Reality provides experience; counterfactuals produce agency.
Arion Volunteer combines the Arion Judgment Core with an open eight-role governance model and a reusable Skill catalog.
It is a small framework for organizing how a collective judgment forms, remains accountable, accepts challenge, and changes when reality contradicts it. Roles keep responsibility stable; temporary Agent instances investigate only the unresolved questions that matter in the current Run.
- the current Arion planning and execution Core;
- eight governance roles with explicit responsibility boundaries;
- eleven reusable
SKILL.mdoperating methods; - a local Volunteer Pack with role-neutral methods, routes, and generated evidence;
- tests for planning, parallel Agents, closure, role assets, and the public privacy boundary.
- no business semantic layer, metric catalog, source definitions, or field dictionary;
- no account profiles, credentials, tokens, sessions, private keys, or secret values;
- no company data, customer data, private knowledge, internal endpoints, or production adapters;
- no runtime state, model artifacts, deployment packages, or inherited private Git history.
The included Pack declares semantic_roots: [] and its only Connector reads
generated local evidence with no secret slots.
| Role | Irreducible responsibility |
|---|---|
| Lead | Own the bounded collective judgment and closure |
| Data Analyst | Establish what the evidence actually supports |
| Strategy Analyst | Compare action counterfactuals and tradeoffs |
| Anti-Fraud Specialist | Test adversarial adaptation and evasion |
| Model Engineer | Bound model uncertainty and failure |
| Risk Platform Engineer | Preserve the judgment contract in implementation |
| Release Engineer | Prepare or apply an explicitly approved reversible change |
| Release Verifier | Independently verify the change claim |
Not every task activates every role. Core creates one Direct Run for one bounded assignment and a TaskGraph only for real parallelism, dependency, waiting, retry, or governed action.
python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[test]'
python scripts/verify_public_boundary.py
pytestfrom planner.domain_pack import load_domain_pack
from planner.execution_planner import plan_runtime
from volunteer_pack import bundled_pack_root
pack = load_domain_pack(
pack_root=bundled_pack_root(),
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.decision.active_roles)This question activates Strategy Analyst and leaves Lead as the accountable
closure gate.
| Path | Purpose |
|---|---|
planner/ |
Task, Working Set, Direct Run, TaskGraph, role activation, execution, and safety |
risk_methods/ |
Public role-neutral method contracts |
volunteer_pack/roles/ |
Human-readable governance-role instructions |
volunteer_pack/skills/ |
Reusable role and team Skills |
volunteer_pack/*.yaml |
Runnable role, route, method, Connector, and executor composition |
tests/ |
Core, Pack, role-asset, privacy, and closure verification |
Read Architecture, Roles and Skills, Quick Start, and Public Boundary for the complete contract.
Arion Volunteer defaults to local execution and does not grant external write authority. Release roles model responsibility boundaries; another project must provide its own explicit authorization, adapters, and independent controls.
Every change is expected to pass tests, dependency auditing, public-boundary checks, and complete-history secret scanning.
Apache License 2.0.