Ainfera Research

Open science for outcome-aware inference routing for AI agents — the public research surface for Ainfera Inference.

Ainfera routes every agent call to the model that will complete the task, learned from outcomes, neutral across providers. This repo holds the method and the results. It does not hold the operated policy, the real labeled outcome corpus, or the production judge — those are the moat and stay closed.

What's here

Path	Contents
preprint/	arXiv preprint source — the routing-delta result + methodology
methodology/	Public-safe specs: q_empirical, judge protocol, exploration floor
benchmark/	Reproducible harness — routing vs baselines on synthetic/public data
eval/	Replay + delta-experiment scripts (operate on synthetic data here)
datasets/	CC0 synthetic outcome sets for reproducibility

What's deliberately NOT here

The real labeled corpus, the operated q_empirical weights, judge prompts, and anti-gaming logic. Those live in private infrastructure. Reproducing our headline number on synthetic data is possible from this repo; reproducing it on our data is not — by design.

Quickstart

pip install -r benchmark/requirements.txt
python datasets/generate.py           # writes datasets/synthetic_outcomes.jsonl
python benchmark/run_benchmark.py      # routing vs baselines → delta table

The thesis in one line

A rational agent never leaves a router whose all-in task cost is always below its own baseline — and the gap widens as the router learns. This repo measures that gap on open data so anyone can check the shape of the claim.

Related

• ainfera-ai/specs — open methodology + API contracts (CC-BY 4.0)
• ainfera-ai/routing — policy templates + production methodology
• ainfera-ai/verify — offline audit-chain verifier

License

Code: Apache 2.0 (LICENSE). Papers: CC-BY 4.0 (preprint/LICENSE). Datasets: CC0 (datasets/LICENSE).