[stevenzheng1990]

Submit Fileverity V4 (step 4000) for official RAID hidden-test evaluation.

• Product: https://fileverity.com
• Contact: hello@fileverity.com
• Coverage: all 672,000 IDs from the official test.csv, each with one continuous AI score in [0, 1]. No test IDs were skipped.
• Training disclosure: the V4 training corpus contains 39,582 records marked as RAID/legacy training replay. Please classify this as a detector trained with RAID data.
• The checkpoint was frozen before hidden-test inference. No hidden-test labels were available, and no fitting or threshold selection was performed on the hidden test set.
• Scores use the same frozen text preprocessing and model for every record. A zero-token input is evaluated through the actual model using CLS/SEP tokens; no fixed fallback scores are used.
• Only predictions.json and metadata.json are included. No model weights, training texts, credentials, or locally calculated results are submitted.

Prediction file SHA256: 2c115fec47fc65ae24ed3b34e91efaa1322cf43b5fca78c2cb9f13953cfff4b0.