Record ECP-7 Batch 18 attenuated replay result

[FatosLocos]

Summary

• reduce only the Batch 17 collision-replay coefficient from 1.0 to 0.1
• preserve all mining, replay, architecture, optimizer, data, and selection settings
• preregister and run ECP-7 Batch 18 on the sealed development split
• record the valid partial improvement with verified artifact hashes
• keep the confirmatory test sealed and preregister the rationale for one replay-pulse follow-up

Result

Batch 18 reaches a new-best 84.08% train exactness and 84.06% universal-translator validation exactness, with 80.71% population validation. Validation and translator thresholds pass together, and hard code use improves for most senders. Known-meaning train thresholds and sender injectivity still fail, so the full joint gate remains closed. Batch 15 retains the highest population validation result.

Verification

• 102 passed
• ecp7 validate: 14336/1024/1024
• both analyzers passed with 65/65 artifact hashes
• 16,384 validation-only episodes per run
• no confirmatory-test metrics or episodes
• valid exact 14-bit local channel
• manifest JSON and artifact hashes independently verified
• git diff --check