Intermittent Sampling¶
Essence¶
Intermittent Sampling is the pattern of making fleeting conditions observable without watching everything all the time. It is useful when a state, failure, behavior, symptom, or environmental condition appears briefly and then disappears before ordinary inspection can see it. The intervention is not merely “check sometimes.” It is a sampling design: decide what counts, when and where to look, how long to observe, how to record context, and what response follows when a sample is positive.
The core insight is that partial observation can be powerful when the observation windows are deliberately chosen. A sparse sample can outperform a large amount of unfocused monitoring if it is timed near likely episodes, randomized against gaming, placed at informative sentinel points, or escalated during burst windows.
Compression statement¶
When a condition appears intermittently and continuous monitoring is costly, sample strategically to catch enough occurrences for diagnosis or control.
Canonical formula: intermittent condition + insufficient continuous observability + bounded sampling windows + capture criteria + follow-up response -> usable evidence without full-time monitoring
When to Use This Archetype¶
Use this archetype when the system repeatedly says, “The problem happens, but never when we are looking.” It fits situations where continuous monitoring is too expensive, intrusive, noisy, or technically impossible, yet waiting for normal review misses the condition. It also fits when evidence is anecdotal and decisions require more disciplined capture.
Do not use it when the condition is stable and easy to observe, when a one-time diagnostic test can reproduce the issue, or when misses are catastrophic enough that sparse observation is unacceptable. In those cases, continuous monitoring, fail-safe design, direct testing, or prevention may be more appropriate.
Structural Problem¶
The structural problem is a mismatch between episode duration and observation availability. The event exists, but not long enough or predictably enough for the normal observation rhythm to catch it. Reports accumulate without proof, failures vanish before inspection, and the system confuses “not seen” with “not real.”
This pattern often appears when bursts, rare events, compliance lapses, symptoms, or defects are real but irregular. The organization may overreact by monitoring everything continuously, or underreact by dismissing reports because scheduled reviews show normal conditions. Intermittent Sampling creates a middle path: bounded observation designed around the intermittent nature of the target.
Intervention Logic¶
The intervention begins by defining the sampling target. A vague concern such as “quality is inconsistent” becomes a more observable target such as “surface defect visible during the first hour after tool change.” The system then chooses a sampling strategy: scheduled, randomized, triggered, sentinel-based, rotating, or burst-window focused.
Each sample needs a window, a detection threshold, and a record. A sample that says only “checked” is weak. A useful sample records when observation was open, what was seen, what context surrounded it, whether the threshold was crossed, and what follow-up happened. The design is then tuned by reviewing false negatives, false positives, coverage gaps, and whether positive detections actually changed action.
Key Components¶
Intermittent Sampling makes fleeting conditions observable without watching everything all the time by designing observation moments around the intermittent nature of the target. The Sampling Target names the elusive state being sought — a failure mode, behavior, symptom, defect, spike, or lapse — so the work does not collapse into a generic inspection ritual. The Sampling Window defines when observation is active and how long it lasts; a window must be long enough to catch the state but bounded enough to preserve the point of sampling rather than drifting into continuous monitoring. The Sampling Schedule or Trigger determines how those windows open — on a planned interval, randomly, after warning signals, or at high-risk moments — translating a hypothesis about intermittency into actual observation chances. The Coverage Model prevents false confidence by clarifying what the sampling plan can and cannot see and whether the samples are representative enough for the decisions they inform.
The remaining components convert observation into evidence and action. The Detection Threshold separates signal from noise and sets the balance between false positives and false negatives, defining what counts as a positive sample. The Event Log preserves sampled observations along with timing, context, and useful non-detections; without a log, intermittent evidence remains anecdotal and disappears between windows. The Follow-up Response specifies what happens after detection — confirm, escalate, diagnose, increase sampling, repair, or change the sampling design — and is what makes intermittent sampling an archetype rather than a measurement habit. Several Optional Components strengthen the design where gaming, blind spots, or risk shifts would otherwise erode it: a randomization rule when predictable sampling would be gamed, a sentinel probe when a few observation points can reveal a wider condition, context metadata to capture load and configuration around each sample, a sampling escalation rule that increases density when risk rises, and a false-negative review that asks what the plan missed and how to improve it.
