EQ-Layer

EQ-Layer is an early-stage, model-agnostic control layer that gives a language model an explicit decision about how to respond before it responds. It tracks affect, intent belief, repair state, and interaction quality across turns, selects a response policy, and compiles that policy into factored controls that are injected into the decode path. The intermediate decisions stay inspectable rather than being folded into a prompt instruction.

The project does not retrain the model and does not claim that affect-aware dialogue policy, intent routing, or repair are new. Its narrower question is whether a transparent control layer can improve a fixed general-purpose model while keeping each decision auditable. See Prior art and novelty for the historical map and the claims this project deliberately avoids.

Status

Early. The evaluation harness runs, but its numbers are development signals, not a real-world EQ benchmark. There is no standard benchmark for EQ; eval/ is a first attempt at measuring one.

Documentation

• Prior art and novelty — the research boundary and relevant pre-LLM work.
• A/B latency and quality benchmark — what the layer costs and what it decides, measured with the layer enabled and disabled.
• Publishing — how the PyPI package and npm launcher are versioned and released.

Install

pip install "eq-layer[ml]"

The zero-dependency CLI route needs no trained artifacts:

python -m eq_layer.cli route --intent-mode heuristic --pretty

The full learned pipeline requires .joblib artifacts produced by the training commands in the repository README; those artifacts are not shipped.