How to read the Dream-RSI study

This page explains the structure of the branching experiment in plain language.

Parent

A parent is the earlier attempt from which a child starts. A root has no parent. A child copies the parent program and gets one improvement opportunity.

Training

The 14 training cases are visible to Pi. They show the task shape and let Pi develop its first solution.

Validation

The 10 validation cases are hidden from Pi. The orchestrator uses their scores to decide which branch should receive an improvement.

Holdout

The 12 holdout cases are hidden from Pi and branch selection. They are used only for the final generalization comparison.

Algorithm

Prepare and split the cases.
Launch five independent root attempts.
Grade each root on training, validation, and holdout.
Improve a predetermined root for the fixed baseline.
Use validation evidence to select a parent for replay.
Improve the selected replay parent.
Compare both children on holdout.

Current limitation

This study is a local simulation, not a real Fabric execution. The onboarding cases are synthetic JSON fixtures, and Pi writes local solution.py programs. It does not yet run real Fabric Business Processes, Worker traces, PDFs, scanned images, or document-extraction tools. The next study should connect the same branching and replay design to isolated customer-32 Fabric development BPs with synthetic PDFs and images.

Open the detailed execution report →