- NOWInterruptA decision cannot wait.
- LATERDigestUseful context, batched.
- QUIETRecordNo action; keep the evidence.
signal = change × confidence × consequence − duplication
§1A monitor is a decision system
Polling is easy. Attention policy is the design problem. Without one, a monitor either goes silent when a source breaks or becomes a notification machine that trains people to ignore it.
The useful architecture separates five jobs: collection, normalization, change detection, interpretation, and delivery. Each can fail independently, so each needs its own health state.
§2Never overwrite the observation
For schedules, listings, advisories, prices, or policy pages, the previous state is part of the evidence. Store snapshots with observed_at, first_seen_at, source identity, and retrieval status. Derive changes from the sequence.
A disappeared item should begin as a candidate change, not an immediate fact. The source may be degraded, identifiers may have changed, or a replacement may exist. Confirmation comes from repetition, provider health, or an independent source.
§3Give every source an honest state
A current response was captured and validated.
The source is healthy and no material delta appeared.
A new, normalized delta passed the change rule.
The source responded, but completeness or freshness is impaired.
No trustworthy current observation exists.
Unknown is not unchanged. That one distinction prevents a broken collector from producing a false all-clear.
§4Separate severity from confidence
A severe event with weak evidence and a small change with high confidence are different operational objects. Keep both dimensions visible.
| Evidence | Consequence | Route |
|---|---|---|
| Weak | High | Collect faster; seek corroboration; label uncertainty. |
| Strong | High | Interrupt with source, timestamp, delta, and required decision. |
| Strong | Low | Batch into the next digest. |
| No material delta | Any | Record silently; preserve source-health evidence. |
§5Use models for interpretation, not reality
A model can classify an update, cluster duplicates, translate, or write a compact digest. It should not invent the underlying observation, send the alert by itself, or silently turn “possibly related” into “caused by.” Core facts, scoring inputs, deduplication keys, budgets, and delivery state stay deterministic.
Models may explain why a change could matter. The database must still show what changed, when it was first seen, which source said it, and whether delivery succeeded.
§6Send a decision receipt
The useful alert is not “something happened.” It is a compact evidence packet:
WHAT CHANGED
Exact normalized delta
EVIDENCE
Source · observed_at · first_seen_at · corroboration
STATE
Confidence · severity · source health · duplicate status
DECISION
Why this crossed the threshold · what needs attention now
LIMIT
What remains unknown or unavailableDelivery itself is another state machine: queued, sent, acknowledged, failed, or dead-lettered. A successful API response proves that a request travelled; it does not necessarily prove that the intended human saw the right message.
The compact rule
- Preserve snapshots; derive changes.
- Keep unknown separate from unchanged.
- Require repeated or corroborated anomalies where false alarms are costly.
- Separate confidence from consequence.
- Interrupt only with a decision-ready evidence packet.