What an alert means

The provider is asked to choose suspicious, routine, or other from a very small event description. This is a coarse advisory classification, not a determination about a person, a session, an exploit, or authorization correctness.

Result	What the plugin does	What an operator should infer
Valid suspicious at or above the confidence threshold	Counts recommendation; calls your callback with fixed event_kind, outcome, assessment fields if one is supplied. Counts alerted only after that callback succeeds.	Investigate with independently governed internal evidence. Do not assume an attack occurred.
Valid routine at or above threshold	Counts routine; no callback.	No advisory flag from this one limited event; not proof of safety.
other, low confidence, invalid/mismatched response	Counts abstain; no callback.	The plugin cannot draw a usable conclusion.
Timeout, transport error, queue full/expired, disabled or unknown event	Records the appropriate aggregate result; no callback.	Monitoring may have a blind spot. Do not reinterpret it as benign traffic.

The default threshold is 0.8 and requires a well-formed response: Jev-family model name, provider-reported usage, sole assessment choice answer, allowed choice and probability fields, finite confidence and consistent probabilities. This checks response shape, not model accuracy. There is no proven precision, recall, or false-positive rate for this project, and no live provider compatibility test has been performed.

Operating guidance

• Decide where your controlled callback sends safe alerts and who reviews them. It must be fast and nonblocking; callback errors are counted, not logged. This library has no durable delivery or built-in deduplication.
• Monitor Metrics.snapshot() for fixed-label results, queue_depth, inflight, provider_calls, and provider-call duration sum/max; Metrics.prometheus() returns an exposition string for a protected scrape endpoint. Sudden drops, abstentions, timeouts, or queue pressure matter as much as alerts. Do not add identities as metric labels.
• Investigate using existing security processes. Use the event kind as a lead, not as an identity or root-cause report. Correlation with richer private logs stays inside your own approved systems.
• Keep the plugin observational. It does not revoke tokens, change policies, block a login, or notify an end user by itself.

See the architecture for the nonblocking path, or measure latency before making any operational claims.