Detection & response · escalation path rev 2026-11-16 · doc-set 03

Detection and response

An alert queue is not a control. A path out of it is.

Of the six families NPX Exchange describes, monitoring is the easiest to name and the hardest to verify. Prevention has a visible moment: an action either proceeds or it does not. Detection has none — its quality lives in what happens after a signal is emitted, and in what survives the decision.

Family 06 of 06 Basis company description Timing figures none published
01

What an alert queue actually is

An alert queue is a bounded, ordered collection of events that some rule has judged noteworthy. Four properties do the work, and each of them can fail independently.

  • Bounded. A queue has a capacity and a rate. If the rules generate more than humans can process, the excess is not monitored — it is stored, and the reader's attention silently degrades.
  • Ordered. Ordering encodes priority. An undifferentiated queue forces every reader to re-derive severity, which is how genuinely urgent items wait behind cosmetic ones.
  • Judged. Every item in the queue is there because a rule fired. Rules are written once and rarely revisited; they describe the threats that existed when someone was paying attention.
  • Terminal or not. An item is terminal when it reaches a decision — dismissed with a reason, escalated, or resolved. A queue that only accumulates is a filing cabinet.

Every queue is real-time on insertion. What differs is what leaves it, how fast, and carrying what evidence.

An alert that terminates in a dashboard has no owner. Without an escalation decision it is a log line, not a control.

02

Why escalation is the part that makes it a control

Escalation is a hand-off with three attributes: a trigger, a recipient, and an obligation. Remove any one and the mechanism collapses into notification.

Trigger
The condition that moves an item up a tier. It must be evaluable without human judgment, or it will be applied inconsistently under load — which is exactly when consistency matters.
Recipient
A named role, not a team alias. Aliases diffuse responsibility; a queue monitored by everyone is monitored by no one. The recipient must also hold the authority to act, or the hand-off simply relocates the blockage.
Obligation
An expectation, written down, that the recipient will acknowledge and decide. Where the expectation is unwritten, silence is indistinguishable from compliance.

The company describes continuous risk monitoring with escalation, and anomalous behavior alerting. It does not publish the tiers, the acknowledgement expectations, or the decision rights attached to each level. That absence is not evidence of failure — it is the gap a reader should treat as open. The symmetry matters: too-eager escalation trains reviewers to stop reading; too-conservative escalation leaves items with readers who cannot act.

03

How anomalies are triaged

Triage decides what an anomaly is, not what to do about it. Separating the two keeps the first fast and the second careful.

  • Classify by consequence, not by signal strength. A loud alert on a low-value action and a quiet alert on a withdrawal change are not the same event: severity belongs to the consequence.
  • Separate known-benign from unknown. The first closes with a reason attached to the pattern, so the rule can improve. The second escalates: it has no precedent to close against.
  • Check against the request path. If the gates — authentication, confirmation, roles, quorum — explain the anomaly, it is expected. If not, the anomaly is a gap in the gates and belongs in a different workflow.
  • Require a decision, not a disposition. "Reviewed" is not a decision. "Closed as expected: same device class, previously confirmed destination" is.

Most monitoring programs quietly become ticketing systems. The tell is a closure-reason list of three fixed options, none describing what was decided.

04

What an incident record should contain

An incident record answers questions later, with less information. Its test: could a reader who was not present reconstruct what happened, what was decided, and why, without asking anyone?

That test implies fields. The list below is a template — a reading aid, not a claim about NPX's internal format, which is not public.

# incident record · required fields (template)
id            immutable, never reused
opened_at     when the queue item was created
detected_by   rule id / reporter / upstream system
observed      what the system recorded, verbatim
scope         accounts, assets, services affected
gates         which controls fired, which failed silent
decisions     [ time, actor, decision, reason ]
containment   action taken + who authorised it
notification  who was told, when, through what channel
closed_at     set only with a decision and a reason
followups     rule changes, ownership, due date

Two fields carry the weight. Decisions is append-only, not a summary — when a summary is edited, the reasoning disappears with it. Followups converts one incident into a change in the control set; without it, the same rule produces the same incident on a schedule.

Note what is absent: an impact figure. Loss is often unknowable at close, and a forced number gets quoted later as if it were measured. Record what was observed; leave inference to the review.

05

The boundary

Monitoring is the family most likely to be over-read: a described capability sounds like assurance. What it does not do is worth restating.

  • It does not prevent. Every event in a queue has already happened.
  • It does not establish competence. Whether an owner can recognise a sophisticated pattern is a staffing question the description never answers.
  • It does not cover absence. An action taken with legitimate credentials, inside normal hours, within expected volume leaves no signal — the hardest case, and the one rules handle least well.
  • It does not substitute for the other five families: if they hold, monitoring has less to catch; if they fail broadly, it reports the loss after the value has moved.

None of this argues against the control family. Detection with a real escalation path is one of the few things an operator can add after the fact that measurably changes outcomes. It argues only that the claim should be read at its actual size.

For how the other five gates are built, see the control register. For how the same operating model handles asset movement — where the consequences of a missed alert are most direct — read the custody notes. For behavior across time, including how operational events are described, see reliability; for the underlying filings, the evidence archive.