jevbooks

← All projects

jev-agent-design-with-topk-logits-choices

6Mikao9/jev-agent-design-with-topk-logits-choices

Jev-native agent runtime with virtual option spaces and progressive refinement.

Agent Decisions100%Runner-up: Classification & RoutingThermostat

What it is

A research prototype for building agent systems around Jev's structured decision interface. It targets tool use, hierarchical memory, and natural language dialogue for decision models, with a Python implementation and benchmarks.

How it uses Jev

Jev makes the final choice among candidate tool parameters, fragments, or tokens. Helper logits propose Top-k tokens; Jev selects the next token or falls back. The saved dialogue and speed experiments use local_top1_proxy, not real Jev per-token decisions.

Primitives:choice

Technique worth stealing

Decision-preserving progressive refinement: PAGE/EXPAND increases candidate coverage, REFINE reduces granularity, Jev always makes the final choice.

Try it

python -m unittest discover -s tests -v

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.87
  • Shows a System One patternyes0.74
  • Handles uncertaintyno0.38
  • Measuredno0.06
  • Runnableno0.12
  • Worth recommendingno0.34
  • Model replicaunclear0.45
  • Problem scopescore on a 0–2 scale1.13
  • About Jevyes0.98

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 23 Sept 2026.