Frontier Infra

Frontier Infra

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.

Models and data here

Jebadiah (Jeb) is our open System One decision model: it answers typed questions (choice, yes or no, score) with a calibrated probability over the options instead of generating text. It is trained and served with AINode, the platform that turns any NVIDIA GPU into a local AI node, so an owner can train, serve and measure a decision model on their own hardware with the same prompt bytes end to end. Public data only; every number on a card cites an evaluation record in the repo.

v2 is out. Merged bf16 weights, not LoRA adapters. The base is now the chat checkpoint (Qwen/Qwen3.5-9B, Qwen/Qwen3.5-4B, thinking off) where v0 and v1 used the -Base checkpoints; the public training pool and recipe are otherwise the same as v1. Training on the chat checkpoint was worth more than every data change we tried: synthetic pools, a stronger teacher's labels, human-labelled yes/no.

• jebadiah-9b-v2: the current 9B, merged weights on the Qwen3.5-9B chat checkpoint
• jebadiah-4b-v2: the current 4B, same recipe on Qwen3.5-4B
• jebadiah-9b-v1 and jebadiah-4b-v1: frozen, LoRA adapters on the -Base checkpoints
• jebadiah-9b-v0 and jebadiah-4b-v0: frozen, the first public versions
• jebadiah-synth-v2: 14,714 synthetic typed-decision questions across 24 families, teacher-labelled on our own fleet, Apache-2.0

Overall score on our 20-set public suite (macro over the zero-shot public sets): 9B v2 73.93 (v1 73.29, v0 72.5); 4B v2 72.49 (v1 70.34, v0 70.9). On JevBench's 231 public items, 9B v2 scores 0.818 (the same as v1) and 4B v2 scores 0.758; Jev 1.13.0 is at 0.866 on the board. We are still behind Jev, and the gap is on the harder sets.

AINode's /v1/systemone speaks TypeSafe's Jev wire format, so an existing Jev client works by changing the endpoint; /v1/decide is AINode's own route. Both live in getainode/ainode. The code and training recipe, with a standalone server in server/, are public at getainode/jebadiah. Applying the model's fitted temperatures on those routes is merged and lands in the next AINode release. Versions are frozen; a new version is a new repository.

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