cookbook/environments/_11_export_provenance/TEST_LOG.md
Fix: both tasks saturated at k=4, so learning_zone() was empty and the file
printed its guard ("No learning-zone tasks; make the tasks harder") instead of
exporting — the provenance sidecar it exists to show never appeared. product-b
replaced with a second independent calibrated chain (expected 10481347) so a zone
row is reliable and the export runs.
Grid (k=4): product-a 3/4 (0.75, zone); product-b 4/4. Exported 3 dataset
rows and 3 sidecar rows with env/policy fingerprints.
inspect_sidecar.py (a=0.50, zone) re-ran clean and unchanged.
Tested 2026-07-20 with OpenAIResponses(id="gpt-5.5", reasoning_effort="low").
Status: PASS
Description: Exported a verified dataset and loaded the generated provenance sidecar.
Result: product-a passed 3/4 (0.75) and product-b passed 4/4 (1.00).
The dataset and sidecar each contained three entries, with both fingerprints
present in the sidecar.
Status: PASS
Description: Validated sidecar fingerprints and passing-attempt references, then joined trusted exported rows to task ids, zero-based attempt indexes, and scores.
Result: Final live run after adding stale-pair cleanup and non-null fingerprint
guards: product-a passed 3/6 (0.50) and product-d passed 6/6 (1.00).
Three exported rows aligned one-to-one with valid passing attempts and matching
fingerprints. The example also printed that the sidecar records provenance but
does not authenticate the JSONL because it carries no dataset digest.
Calibration: The first post-guard rerun saturated at 4/4 on both product-a
and product-c. That grid was rejected; product-c was replaced by the harder
product-d boundary and K was raised to six before this PASS was recorded.