Record sealed ECP-7 joint-collision batch

[FatosLocos]

Summary

• preregister ECP7-B4-I as one direct full-message collision intervention on top of Batch 3
• add a factor-agnostic straight-through loss that counts distinct minibatch inputs sharing a complete message
• exclude duplicate sampled inputs without using factor identities or target slots
• run the unchanged ECP-6 positive control and B4 intervention on the sealed development split
• preserve the regression and compact integrity evidence
• identify sender consensus on the stronger Batch 3 base as the next isolated question

Development result

The positive control again achieved 100% on every registered measure. ECP7-B4-I regressed relative to Batch 3:

• unique messages per sender: 426–579, down from 585–972
• hard-message entropy: 7.46–7.84 bits, down from 8.20–8.73
• population validation mean: 0.4211%, down from 1.9226%
• universal-translator validation: 0.6104%, down from 2.2705%

The local collision loss reached 0.03125 at the selected checkpoint, but remained a sparse proxy for global codebook occupancy and disrupted previously learned size and texture factors. The intervention is recorded as a negative result without tuning its weight.

Validation

• 58 pytest tests pass
• B4 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