replication-pack

To package an AI experiment into a reproducible bundle, provide a run directory or let the system select the latest run. The workflow collects logs, setup, and environment data to create a deterministic ZIP file.

A deterministic replication package is a self-contained bundle that allows an AI experiment to be shared and cited. It includes prompts, model IDs, responses, thinking traces, and cost logs for exact reproducibility.

Environment snapshotting works by recording the setup and configuration of your run directory. This data is added to a MANIFEST alongside run metadata and a SHA256 hash to ensure the bundle remains self-contained.

Yes, the experiment packaging process strips credentials from the bundle. It ensures all data is based only on logged records, keeping the shared ZIP file secure while preserving provenance and cost logs.

Yes, you can package runs from a .research directory, as well as experiments or runs directories. The workflow automatically identifies these locations to collect raw calls and environment data for the bundle.

The main limitation is that the bundle is strictly based on logged records. If your raw calls or cost logs are incomplete, the resulting deterministic ZIP will lack that missing provenance data for reproduction.