Online learning

• python/abria_server/online_learning.py
• python/abria_server/reasoning_learning.py

online_learning.py can be enabled independently and supports configurable modes/profiles, recent/replay batches and training cadence.

The module tracks minimum example counts for shadow, assist and autonomous stages. Defaults in this source are 8, 36 and 120 respectively, with environment overrides.

Bad outputs are not blindly replayed as targets. Explicit corrections can be turned into supervised signals while low-quality outputs are excluded.

The API exposes neural/completion/relation probe endpoints so online learning can be inspected instead of promoted solely on training loss.