m/market labResearch, with a record.
Source on GitHub ↗

Decision runners

A DecisionRunner connects one decision method to an experiment's execution and lifecycle. A runner is an App object, not a thread and not a process entrypoint. Platform creates it; ResearchController owns the experiment's single runner.

  • base.py: common DecisionRunner lifecycle interface.
  • policy/: Human/Rule observations, policies, Controlled runner, Live runner, policy invocation and authority handoff.
  • agent/: AgentDecisionRunner connects the Agent host, platform tools and common execution pipeline.
  • service.py: shared atomic decision audit/state/action commit logic, used by the financial transaction path. It is not a scheduling loop.
Human/Rule Controlled → policy/runner.py       → PolicyDecisionRunner
Human/Rule Live       → policy/live/runner.py  → LivePolicyDecisionRunner
Agent                → agent/runner.py        → AgentDecisionRunner

Background runners share experiment execution and work scheduling. Agent does not pass through LivePolicyDecisionRunner. The Agent program runs in a child hosted by agents/hosting.

The package's existing public exports (for example from sim_platform.decisions import DecisionConfig, HumanPolicy) remain available. Internal modules have one canonical location; there are no forwarding files at the former paths.

Maintained source: sim_platform/decisions/README.md