HOW INTELLIGENCE WORKS · CHANGE DETECTION · eduKateSG
How a Mind Notices When the World Has Shifted Enough to Require a New Model
Change detection is the intelligence process that notices when incoming observations no longer fit the expected pattern strongly enough to justify reopening the current state estimate, rule, model or plan.
Expected pattern → new observation → discrepancy → noise check → persistence check → change signal → reopen model → update state, boundary or plan.
This article belongs to the How Intelligence Works series. State Estimation owns reconstruction of the current hidden state; Pattern Recognition owns recurring structure. Change detection owns the transition question: has the world moved far enough that the old pattern should no longer be trusted automatically?
The World-Has-Moved Problem
A student who was stable begins making a new kind of error. A familiar process takes longer. A sensor pattern drifts. A market, weather system, class dynamic or software environment changes enough that yesterday’s assumptions begin producing today’s mistakes.
Intelligence needs a way to distinguish ordinary variation from meaningful change.
The danger is not only failing to notice change. It is also changing the model every time the world merely wiggles.
1. Change Begins as Prediction Error
Every working model carries expectations. Change detection begins when observation deviates from those expectations.
One discrepancy may be noise. Repeated, structured or high-consequence discrepancies deserve more attention because they suggest that the generating process itself may have changed.
2. Change and Noise Are Different
Noise is fluctuation around a stable process. Change is alteration in the process, state or relation that generated the observations.
| Signal | Likely interpretation |
|---|---|
| Small isolated deviation | Possible noise |
| Persistent drift | Possible gradual change |
| Sudden sustained jump | Possible regime shift |
| New error type | Possible structural change |
| Multiple independent signals shift together | Stronger evidence of system change |
3. Thresholds Decide When Intelligence Reopens the Model
A highly sensitive threshold catches change early but produces false alarms. A conservative threshold avoids false alarms but may detect real change too late.
The right threshold depends on cost, reversibility and speed. A high-stakes system may tolerate more investigation to catch important change earlier.
Change detection is an argument about how much surprise is enough to justify attention.
4. Change Detection and Conflict Monitoring Are Different
Conflict Monitoring detects incompatible active routes. Change detection compares the current stream against an expected baseline and asks whether the generating environment itself may have shifted.
A detected change may later create conflict among old and new models, but the initial signal is temporal and comparative rather than merely competitive.
5. Change Detection in Mathematics
In sequences, graphs and dynamic problems, solvers look for points where behaviour changes: increasing becomes decreasing, linear becomes nonlinear, one constraint becomes active or a piecewise rule switches.
Recognising the transition prevents the learner from extending one local pattern across the entire domain.
6. Change Detection in Learning
Learning systems should notice both improvement and deterioration. A student’s error pattern may shift after a new topic, a period of absence, increased time pressure or successful repair.
The useful question is not simply “Is the score different?” but “Has the mechanism of performance changed?”
A new misconception, stronger retrieval or changed response strategy can matter even when the total score moves only slightly.
7. Change Detection in Operations
Operations depend on baselines: normal cycle time, vibration, error rate, demand, temperature or throughput. Detection systems watch for departures large enough to justify inspection.
Early change detection can convert an emergency into maintenance by making drift visible before the system crosses a failure threshold.
8. Change Detection Should Trigger Investigation, Not Instant Storytelling
Detecting that something changed is not the same as knowing why it changed. The mind is tempted to attach the first plausible cause immediately.
A strong system separates detection from explanation: first establish that the shift is real, then generate and test competing causes.
9. Change-Detection Failure Atlas
| Failure | What happens | Repair |
|---|---|---|
| Change blindness | Old model continues after environment shifts | Monitor residuals and edge signals |
| Noise chasing | Every fluctuation triggers a new model | Require persistence or corroboration |
| Baseline drift | Normal slowly changes without being redefined | Refresh reference state |
| Late detection | Shift is recognised only after failure | Move sensors upstream |
| Single-signal dependence | One bad sensor creates false change | Use independent evidence |
| Cause collapse | Detection is mistaken for explanation | Separate signal from hypothesis |
| Adaptation lag | Change is noticed but action stays old | Link signal to model review |
10. Change Detection and Belief Revision Are Different
Belief Revision changes the model after evidence deserves an update. Change detection supplies one common trigger by showing that the previous baseline no longer explains the stream adequately.
Detection says, “reopen.” Revision decides what changes.
11. Teams Need Shared Baselines
Teams cannot detect change consistently if each member carries a different definition of normal.
Strong teams specify the baseline, threshold, evidence source and owner who decides whether a detected shift deserves escalation.
You cannot agree that something changed until you know what everyone thought normal was.
12. Institutions Need Change Sensors at the Edge
Large institutions often receive change late because local signals are compressed, filtered or normalised before reaching decision-makers.
Healthy systems preserve channels for anomalies, complaints, frontline observations and independent metrics so that regime shift can become visible before official reports fully reflect it.
13. Artificial Intelligence and Change Detection
AI systems operating over changing data need to detect drift in users, sources, tools, environment and task distribution.
A system that performs well on yesterday’s distribution may quietly degrade when inputs shift. Reliable deployment therefore monitors behaviour, error types, confidence, data freshness and the distance between current inputs and the conditions under which the model was evaluated.
Change signals should trigger reevaluation rather than automatic self-justification.
14. The Change Detection Audit
- Baseline: What counts as normal?
- Expectation: What should the current model predict?
- Discrepancy: What changed in the observations?
- Noise: Could the shift be random variation?
- Persistence: Does the shift continue?
- Corroboration: Do independent signals agree?
- Threshold: Is the change large enough to matter?
- Consequence: What breaks if the old model continues?
- Investigation: Which hypotheses could explain the shift?
- Update: Which state, boundary, model or plan should reopen?
15. CivDJ Reading: Hear When the Mix Has Drifted
In the CivDJ frame, change detection watches whether the Receiver, case state or operating environment has moved far enough that the current Master mix no longer fits.
The Tumbler should not keep increasing the volume of an old channel when the receiver state itself has changed.
A good mixer notices when yesterday’s balance has become today’s distortion.
16. Return to the Moment the Old Map Stops Fitting
Change detection is intelligence noticing that continuity has broken.
It protects the system from two opposite failures: freezing the old model while the world moves, and rebuilding the model every time noise appears.
The mature system can say, with evidence, “something important is different now.”