Make entity_observer reasoning nullable while keeping it required, and mark the audit-only reasoning and summary fields as truncatable so canonical validation can handle over-cap audit prose. Update the prompt to ask for one short reasoning clause on add/update/drop, null reasoning on keep, and keep operations only when deliberately re-affirming an observation against fresh evidence. This targets the measured 24.3% per-call schema_invalid failure rate on the daily entity_observer talent.