daf-jev: Composable Python Toolkit

The second. The record presents a live-API characterization of the model bundled with the toolkit, not an external evaluation. It does say every manuscript number is generated from benchmark artifacts with a machine-checked provenance chain, but the roughly 18x speedup and 4x token reduction come from batching questions inside this toolkit, so re-measure on your own workload.

Both. The record publishes software release 0.3.0 as a 2.1 MB source zip alongside a 715.6 KB rendered manuscript PDF, and the abstract lists question builders, confidence gates, tiered routing, a concurrent corpus evaluator, a calibration module, a CLI, an agent skill and an MCP server over a thin typed client.

No. What is documented is that reported confidence was self-consistent across repeated evaluations, which is repeatability rather than correctness. Self-consistency means the same input yields a similar number again; whether that number tracks reality is what you have to establish against labelled data of your own.

It is a single-author software record by Daniel Ari Friedman of the Active Inference Institute, published under MIT on Zenodo with a DOI and versioned releases. The captured record is 0.3.0 and Zenodo notes a newer version exists, so read that before pinning anything.

From Zenodo by DOI: 10.5281/zenodo.22817425 for release 0.3.0, and the concept DOI 10.5281/zenodo.22816187 which resolves to whatever the newest version is. That version record carries two files, daf-jev_combined.pdf at 715,620 bytes and daf-jev-0.3.0-source.zip at 2,095,260 bytes, each with an md5 checksum published alongside it.

A decision-composition library sits between the question builders and the transports: confidence gates, tiered routing and composite scoring, so a pipeline is assembled from parts rather than written inline. The same capability is then exposed through four surfaces, a typed client, a CLI, an agent skill and an MCP server, with a concurrent corpus evaluator and a calibration module alongside for batch work.

Not at volume, and the record is small. It shows 221 downloads, 179 of them unique, with 51 on 0.3.0 itself, published 2026-09-17, and the concept record already points at a newer version. It is a single-author software record with no source repository linked, so read the numbers as a signal about the artifact rather than about the toolkit.