Behavior checks need to distinguish completed work from valid observations, keep authority and effects separate, and survive interrupted local runs. - Add neutral v3 observation and cohort schemas with qualified verdicts, source provenance, controls and compatible historical readers. - Journal action intents, dispatch, responses, reconciliation and effects; permit explicit recovery only after an independent non-commitment read. - Record delivery stages and enforce revisioned revocation at the final connector and browser dispatch boundary. - Add bounded conversation session leases while preserving task isolation by default and keeping retained browser state private. - Add explicit clocks, bounded availability checks and durable schedules with distinct attempts, effects, end conditions and cleanup evidence. - Separate credential receipt, treatment, consumption and disclosure using generated synthetic fixtures and existing private-fill protections. - Require exposure and independent controls for decisive behavior verdicts. - Bind optional answer annotations to task, action, object, time and phase while preserving relevant corrections and default lexical scoring. - Add a single-writer sandbox runner with durable send slots, conservative reconciliation and an allowlisted aggregate publication view. Document the interfaces, limitations, nine architecture proposals and 13 generic behavior families. Keep generated validation records outside Git. Review: claude -p; apply material findings and verify the final corrections. Validation: 4,569 tests passed (32 deselected); 398 focused tests passed; eight scripted connector CI examples and three readiness suites passed. Browser/timeout tests passed repeated runs and all-core CPU load. Ruff, schema consistency, repository checks and publication checks passed.