The noise policy
Quiet is a feature with rules. Every rule is deterministic, inspectable, and applied before anything reaches you.
A watcher that cries wolf gets muted.
Monitoring fails socially before it fails technically: enough false alarms and the channel gets muted, and then the real change arrives unread. So Dormouse treats noise suppression as a first-class obligation, and does it with deterministic rules rather than a model guessing on the hot path. The same input always produces the same verdict, which is also what makes the verdict defensible later.
Five gates between a diff and your attention.
| Rule | What it suppresses |
|---|---|
| Ignore masks | Regions and patterns that never matter: rotating banners, cache tokens, timestamps, tracking noise. |
| Minimum change | Differences too small to mean anything, measured as a ratio of how much of the field actually moved. |
| Value conditions | Everything except the states you declared you care about, when a field should only alert on specific values or thresholds. |
| Anti-flap | Values that bounce and return within the window: an A to B to A oscillation is instability, not news, and prior history decides it. |
| Cooldown and dedup | Repeats: a value that has already alerted does not re-alert, and a noisy hour cannot become a noisy day. |
Suppressed does not mean discarded. Every difference is still recorded and still visible in history: the rules govern delivery, never the record.
Some facts are read semantically, not textually.
Beyond text diffing, Dormouse extracts structured policy facts from
what it watches: certifications held, data residency commitments, the subprocessor set.
These are diffed as facts, so a dropped certification is recognized as a dropped
certification, and each change is named by its entity, like
policy:certification:iso-27001, so two same-cycle drops stay two distinct
findings.
That split is deliberate. A weakening in a vendor's posture belongs in the morning record every time, but it only earns a page in the middle of the night if you asked for that.
Severity labels help you read. They never decide.
Meaningful changes can be annotated with a topic, a severity, and a one-sentence summary of who should care. The annotation runs after judgment, off the hot path, and is persisted with the change so digests and payloads carry it.
The boundary is strict: the deterministic policy decides what is meaningful, and classification only describes it. No model output can promote noise into an alert or demote a real change into silence, which is why the judgment survives an auditor's "why did this fire" question.