How trust evolves
Two event types. One trust signal.
Trust is not static. It shifts in response to definition changes (severity events) and agent decisions (inference / decision events). Trigger both kinds below and watch the same consumption confidence signal absorb each one.
Definition changes
When a metric definition changes, the severity classifier counts blast radius across static signals (lineage, freshness, ownership, discoverability) and responds proportionally.
Select one or more definition changes. Blast radius and consumption confidence response update below.
Upstream sources and transformations
Recency of the underlying data
Accountable human or team
Name, description, tags, docs
Select one or more changes to preview the impact
Agent decisions
Every agent action produces two correctness events, inference (the read on the world) and decision (the action that followed). The quadrant separates them so robustness becomes measurable and decision logic becomes falsifiable.
Agent retrieved Monthly Recurring Revenue at $4.2M with trust score 87. Question: should it auto-act or escalate?
Was the read on the world correct?
Was the action that followed correct?
Select an inference and a decision to preview the impact
One signal, two streams
Both event types feed the same consumption confidence signal. The architecture has one trust state that absorbs both schema types.
No events yet. Apply a definition change or log an agent decision above.
Severity events and decision events have different schemas. The architecture treats them as different event types. They share one trust signal.
That is the unified abstraction.