Export Provenance

Keep verification evidence beside a portable SFT dataset. The JSONL contains only conversational messages; scores, attempt indexes, and environment and policy fingerprints live in a deterministic .meta.json sidecar.

Files

basic.py — export a learning-zone dataset and locate its sidecar.
inspect_sidecar.py — read the sidecar and join each JSONL row back to its task, attempt, and score.

When to use

Use this after _10_export_sft/ when a dataset needs recorded provenance. Keep the JSONL and sidecar together in a trusted artifact store when moving them. The next folder, _12_trainer_loader/, shows the separate consumer boundary: a trainer loader reads only the message rows.

The sidecar records where the exported rows came from, but it does not contain a digest that authenticates the JSONL. Treat the pair as one trusted artifact; a same-length replacement JSONL would not be detected by the sidecar alone. It also does not claim that training happened or that the dataset will improve a model.

Run

bash

python cookbook/environments/_11_export_provenance/basic.py
python cookbook/environments/_11_export_provenance/inspect_sidecar.py

Requires OPENAI_API_KEY. Every example uses gpt-5.5 through OpenAIResponses.