ifReflex.cpp
llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
Overview
ifreflex.cpp is native C++ inference for System One decision models, built on llama.cpp. You send state and a map of typed questions, and it answers every one in a single prefill pass, returning the full probability distribution over your options with zero generated tokens. Answers are read directly from next-token logits. The server speaks the same POST /v1/systemone contract as TypeSafe's Jev, so Jev clients can point at a local server unchanged.
One binary covers reflex, SemIf, RWKV-Jev and any stock instruct GGUF, with pre-converted Qwen GGUFs from 0.8B up. It targets people who want Jev-style decisions on a personal computer without a GPU or a hosted API. The state prefix is decoded once and each question (and each option permutation for letter readouts) branches from it; a permutations option averages over option orders. It supports choice, score and noul, bearer-key auth, CORS restriction, batch JSONL input and raw uncalibrated output.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should handle this? | billing · technical support · sales |
| Yes / no | Is a refund requested? | yes · no |
How it works
- Start the server with a GGUF model and prompt format
- Client posts state and questions to /v1/systemone
- State prefix is prefilled once; each question branches from it
- Probabilities are read from logits (optionally averaged over permutations)
- Response matches Jev's answers and usage format
Strengths
- Drop-in local replacement for the Jev API contract
- No generation and no GPU required
- Works with many open models via GGUF
- Data never leaves the machine
Caveats
- Accuracy depends on the chosen open model, not Jev
- Raw probabilities may be uncalibrated without the built-in calibration
Getting started
`build/ifreflex-cli --server --port 8080 --model Qwen3.5-4B-Q4_K_M.gguf --prompt reflex_markdown --permutations 2`, then POST to http://127.0.0.1:8080/v1/systemone.At a glance
- Type
- Project
- Status
- beta
- Language
- C++
- Runtime
- local HTTP server, CLI
- Has code
- yes
Use cases and integrations
How curated lists describe it
Any LLM can serve as SystemOne decision model on your Personal Computer.
wh000wh000/awesome-jev-live
Frequently asked questions
Does ifreflex.cpp use Jev?
No, it runs open models locally but exposes the same /v1/systemone contract so Jev clients work against it.
Does it need a GPU?
No. It runs prefill-only inference on CPU via llama.cpp.
Which models can I use?
reflex, SemIf, RWKV-Jev and any stock instruct GGUF, such as the listed Qwen models.
Related projects
MLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.
ProjectprototypePythonLocal server (xks) implementing Jev's /v1/systemone contract on an open-weight LLM, accelerated with Xeon Phi cards.
ProjectresearchRustEducational C11 library and CLI that turns DiffusionGemma 26B into a CPU-first OpenJev-style decision engine reading candidate logits.
ProjectresearchCExperimental engine that makes an existing LLM do Jev-like classification and scoring by reusing context and building JSON.
ProjectresearchPythonReads typed decisions from one forward pass of PrismML's 1.75-bit Bonsai 27B under llama.cpp, using the OpenJev/SemIf logit method.
ProjectresearchPythonJev-class decision server on an RWKV-7 1B checkpoint that reads typed answers from a constrained softmax behind /v1/systemone.
ProjectresearchPython