Record ECP-7 Batch 20 hard-meaning replay

[FatosLocos]

Summary

• add deterministic training-only mining of meanings failed by any population link
• replay ordinary reconstruction updates with one preregistered task-scale schedule
• run sealed seed-11 positive-control and intervention development arms
• record the valid negative result, compact artifact hashes, and justified Batch 21 direction

Result

Batch 20 reached 83.77% train exactness, a new-best 83.45% mean and 82.13% worst-link validation, and 83.96% universal-translator validation. The hard pool shrank from 3,452 to 2,848 at the selected checkpoint, but the remaining errors became more population-wide. Validation and translator gates pass; train thresholds and sender injectivity fail. The ECP-7 confirmatory split remains sealed.

Verification

• 109 passed
• configuration and schemas valid: 14336/1024/1024
• both post-hoc analyses match 65 artifact hashes
• 16,384 validation-only episodes per arm
• no confirmatory-test metrics
• all messages use valid local alphabets and exactly 14 declared bits