symfonic.capabilities.memory.phases.evidence_rules

evidence_rules

Which episodic rows phase 12 is allowed to learn from, and how it groups them.

Split from :mod:.procedural, which crossed its 300-line budget coming into this package. The split is a real seam rather than a byte count: everything here decides what counts as evidence -- whose turn it was, when it happened, and which action a line narrates -- and the module it left decides what to do with the groups that come out.

The speaker filter is the load-bearing one. Under the default policy an assistant-narrated row is skipped, and so is a row with no speaker at all: the first would let the agent promote its own narration into a skill and read it back next turn, and the second is a pre-v7.7.2 write that packed both speakers into one line, where "I ran the deploy" may be something the agent said.