Ainfera Research — Open Science for Inference Routing

Open science for outcome-aware inference routing for AI agents. Public research surface for Ainfera.

Overview

This repository contains the research artifacts behind Ainfera's outcome-aware routing methodology. Everything is reproducible — benchmarks, datasets, evaluation harness, and methodology documentation.

What's Here

Directory	Description
benchmark/	Benchmark harness — run_benchmark.py
methodology/	Methodology docs — judge protocol, empirical Q, exploration floor
datasets/	Dataset generation + licensing
eval/	Delta evaluation, replay, comparison
labs/	Lab experiments
preprint/	Preprint manuscripts (LaTeX)

Methodology

Ainfera's routing research is built on three principles:

1. Outcome-aware — routing decisions learn from observed outcomes (latency, quality, cost), not static model metadata
2. Reproducible — every benchmark run is deterministic with pinned seeds and recorded env state
3. Open — methodology, datasets, and evaluation code are public under CC-BY 4.0

Key Documents

• methodology/judge-protocol.md — LLM-as-judge evaluation protocol
• methodology/q-empirical.md — Empirical quality scoring
• methodology/exploration-floor.md — Exploration vs. exploitation in routing

Quick Start

# Install dependencies
pip install -r benchmark/requirements.txt

# Run the benchmark suite
python benchmark/run_benchmark.py

# Evaluate routing delta (before vs after)
python eval/delta.py --baseline results-baseline.json --candidate results-candidate.json

Citation

See CITATION.cff.

License

Creative Commons Attribution 4.0 International — see LICENSE.