| Component | Description |
|---|---|
| Sampling Target ↗ | The sampling target names the intermittent state being sought. It can be a failure mode, behavior, symptom, defect, spike, lapse, or environmental condition. A clear target prevents the sample from becoming a generic inspection ritual. |
| Sampling Window ↗ | The sampling window defines when observation is active and how long it lasts. Window length matters because an intermittent event may be visible for seconds, minutes, hours, or a brief phase of a process. The window should be long enough to catch the state but bounded enough to preserve the point of sampling. |
| Sampling Schedule or Trigger ↗ | The schedule or trigger determines how sample windows open. Some samples happen on a planned interval, some randomly, some after warning signals, and some at high-risk moments. This component is where the intervention translates a hypothesis about intermittency into actual observation chances. |
| Coverage Model ↗ | The coverage model explains what the sampling plan can and cannot see. It may be statistical, heuristic, or practical. Its purpose is to prevent false confidence by clarifying whether samples are representative enough for the decisions being made. |
| Detection Threshold ↗ | The detection threshold defines what counts as a positive sample. It separates signal from noise and determines the balance between false positives and false negatives. A threshold can be a number, a categorical criterion, a behavioral marker, or a procedural condition. |
| Event Log ↗ | The event log preserves sampled observations. It records both detections and useful non-detections, plus timing, context, and follow-up. Without a log, intermittent evidence remains anecdotal and disappears between sample windows. |
| Follow-up Response ↗ | The follow-up response specifies what happens after detection. The response might confirm, escalate, diagnose, increase sampling, repair, intervene, or change the sampling design. Sampling becomes an archetype-level solution only when captured evidence changes action. |
Common Mechanisms¶
| Mechanism | Description |
|---|---|
| Spot Checks ↗ | Spot checks implement short bounded observations. They are useful when the target is frequent enough that a brief check has a realistic chance of seeing it. A spot check becomes Intermittent Sampling only when it is tied to a target, threshold, evidence record, and follow-up decision. |
| Randomized Audits ↗ | Randomized audits use unpredictability to reduce gaming and patterned blind spots. They are especially useful when actors could behave differently during known inspection times or when regular checks align poorly with the intermittent condition. |
| Canary Probes ↗ | Canary probes are lightweight signals or tests placed in the system to reveal whether an intermittent state may be occurring. They often support triggered sampling: the canary gives a warning, and a richer observation window opens. |
| Diagnostic Sampling ↗ | Diagnostic sampling collects targeted measurements when a condition is suspected but not continuously visible. It is common in technical diagnosis, medicine, quality control, and field investigation. |
| Burst Capture Logging ↗ | Burst capture logging increases observational density around suspected bursts. It is a mechanism for cases where episodes cluster around load spikes, seasonal events, incidents, or other high-risk windows. |
| Rotating Inspections ↗ | Rotating inspections sample different units, sites, teams, or subsystems over time. They broaden coverage without requiring every location to be inspected at once. |
| Sentinel Surveys and Temporary Sensors ↗ | Sentinel surveys gather evidence from selected people, locations, or institutions that are likely to show early signal. Temporary sensors instrument a system only during a suspected window or for a bounded period. Both are mechanisms that instantiate the broader sampling design. |
Parameter / Tuning Dimensions¶
The main tuning dimensions are sample frequency, window length, target specificity, trigger sensitivity, threshold strictness, randomness, sentinel placement, context depth, escalation criteria, and review interval. Each dimension changes a tradeoff.
Higher frequency improves capture probability but increases burden. Longer windows improve capture but approach continuous monitoring. Lower thresholds catch more episodes but create false positives. Randomness improves anti-gaming and coverage but can be harder to coordinate. Rich context improves diagnosis but raises privacy, complexity, or storage costs.
Invariants to Preserve¶
The first invariant is representativeness: the samples must be good enough for the action they inform. The second is bounded burden: the system should know when intermittent sampling is cheaper or safer than continuous monitoring and when it has crossed that line. The third is actionability: positive samples must trigger a defined response. The fourth is uncertainty honesty: sparse sampling must never pretend to be complete coverage.
