@meta
  v: 1
  route: /jebadiah
  generated: 2026-09-25T12:05:12.297Z
  ttl: 1h

@intent
  purpose: Product page for Jebadiah (Jeb), Frontier Infra's open System One decision model: typed questions in (choice / noul / score), calibrated probabilities out. Ships as frozen-version LoRA adapters (jebadiah-9b-v1, jebadiah-4b-v1, plus the v0 pair) and the jebadiah-synth-v2 dual-teacher dataset on Hugging Face, Apache-2.0, trained and served with AINode (/v1/decide, /v1/systemone).
  audience: ai-agent, developer, ml-engineer, agent-pipeline-builder
  capability: read, download, adopt, reproduce

@state
  versions[2]{version,models,training}:
    v0,frozen,11072-public-questions
    v1,frozen,14900-questions-ordinal-score-human-rubrics
  adapters[4]{repo,size,base}:
    jebadiah-9b-v1,9B,Qwen3.5-9B-Base
    jebadiah-4b-v1,4B,Qwen3.5-4B-Base
    jebadiah-9b-v0,9B,Qwen3.5-9B-Base
    jebadiah-4b-v0,4B,Qwen3.5-4B-Base
  question_types[3]{type,returns}:
    choice,probability-over-N-options
    noul,P(yes)
    score,ordered-rubric-placement
  dataset: jebadiah-synth-v2 (14714 questions, 3253 states, 24 families, dual-teacher, Apache-2.0)
  serving: AINode /v1/decide + /v1/systemone on any NVIDIA GPU
  calibration: temperatures.json applied offline today (ainode issue 276)
  license: Apache-2.0

@actions
  - id: huggingface_org
    method: GET
    href: https://huggingface.co/frontier-infra
  - id: model_9b_v1
    method: GET
    href: https://huggingface.co/frontier-infra/jebadiah-9b-v1
  - id: model_4b_v1
    method: GET
    href: https://huggingface.co/frontier-infra/jebadiah-4b-v1
  - id: dataset_synth_v2
    method: GET
    href: https://huggingface.co/datasets/frontier-infra/jebadiah-synth-v2
  - id: ainode_repo
    method: GET
    href: https://github.com/getainode/ainode
  - id: ainode_docs
    method: GET
    href: https://docs.ainode.dev
  - id: base_model_9b
    method: GET
    href: https://huggingface.co/Qwen/Qwen3.5-9B-Base

@context
  > Jebadiah answers three typed question shapes — choice (pick one of N named options), noul (a yes/no statement, returned as P(yes)), and score (place a state on an ordered rubric) — with a probability distribution over the labels instead of generated text. Output is calibrated per question type when the shipped temperatures are applied (temperatures.json; AINode issue 276 tracks applying them on the route — served routes return the raw distribution today, and a standalone script in each repo reproduces every calibrated number on the card). Training is public data only: the dual-teacher jebadiah-synth-v2 pool (14,714 questions, DeepSeek V4 Flash + Qwen3.8 27B teachers, both distributions kept in provenance) plus licensed human-rubric sources (HelpSteer2, SummEval) in v1. Versions are frozen: a new version is a new repository, so a decision recorded with v1 stays reproducible. Apache-2.0. Made in Texas.

@nav
  self: /jebadiah.agent
  parents: [/.agent]
  peers: [/.agent, /sdk.agent]