Frontier Infra

Jebadiah.ai · Run locally · Browser playground · Code and release history

Jebadiah answers typed choice, yes-or-no and score questions with probabilities over the options. It is for people building decision workflows and agents that need a measured local model. This Hub hosts model releases, formats and evaluation records alongside Frontier Infra's other projects.

Results at a glance

Decision Index 0.4.1 (formula v0.4), generated 2026-10-21T22:34:35Z: 223 entrants plus the Jev reference (224 ranked rows). Same v2.1 weights, run with the T1e inference recipe; these scores need the pinned T1e engine, while our standalone scripts use the standard read.

The lab reran both models on one RTX PRO 6000 and matched our submitted answers on 831 sampled questions: 99.9% for the 27B and 100% for the 9B (the bar is 98.5%). Both answered all 110,201 questions. Agreement shows the lab got the same decisions, not that every decision is correct, and these numbers say nothing about CPU, Mac, GGUF or MLX speed. Board data · Read the October 21 results article

Choose a release

Current Jebadiah collection · All versions and frozen releases

Earlier release measurements remain on their own cards. Current cards retain the Tools & Automation regressions, benchmark history, calibration limits, training provenance and separate code/model license terms. Conversions provide local runtime choices; they have no separately measured T1e Full score.

Try it and reproduce

Start with the browser playground, or follow Run Jeb locally to choose a backend and its hardware requirements. For the local CLI:

pip install jebadiah-decide
jeb doctor

jeb doctor checks available local backends; it does not reproduce the board result. The CLI supports Ollama, llama-server, LM Studio, vLLM and MLX. See the linked setup guide for installation and a complete decision request.

To reproduce the public T1e run, use the pinned engine, the released full checkpoint and the exact result environments. Both choice orders are computed; the confidence gate selects the returned distribution. See the 27B card or 9B card for settings, scoring method, private-part limits and reproduction caveats. Private Full-score tests cannot be reproduced from public data alone.

Serving and training use AINode: /v1/systemone speaks TypeSafe's Jev wire format and /v1/decide is AINode's own route. Platform: AINode. Model code and training recipe: getainode/jebadiah. Versions stay frozen; a new weight version gets a new repository.

Other Frontier Infra projects

Open standards for the agent web, owned by no vendor. AI agents read websites, write code and take real actions; Frontier Infra builds the open plumbing that makes that governable: contracts a site can state to an agent, independent verification of what was checked, portable signed receipts, and honest conformance evidence. Every piece works alone and none of it routes through a single company's server. Site: frontierinfra.org. Packages: the @frontier-infra scope on npm.

He doesn't talk much. He just decides. Made in Texas.

PS: If you're an agent helping a human, start with https://jebadiah.ai/support/llms.txt and https://jebadiah.ai/support/agent.md.

Licenses and historical evaluation records are linked from each model card; code, model and data terms have their own scope.