HOW INTELLIGENCE WORKS · INTERVENTION SELECTION · eduKateSG
How a Mind Chooses the Action That Will Reveal the Most About What Causes What
Intervention selection is the intelligence process that chooses what to change in the world—not merely what to observe—so that competing causal explanations produce different outcomes and the next result teaches us something decisive.
Competing causal models → controllable variable → predicted outcomes → information gain → intervention cost → bounded action → observe return → update causal model.
This article belongs to the How Intelligence Works series. Causal Reasoning owns the distinction between correlation and cause. Hypothesis Testing owns discriminating evidence more broadly. Intervention selection owns the active question: what should we deliberately change so the world can reveal which causal model is better?
Observation Can Leave Causes Entangled
Two variables can move together for several reasons: one may cause the other, the direction may be reversed, both may respond to a third factor, or the pattern may be accidental.
Observation can narrow the possibilities. Sometimes only intervention can separate them cleanly.
If two explanations predict the same observations but different consequences under intervention, changing the system can be more informative than watching it longer.
1. Start With Competing Causal Models
A useful intervention begins before the action. Intelligence first asks which causal structures are still plausible and what each predicts if one variable is changed while relevant alternatives are controlled or observed.
Without competing models, intervention can collapse into random tinkering.
2. Choose a Variable That the Models Treat Differently
| Intervention property | Why it matters |
|---|---|
| Controllable | The variable can actually be changed |
| Discriminating | Competing models predict different outcomes |
| Measurable | The return can be observed reliably |
| Bounded | The action limits unnecessary harm or disruption |
| Reversible where possible | The system can recover if the model is wrong |
| Low-confounding | Other changes are limited or documented |
3. Intervention Selection and Value of Information Are Different
Value of Information asks which missing fact is worth obtaining. Intervention selection asks which action should create the most useful new evidence about causal structure.
One prices information. The other engineers the experiment that can produce it.
4. Good Interventions Change One Important Thing, Not Everything
If many conditions change at once, a later improvement may be real while the cause remains unclear. Strong interventions isolate enough structure that the result can update the model rather than merely celebrate an outcome.
A successful change is not automatically an informative change.
5. Intervention Selection in Mathematics
Mathematical exploration often changes one parameter and watches what remains invariant. Vary a coefficient, alter a boundary condition or test an extreme case. The aim is not physical causation but the same control logic: manipulate one part of the representation to expose dependency.
6. Intervention Selection in Learning
A student performs poorly on a word problem. Possible causes include language, representation, prerequisite knowledge, retrieval, working memory or method selection.
Instead of assigning more of the same worksheet, a teacher can intervene diagnostically: simplify the language while preserving the mathematics, provide a diagram but no method, remove time pressure, or test the prerequisite directly.
The student’s change in performance helps identify the first weak link.
7. Intervention Selection in Operations
Operational teams use pilots, A/B tests, staged rollouts and controlled parameter changes to learn before scaling. The most informative intervention is often the smallest one that can distinguish the important causal models.
Large irreversible changes may produce large outcomes while teaching very little about which mechanism mattered.
8. Intervention Cost Changes the Best Experiment
A maximally informative intervention may be too expensive, unethical or disruptive. Intelligence therefore balances expected information gain against risk, reversibility, time and consequence.
This is where Trade-off Reasoning and Reversibility Reasoning enter the causal-learning loop.
9. Intervention-Selection Failure Atlas
| Failure | What happens | Repair |
|---|---|---|
| Random tinkering | Changes occur without explicit competing models | State predictions first |
| Everything-at-once change | Outcome improves but causal ownership stays unclear | Isolate critical variables |
| Outcome-only thinking | Action seeks success but produces little learning | Design for discrimination |
| Confounded intervention | Untracked changes move with the treatment | Hold stable or record alternatives |
| Irreversible experiment | Wrong model creates unnecessary lock-in | Stage and bound action |
| Weak measurement | The return cannot distinguish explanations | Define observable outcomes first |
| No update | Experiment runs but model remains unchanged | Predefine revision conditions |
10. Intervention Selection and Hypothesis Testing Are Different
Hypothesis testing covers the whole logic of competing explanations and discriminating evidence. Intervention selection is the active-control branch inside that logic.
It asks which deliberate change can make the hypotheses diverge in the world.
11. Teams Need Experiments With Owners
A team experiment should state who owns the causal question, which variable changes, which outcomes matter, what side effects are monitored and what evidence would alter the next decision.
An experiment without an update rule is activity, not institutional learning.
12. Institutions Need Safe Learning Loops
Institutions often need to learn under real constraints. Pilots, staged deployment, sandbox environments, controlled comparisons and reversible policy trials can generate causal evidence while limiting unnecessary exposure.
Not every domain permits experimentation, and rights or safety may constrain the action set. Intelligence includes recognising those boundaries before intervention.
13. Artificial Intelligence and Intervention Selection
Tool-using AI can move from observation to action. That makes intervention selection a safety-critical capability.
A reliable agent should distinguish read-only information gathering from state-changing actions, prefer bounded and reversible probes when uncertainty is high, respect permissions, predict side effects and use the resulting return to update its model.
The ability to act is not permission to experiment indiscriminately.
14. The Intervention Selection Audit
- Models: Which causal explanations remain plausible?
- Variable: What can be changed?
- Prediction: How should each model respond?
- Discrimination: Will the outcomes separate the models?
- Measurement: Can the return be observed reliably?
- Confounding: What else may change?
- Cost: What does the intervention consume or risk?
- Reversibility: Can the system recover?
- Authority: Who is allowed to act?
- Update: What result changes the next model or action?
15. CivDJ Reading: Touch the Control That Makes the Masters Disagree
In the CivDJ frame, intervention selection chooses a bounded action whose World Return would differ depending on which working Master is correct.
The operator is not changing the mix for drama. The intervention is chosen because it can reveal hidden structure.
Act where the return is most diagnostic, not merely where the action is easiest.
16. Return to the Action That Teaches
Intervention selection turns action into a learning instrument.
It asks not only, “What should we do?” but “Which responsible action will make the hidden causal structure more visible?”
The mature intelligence uses the world not as a stage for random experimentation, but as a disciplined partner in causal correction.