GOLD STANDARD · RADAR · HOW X WORKS
Find the decision hidden inside the model
A conditional-extremes CFAR monitor needs to notice regime change quickly without rebuilding the threshold for every noisy run. Sentinel windows turn that tension into an explicit trade-off among detection delay, false recalibration and the cost of waiting.
This guide connects the Gold Standard of Radar, Gold Standard of Education, and How X Works hub.
Chapter contents and reading routes
Choose the route closest to your present task. Every chapter remains expanded, so the article also reads naturally from start to finish.
Orient sentinel monitoring
Build the change score
Set delays and costs
Transfer and govern
Use evidence and next routes
CHAPTER 1 OF 12 · ORIENT THE DECISION
1. A sentinel window watches the model, not the target
A sentinel stream should contain clutter or reference cells whose behaviour reveals whether the fitted tail model remains valid. It is not simply another detection channel. Its job is to monitor conditional calibration, support and dependence before operational thresholds become misleading.
Choose sentinel cells, covariates and exclusions before evaluating alarms. If the stream is contaminated by true targets or post-alarm filtering, a change score can confuse mission activity with clutter drift.
CHAPTER 2 OF 12 · ORIENT THE DECISION
2. Detection delay and false alarms are coupled
A low change threshold responds quickly but triggers more false recalibrations. A high threshold is stable but allows a changed regime to operate longer under stale thresholds. Quickest-detection methods formalise this tension using delay subject to a false-alarm constraint.
For CFAR governance, add replay, fitting, validation, deployment risk and the alarms incurred while waiting. The best rule is not the one with shortest delay in isolation.
CHAPTER 3 OF 12 · BUILD THE METHOD
3. Construct scores in conditional tail space
Transform sentinel observations using the fitted conditional marginal and dependence model. Under calibration, exceedance indicators or probability-integral-transform residuals should behave according to a known reference, within estimation uncertainty.
Build a score from excess breaches, dependence residuals or a likelihood ratio between baseline and plausible shifted regimes. Keep covariate-support alarms separate so extrapolation does not masquerade as an estimated within-support change.
CHAPTER 4 OF 12 · BUILD THE METHOD
4. A worked two-window example
Suppose each window contains 200 sentinel episodes and the validated model expects about four tail breaches. Window A has six; Window B has nine, with seven concentrated in a newly common sea-state band. A one-window rule at eight triggers only on B.
A cumulative score can add each window’s excess above four and reset after sustained normality: A contributes two and B contributes five, crossing a six-point trigger. Real calibration must simulate dependence; the arithmetic shows why persistence can detect repeated moderate change.
| Evidence | Single-window view | Sequential view |
|---|---|---|
| 6 breaches | Below trigger | Adds moderate evidence |
| 9 breaches next | Triggers | Cumulative evidence crosses |
| Normal windows follow | No action | Score resets after recovery |
CHAPTER 5 OF 12 · BUILD THE METHOD
5. Choose window length from dynamics
Short windows localise abrupt shifts but contain few extremes and produce noisy scores. Long windows stabilise tail evidence but dilute rapid changes and delay localisation. Overlapping windows add resolution but create strong dependence between alerts.
Use physical timescales: scan cycles, weather transitions, processing releases and response time. Evaluate candidate windows on historical and simulated shifts, reporting delay as a distribution rather than one mean.
CHAPTER 6 OF 12 · TEST THE RESULT
6. Price a false recalibration honestly
A false trigger consumes review, shadow fitting and validation capacity. It can introduce a worse model if a small sample is mistaken for a new regime. Declining to recalibrate also has a cost when stale thresholds inflate false alarms or erode detection.
Record staff hours, replay load, false-alarm burden, missed-coverage exposure and deployment risk. Cost informs the action ladder after the safety floor is protected.
CHAPTER 7 OF 12 · TEST THE RESULT
7. Use a staged trigger ladder
The first level widens monitoring. The second freezes automatic release and fits a shadow model. The third approves recalibration only after support, coverage and operational replay pass. A severe unsupported regime routes directly to conservative fallback.
Persistence and cooldown prevent repeated refits from one episode, but cooldown must not silence independent severe evidence. Define reset, escalation and override rules explicitly.
| Level | Trigger evidence | Action |
|---|---|---|
| Watch | One moderate score | Retain data and shorten review |
| Shadow | Persistent score | Fit without deployment |
| Release | Coverage and replay pass | Deploy with rollback |
| Fallback | Severe unsupported regime | Use conservative thresholds |
CHAPTER 8 OF 12 · TRANSFER AND GOVERN
8. Changed-context transfer: coast to airport
A coastal radar may organise windows by sea state and tide; airport surveillance may use precipitation, runway configuration and interference. The stable logic is a conditional tail model, sentinel evidence, change statistic and controlled update path.
Transfer requires rebuilding contamination rules and timescales. A window spanning one tide cycle may have no meaning for short convective cells around an airport.
CHAPTER 9 OF 12 · TRANSFER AND GOVERN
9. Validate delay with injected shifts
Replay baseline data and inject changes in tail level, dependence and covariate mixture. Include gradual drift, abrupt jumps and temporary excursions. Measure detection probability, time to trigger, false-trigger interval and post-reset performance.
Include shifts the detector was not tuned for. A rule that excels only on its assumed likelihood ratio may miss structural changes preserving breach counts but altering spatial clustering.
CHAPTER 10 OF 12 · TRANSFER AND GOVERN
10. Limits and misconceptions
A change alarm does not identify cause. Sensor calibration, target contamination, processing changes and environment can produce similar scores. A quiet monitor does not prove coverage if current covariates lie outside training support.
Repeated overlapping windows do not create independent evidence. Calibrate thresholds through simulation or a dependence-aware method, with human approval for consequential releases.
CHAPTER 11 OF 12 · CHECK AND CONTINUE
11. Questions operators often ask
Should a trigger automatically refit? Usually not; separate detection, diagnosis, shadow validation and release. What false-alarm rate is acceptable? It depends on recalibration cost and stale-model risk, expressed over an operational horizon.
Can two weak windows outweigh one large breach? Yes, when persistence is expected under true change and the score is calibrated. A safety-critical single event may still justify immediate fallback.
CHAPTER 12 OF 12 · CHECK AND CONTINUE
12. Evidence, specialist routes and the next check
The conditional-extremes foundation is Heffernan and Tawn, with covariates in non-stationary conditional extremes. CUSUM quickest-detection research states the delay-versus-false-alarm objective; applied CUSUM guidance shows practical tuning.
Continue with regime-shift and support checks, rolling conditional coverage, and threshold stability. Inject one abrupt and one gradual shift, then compare delay and false-recalibration burden.