Problem
Retail brokerages send intermittent, privacy-constrained updates. The goal is stable inference without any single dominant participant.
Scarcity components exercised
Constraint ledger (access + privacy), aggregation patterns (reliability weighting), simulation harness (client dropout), resilience checks (participation + gradient divergence).
What worked
Synthetic clients surfaced failure modes early; reliability weighting prevented over-indexing on spiky brokers.
What failed
When a high-volume client went silent, stale updates still warped the model. Validation data missed weekend/holiday behavior.
Next steps
Increase personalization per broker, add staleness decay, broaden validation cohorts.