Arion Judgment Core

English | 简体中文

Reality provides experience; counterfactuals produce agency.

Arion Judgment Core is a small framework for organizing how a collective judgment forms, remains accountable, accepts challenge, and changes when reality contradicts it.

It is not a fixed role pipeline. A Domain Pack supplies governed roles, question profiles, routes, and executors. Core keeps the durable judgment state and creates temporary Agent instances only for the open questions that currently matter.

Arion was distilled from a Chief Risk Officer project. Its original assembly kept eight responsibilities distinct: Lead, Data Analyst, Strategy Analyst, Anti-Fraud Specialist, Model Engineer, Risk Platform Engineer, Release Engineer, and Release Verifier.

Those roles are not hard-coded into Core. They demonstrate why governance roles and Agent instances must remain different: evidence, policy counterfactuals, adversarial adaptation, model uncertainty, implementation, controlled action, and independent verification cannot safely collapse into one generic executor.

The public repository excludes private business vocabulary, metrics, systems, thresholds, models, and release procedures. See the sanitized risk-management walkthrough for one complete collaboration.

• Incompleteness: unknowns remain visible instead of being erased by workflow completion.
• Asymmetry: action and inaction can carry different, irreversible costs.
• Revisability: every closed judgment records what would reopen it.
• Responsibility conservation: execution can be delegated; accountability cannot disappear.

• Task preserves the long-lived objective and accountability boundary.
• Working Set preserves experience, unknowns, hypotheses, counterfactuals, obligations, and the current judgment.
• Run is one bounded attempt to revise that Working Set.
• Direct Run handles one bounded assignment.
• TaskGraph appears only for multiple assignments, real dependencies, waiting, retries, or governed actions.
• A governance role may own multiple parallel Agent instances.

python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[test]'
pytest -q

from pathlib import Path

from planner.domain_pack import load_domain_pack
from planner.semantic_execution_planner import plan_runtime

pack = load_domain_pack(
    pack_root=Path("open_source/reference_pack"),
    manifest_path="manifest.yaml",
)
planned = plan_runtime(
    domain_pack=pack,
    question="Explore a counterfactual and competing hypothesis.",
    run_id="example-run",
)

print(planned.orchestration_mode)
print(planned.capability_plan.to_dict())

The included reference pack is entirely synthetic. It demonstrates the one-way dependency Domain Pack -> Arion Core without business data, production identity, or private operating knowledge.

Core does not ship business vocabulary, data connectors, model pipelines, release procedures, credentials, runtime payloads, or a default governance organization. A Domain Pack may narrow Core safety rules but may not weaken them.

See the risk-management walkthrough, Chinese Guidebook, Principles, Architecture, and the Domain Pack contract.

Arion Core runs tests, dependency auditing, and complete-history secret scanning for every proposed change. Please report suspected vulnerabilities through the private process in SECURITY.md.

Apache License 2.0.