Record sealed ECP-7 learning rate decay batch

[FatosLocos]

Summary

• add a validated late linear learning-rate schedule with per-checkpoint logging
• preregister and execute sealed ECP7-B14-I on the complete Batch 10 base
• record the valid negative result, integrity evidence, and exact artifact hashes
• retain Batch 10 and narrow ECP7-B15 to a duration-only 30,000-step extension

Result

The unchanged positive control remained at 100%. ECP7-B14-I reached 79.13% population train exactness, 65.49% validation exactness, and 71.02% translator validation exactness. The translator threshold passed, but train, validation, and injectivity did not. Late learning-rate decay therefore regressed the strongest Batch 10 trajectory.

The confirmatory ECP-7 test split remains sealed.

Verification

• 88 passed
• ECP7-B14 configuration validates with split sizes 14336/1024/1024
• both sealed analyzers passed with 65/65 artifact hashes and no test metrics
• all eight hashes recorded in batch14-manifest.json match the local run artifacts