jev-agent-design-with-topk-logits-choices
Research runtime that lets a bounded decision model like Jev operate over large tool, memory and context spaces via a virtual option space.
Overview
Decision-Native Agent Runtime is an open research runtime for bounded, non-generative decision models. Its central abstraction is a Virtual Option Space: the logical space of tools, memories, context blocks and refinements can be much larger than the options that fit in one decision request, so the runtime pages and refines that space while keeping stable IDs, revisions and validation. Jev is the first live backend, but the API is meant to be backend-independent.
Per the logicrw list, the design covers tool integration, speculative parameter proposals, top-k proposals from an external helper model's logits with Jev-controlled fallback, decision-aware hierarchical memory and dependency-aware replanning, plus a conversation prototype using helper logits and dynamic top-k token selection. A core profile runs locally with no GPU, weights or key as a deterministic protocol test; an optional 0.8B helper proposes candidates; and a live Jev mode reads the key from stdin. The CLI requires explicit :plan and :approve steps before any side effect.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which option from the current page of the virtual option space should be selected?(inferred) | — |
| Yes / no | Should this helper-proposed candidate be accepted or should Jev fall back?(inferred) | yes · no |
How it works
- Represent tools, memories and context as a large virtual option space
- Page a bounded subset into each Jev request
- Optionally use a small helper model's top-k logits to propose candidates
- Jev selects; runtime refines and replans
- Side effects require explicit :approve
Strengths
- Lets bounded decision models address large option spaces
- Explicit approval boundary keeps actions auditable
- Core profile runs without any model or key
Caveats
- Research prototype; autonomous use is a future target
- Setup documented for Windows PowerShell
Getting started
`pip install -e .` then `python -m jev_agent.cli --workspace .\agent_workspace`; add `--live --key-stdin` for Jev.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- CLI research runtime
- Has code
- yes
Use cases and integrations
How curated lists describe it
Research design for a Jev-native agent system: tool integration, speculative parameter proposals, external helper logits Top-k proposals with Jev-controlled fallback ,decision-aware hierarchical memory, and dependency-aware replanning.Feature:Jev naturallanguage conversation prototype using external helper logits and dynamic Top-k token selection.
logicrw/awesome-jev-projects
Frequently asked questions
What is a Virtual Option Space?
A logical set of options larger than one request can hold, paged and refined by the runtime with stable IDs.
Does it need Jev?
No for the core profile; live experiments use Jev as the first backend.
Can it act autonomously?
Not yet; the CLI requires :plan and :approve for every side effect.
Related projects
DeepSeek Harness jev_decide tool that brings the Jev decision paradigm to any OpenAI-compatible LLM, or to real Jev with a key.
ProjectprototypeTypeScriptReads candidate logits from existing MLX-LM/VLM models on Apple Silicon to map state and a request to an allowed action.
ProjectprototypePythonResearch runtime where every agent step is a bounded typed decision, with a paged virtual option space over tools and memory.
ServiceresearchPython- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Nautilus Compass agent memory layer whose jev-trust middleware logs Jev decisions and measures calibration in your domain.
ServicebetaPython