Record sealed ECP-7 context-invariance batch

[FatosLocos]

Summary

• preregister ECP7-B8-I as one algebraic context-invariance intervention on the B7 parallel sender
• activate the inherited training-only quadruple loss without assigning factors to message slots
• keep the rejected Batch 4 through Batch 6 losses disabled
• run the unchanged ECP-6 positive control and B8 intervention on the sealed development split
• preserve compact integrity evidence and select a generic position-aware decoder as the next isolated question

Development result

The positive control again achieved 100% on every registered measure. ECP7-B8-I reduced its relaxed transition surrogate but suppressed the discrete gains from Batch 7:

• selected algebraic loss: 0.05265 at scheduled weight 0.75
• train exactness: 0.5349%, down from 9.5965%
• unique messages per sender: 857–1,135, down from 3,118–3,415
• population validation mean: 0.3235%, down from 0.5066%
• universal-translator validation: 0.6836%, down from 2.7588%

The intervention is recorded as a negative result without tuning its weight. The confirmatory split remains sealed.

Validation

• 70 pytest tests pass
• B8 and positive-control configs validate at 14336/1024/1024
• both smoke arms complete through isolated workers
• all 65 captured artifact hashes match for each full run
• all 16,384 episode rows per arm are validation-only
• no confirmatory-test metrics exist
• every message respects the exact 14-bit local channel