# Recorded Codex AI analysis

This run uses actual semantic judgments produced by the AI assistant in the Codex conversation on October 14, 2026. The website displays the recorded output; it does not generate new responses for visitors. No external automation platform or additional model API was used.

The 180 source records are synthetic. Only 17 normalized comment texts are distinct. Repetitions are linked, not counted as new inference. Portuguese and English texts remain distinct; this is not a sample of 17 independent customers.

CLASSIFIED.json contains the model-authored output. None of its judgments has been human-approved. Confidence is null because it was not calibrated. Two comments require priority review; every record remains pending human review. Counts are descriptive, not an accuracy or production-impact claim.

To repeat: freeze RAW_FEEDBACK.json and the taxonomy, supply PROMPT.txt and the distinct comments to Codex, save a new version of the model output, then run the validation tests. Running the tests does not rerun AI. A repeat model run may produce different judgments. The previous rules-based work remains in ../portfolio-feedback-case/feedback-validation as a baseline.

Independent evaluation is pending. The old development examples are not an independent human-labelled holdout. A human reviewer must label unseen multilingual, ambiguous and multi-topic examples before an accuracy comparison is reported.