DedeuceRL

Benchmark LLMs on active hidden-system identification.

DedeuceRL is organized around four layers:

• dedeucerl.kernel: pure hidden-system semantics
• dedeucerl.ir: executable action spaces, hypothesis/equivalence contracts, task contracts, and renderers
• dedeucerl.runtime: budget, traps, events, tool dispatch, and replay
• dedeucerl.surface: prompts, provider tool schemas, datasets, CLIs, Verifiers, and MCP

Install

pip install dedeucerl
pip install "dedeucerl[openai]"
pip install "dedeucerl[all]"

Quickstart

dedeucerl-generate --skin mealy --seeds 0-9 --budget 25 --n-states 3 -o tasks.json
dedeucerl-eval --skin mealy --split tasks.json --model heuristic:none --out results.jsonl
dedeucerl-aggregate results.jsonl --format markdown

mealy is the current reference kernel. Protocol/APIEnv/ExprPolicy are planned to return as kernels after the architecture stabilizes.

MCP Mode

Expose one stateful episode to any MCP host over STDIO:

dedeucerl-mcp serve --task mealy --seed 42

The MCP initialization supplies the task instructions and tools. DedeuceRL automatically records the semantic tool trace, scores terminal episodes, and writes result.json plus trace.jsonl under .dedeucerl/runs/<run-id>/. Terminal responses include the score, while discovery-only processes create no artifacts. No provider adapter, manual run-spec tool, or manual scoring step is required.

Prime / Verifiers

uv tool install prime
prime lab setup

Use dedeucerl.vf_env as the Verifiers entrypoint.

License: MIT.