The first fresh-test evaluation read all1200cases and computed predictions.
It failed while writing fresh-test-predictions.jsonl:
TypeError: Object of type int64 is not JSON serializable
at evaluate_steward.py line77 json.dumps(a,separators=(',',':')).
No test report or aggregate metric was written. No checkpoint, model, feature,
target, method, metric, or selection was altered. The original evaluator is
preserved as evaluate_steward-before-json-fix.py and its pretest hash remains.
The corrected serializer converts NumPy integers to ordinary integers. The
same frozen evaluation is rerun, verifying all original pinned files (using
the preserved originals for two separately documented helper fixes).
Test bytes must also match the earlier data-generation manifest. This is a
local author-maintained sequence, not independently timestamped preregistration.

The first retry stopped before loading test because the broad original
tools/*.py freeze had also captured make_steward_art.py while another agent
was still developing that publication helper. The retry verifier now records
differences in explicitly listed non-model publication/audit helpers instead
of treating them as model mismatches. All actual model, training, feature,
checkpoint, teacher, and original evaluation files remain hash-verified.
The original broad freeze is unchanged; incidental differences are retained
in freeze-incidental-files.json.
