# Human label-preservation review — PASSED by user confirmation

The run creates PRIVATE/<out-name>-human-review.json before inference, containing96original samples/320template-sample pairs with original data labels, complete messages and blank human_decision,reviewer,reviewed_at,note. No model answers. Public results/human_review_pending.json has only row IDs/hashes and blank decisions; original data are not redistributable. Reconstruct by running run.py with verified downloaded resources. The local article has a private HTML convenience view.

A human reviewer must verify: input text is complete and unaltered; the new instruction still requests movie-review sentiment or the same news-topic classification; label meaning/order and A/B/C/D mapping are unchanged; output-contract changes do not change the target semantic label; ambiguity introduced by the template is recorded. Inspect all relevant variants side by side for each sample, with no reference to model success. Record preserved/changed/uncertain per pair, name/identifier and actual timestamp. This is transformation validation, not relabeling the original dataset or auditing model answers.

Do not auto-fill decisions. A completed row form or an explicit human confirmation can document review completion. Here the user directly confirmed “可以,真人审查通过了” on2026-10-05 in response to the outstanding092review request. See human_review_confirmation.json for the recorded confirmation time and exact packet hashes. Separate per-row annotations and the precise time the review was performed were not supplied; the original blank form is retained, not backfilled. If a variant is invalid, preserve the original numerical evidence and register any new protocol before rerunning; do not silently drop losing cases. Human review after this exploratory run is retrospective, not preregistered blinded acceptance of an unseen final test. Future confirmatory work needs review before inference on reserved rows.