Watch & Monitor

Watch solves the staleness problem that plagues every knowledge system. When you pin an answer or artifact, VaultCrux fingerprints every source chunk that contributed to it - the specific text segments, their embedding vectors, and their positions within the parent documents. These fingerprints form the baseline. Watch then monitors the underlying corpus on a configurable schedule (daily, weekly, or on-change) and compares the current state of those chunks against the baseline.

When a source document is re-ingested or updated, Watch identifies exactly which chunks have changed and computes a staleness score. The score is not binary - it accounts for the magnitude of the change (minor formatting vs. substantive content revision), the position of the changed chunk relative to the pinned answer (did the change affect the most-cited passage or a peripheral one?), and whether the change was additive (new information) or contradictory (information that conflicts with the original).

Alerts are severity-graded. An informational alert means source material has changed but the change is unlikely to affect the pinned answer. A warning means substantive changes were detected in chunks that directly support the answer. A critical alert means contradictory information has been detected - the pinned answer may now be factually incorrect given the updated corpus.

Watch alerts integrate with the broader monitoring infrastructure. They surface in get_session_context so agents see them at session start. They feed into get_pressure_status so the system-wide health view reflects answer reliability. And they are checked by verify_before_acting - if an agent proposes an action that depends on a pinned answer with active Watch alerts, the verification gate flags it.

For teams, Watch provides accountability. The alert history for any pinned answer shows exactly when staleness was detected, who was notified, and whether the answer was re-evaluated or replaced. This is the audit trail that compliance teams need: not just "was this answer correct when generated?" but "was it still correct when acted upon?"

The monitoring schedule is deliberate. On-change monitoring catches updates as they happen but requires real-time ingestion pipeline integration. Daily and weekly schedules are better suited for corpora that update in batches (document management systems, regulatory filings, vendor contract repositories). You can mix schedules across different watches - real-time for critical operational knowledge, weekly for reference material.