cookbook/environments/_10_export_sft/TEST_LOG.md
Tested 2026-07-20 with OpenAIResponses(id="gpt-5.5", reasoning_effort="low").
Status: PASS
Description: Ran two arithmetic tasks four times, selected both learning-zone rows, and exported only passing text conversations.
Result: product-a passed 1/4 (0.25) and product-b passed 3/4 (0.75).
The exporter wrote four conversations and skipped four failed attempts.
Status: PASS
Description: Verified that the default exporter keeps passing attempts and excludes failed attempts within selected learning-zone tasks.
Result: product-a passed 4/6 (0.67) and product-d passed 5/6 (0.83).
The JSONL contained exactly the nine passing attempts; three failed attempts
were excluded.
Calibration: The first task set (product-a, product-c, k=4) saturated at
4/4 on both rows. product-c was replaced with the harder product-d and k was
raised to six before this PASS was recorded.
Status: PASS
Description: Removed any stale artifact pair, exported the real learning-zone selection, then exercised an explicit empty selection without calling the exporter.
Result: Final live run after adding stale-output cleanup: easy-anchor passed
4/4 (1.00), product-a passed 1/4 (0.25), and product-b passed 4/4 (1.00).
The real selection wrote one row; the synthetic empty selection was detected
and skipped. If a future real run has an empty zone, assertions verify that no
prior JSONL or sidecar remains at the advertised paths.