Target Outcomes¶
A successful design catches more intermittent episodes with usable context. It reduces reliance on anecdotes, lowers the cost of observation compared with full monitoring, and improves response timing. It should also improve over time as false negatives, false positives, and coverage gaps are reviewed.
The target outcome is not perfect knowledge. The target is enough well-designed evidence to support better diagnosis, preparation, governance, or control.
Tradeoffs¶
Intermittent Sampling trades completeness for feasibility. It accepts that some events will be missed in exchange for lower cost, lower burden, or less intrusive observation. It also trades predictability against gaming: planned windows are easier to staff, but randomized windows may reveal more truth. It trades sensitivity against follow-up load: permissive thresholds catch more episodes but may overwhelm response capacity.
Ethically, sampling can become surveillance when people are the target. The design should minimize unnecessary data collection, distribute burden fairly, and make the purpose and governance of sampling clear where disclosure does not defeat the method.
Failure Modes¶
A common failure mode is the systematic miss pattern, where sampling always happens at the wrong time. The mitigation is to review misses, vary windows, add triggers, or randomize timing.
Another failure mode is false reassurance. Sparse non-detections can be misread as proof that no problem exists. The mitigation is to report coverage limits and known blind spots.
A third failure mode is sampling theater. The system performs checks to show diligence, but no one defines thresholds, records context, or responds to detections. The mitigation is to tie each sample to a follow-up decision.
A fourth failure mode is over-sampling creep, where bounded sampling expands into continuous monitoring without governance. The mitigation is to define observation budgets, sunset rules, escalation criteria, and privacy review.
Neighbor Distinctions¶
Intermittent Sampling is distinct from Observability Instrumentation because it is about bounded observation windows for transient conditions, not measurement infrastructure in general.
It is distinct from Statistical Sampling because the central problem is catching intermittent visibility, not estimating a population parameter. Statistical validity may matter, but it is not the whole archetype.
It is distinct from Intermittent Failure Capture because failure capture preserves detailed state during rare failure episodes. Intermittent Sampling may only detect that an episode occurred or provide screening evidence.
It is distinct from Perturbation Testing because perturbation testing induces a condition. Intermittent Sampling observes naturally occurring episodes.
It is distinct from Periodic Review and Reset because review-and-reset corrects drift; intermittent sampling primarily creates evidence about transient states.
Variants and Near Names¶
Scheduled Intermittent Sampling uses planned intervals. Randomized Intermittent Sampling uses unpredictability to prevent bias or gaming. Triggered Intermittent Sampling opens sample windows after warning signs. Sentinel Sampling chooses high-signal observation points. Burst-Window Sampling concentrates observation around suspected surges.
Near names include spot sampling, intermittent audit, event sampling, diagnostic sampling, random inspection, and burst capture sampling. These names should usually collapse into the parent unless they introduce distinct components and failure modes.
Intermittent Failure Capture remains a promotion candidate because it may deserve its own archetype. It emphasizes state snapshots, evidence preservation, and diagnosis during rare episodes, which may go beyond sampling design.
Cross-Domain Examples¶
In cloud reliability, a team may enable detailed traces only after warning-level latency spikes. This catches transient failure states without paying the cost of continuous deep tracing.
In manufacturing, rotating spot checks during different shift phases can reveal sporadic defects that final inspection misses.
In clinical care, a symptom diary can capture episodic dizziness, pain, mood, or side effects that are absent during appointments.
In compliance, randomized audits can reveal intermittent lapses that predictable reviews would miss.
In environmental monitoring, temporary sensors can sample noise, emissions, or flow during suspected peak windows instead of permanently instrumenting every possible location.
Non-Examples¶
A permanently installed sensor measuring a stable condition is continuous monitoring, not Intermittent Sampling. A quarterly review that resets goals is Periodic Review and Reset. A controlled stress test that intentionally creates a failure condition is Perturbation Testing. A dashboard that merely visualizes already collected data is a mechanism, not this archetype.