2AI/wald-4bPython
org2AI/wald-4bPython1GitHub stars
Demo of the LangChain Jev harness pattern in LangGraph: a fast classifier routes events and gates a risky tool call, leaving open-ended reasoning to a chat LLM. Entry is POST /events; output lands in JSONL files.
Jev Agents & MCP#community#github
1GitHub stars
Portable Python client bringing typed Jev judgments (noul, choice, score) to coding-agent harnesses over MCP or shell, with a shared budget ledger plus harness and memory-system recipes.
Jev Agents & MCP#community#github
Community
Coding-Dev-Tools/jev-decisionPython
Coding-Dev-Tools/jev-decisionPython1GitHub stars
Engineering playbook and skill for placing Jev typed judgments at bounded seams while keeping validation, permissions, persistence, and policy in deterministic code.
Jev Agents & MCP#community#github
1GitHub stars
Standalone Agent-Workflow adapter that projects bounded redacted evidence into typed Jev choice, noul, and score questions and returns secret-free advisory receipts.
Jev Agents & MCP#community#github
0GitHub stars
Code review demo pairing Jev per-rule yes/no checks with a Meko datapack that stores guidelines, rulings, and decision traces, exempting only the ruled case on re-review.
Jev Agents & MCP#community#github
From the Jev community
Posts from builders shipping with Jev right now.