Testing Autonomous Systems in Possible Worlds: Inside PEIRAVELA’s Experiment Control Plane

When agents, workflows, robots, or security controls begin influencing real business, teams must determine how the system behaves if the environment changes. Traditional testing validates predefined inputs and outputs, making it difficult to observe behavior under abnormal, boundary, or counterfactual conditions.

PEIRAVELA defines itself as a “possible‑world experiment control plane.” It creates governed experiment worlds, runs scenarios and counterfactual branches, injects controllable perturbations, and generates complete raw simulation evidence.

Typical experiment loop

create world

derive branch

execute scenario

observe result

compare control variables

destroy environment

validate evidence

replay if necessary

Each “world” is an isolated object with a clear lifecycle. Every execution binds a scenario, perturbation, control variables, and a random seed, allowing runs to be distinguished and compared.

PEIRAVELA supports a matrix of scenario × perturbation × control × seed . Teams can keep the scenario and seed constant while varying a governance policy, model version, or resource limit to pinpoint the factor causing a change.

Recorded facts per run

scenario and control configuration summary

random seed and clock

memory, CPU, and runtime

observer health status

environment cleanup results

simulation‑to‑reality calibration and fidelity information

explicit statements of missing information

It is the original evidence producer, not the ultimate judge.

The platform records what happened but does not claim pass/fail, trust, or release decisions. Pass/Fail, assurance claims, and release gates are handled by external evaluation or engineering assurance systems, preventing the platform from both generating and interpreting its own data.

Evidence packages can be independently verified. An open repository provides a verification method that relies only on tar archives and SHA‑256 hashes, allowing third parties to check contents without installing PEIRAVELA.

At the execution layer, PEIRAVELA can simulate numeric results, measure real process CPU and memory, run constrained local processes, and enforce default network denial and cgroup resource limits inside Linux containers. This enables a gradual transition from logical simulation to controlled real execution.

The architecture separates three responsibilities:

simulation system creates and controls worlds

evidence system faithfully records observations

evaluation system makes judgments based on evidence

This separation is important for agents, digital twins, robots, and security systems, ensuring teams know which world ran, what conditions changed, what was observed, and whether results can be independently validated.

Project repository: https://github.com/axisrobo/peiravela-open