Context Feedback Discipline
Closing the RAG loop so the adaptive source-weighting cycle can learn.
Every context-assembly call must be paired with a context-feedback call. Without the pairing, the adaptive source-weighting eval cycle has nothing to learn from and the dashboard's "used" bar never renders.
What context assembly does
Assembly fans out across all memory sources — the psychic cache, persistent memory, agent memory, and the vault index — applies per-caller source weights, allocates a token budget, and returns a context block plus an assembly_id.
The feedback pairing
At task completion, LISA and every dispatched agent submit feedback with:
assembly_id— the id returned from assembly.used_source_ids[]— the sources that materially shaped the output (empty array if nothing was used).tokens_used— the count actually consumed from the assembled context.
Why it matters
Source weights drift over time via the Dream eval cycle based on these signals. The effectiveness of a source is scored from a blend of the confidence of the work it fed, how often it is actually used, and how often it appears. Feedback is the only input to that loop — so calling assembly and then forgetting to submit feedback leaves the row half-written and the loop learns nothing.